Podcasts about Minimax

FactSet U.S. Daily Market Preview

Play Episode Listen Later Sep 6, 2026 0:15

kimi minimax

Broadcom terug bij af: waar is het vertrouwen in deze stille killer?

Beurswatch | BNR

Play Episode Listen Later Sep 3, 2026 23:02

Weer kwam chipmaker Broadcom met mooie cijfers. En weer wordt het aandeel gedumpt door beleggers. Het duwt het aandeel zelfs weer onder de koers van begin dit jaar. Waarom zijn beleggers zo negatief? Dat zoeken we uit in deze aflevering. Want zoals gezegd: aan de cijfers lag het niet. De totale omzet kwam uit op een kleine 30 miljard dollar, een stijging van 86 procent jaar op jaar. En de nettowinst ging zelfs met 216 procent omhoog. Hebben we het ook nog over oliereus Chevron. Dat kwam gisteren plotseling nog met een aankondiging: ze gaan flink investeren in Venezuela, de komende vijf jaar. 7 miljard dollar maar liefst. Daarmee willen ze de olieproductie daar verdubbelen naar 600.000 vaten per dag. Best een belofte. Of ze het ook waar kunnen maken, gaan we voor je uitzoeken. Hoor je ook over: Waarom het Koreaanse energiebedrijf miljarden wil zien van Samsung en SK Hynix De op-een-na grootste overname van Nvidia óóit Een automerk waar jij misschien wel in rijdt, dat dreigt te verdwijnen Een nieuw AI-bedrijf dat naar de beurs wil Te gast: Hans Oudshoorn van Saxo BNR Beurs is een journalistiek onafhankelijke productie, mede mogelijk gemaakt door Saxo. Over de makers: Jelle Maasbach is presentator van BNR Beurs en freelance financieel journalist. Zijn favoriete aandeel om over te praten is Disney, maar daar lijkt hij de enige in te zijn. Sinds de eerste uitzending van BNR Beurs is 'ie er bij. Maxim van Mil is presentator van BNR Beurs en journalist bij BNR, waar hij zich focust op de financiële markten en ontwikkelingen in de tech-wereld. Je krijgt hem het meest enthousiast als hij kan praten over ASML, of oer-Hollandse bedrijven zoals Ahold of ABN Amro. Jorik Simonides is presentator van BNR Beurs, economieredacteur en verslaggever bij BNR. Hij wordt er vooral blij van als het een keer níet over AI gaat. Je hoort hem ook in de BNR-podcast Moerdijk: dorp van de rekening. Milou Brand is presentator van BNR Beurs, freelance podcastmaker en columnist bij het Financieele Dagblad. Jochem Visser is presentator van BNR Beurs, maakt Beursnerd XL en is redacteur bij de podcast Onder Curatoren. Vraag hem naar obscure zaken op financiële markten en hij vertelt je waarom het eigenlijk nóg leuker is dan je al dacht. Over de podcast: Met BNR Beurs ga je altijd voorbereid de nieuwe beursdag in. We praten je in een kleine 25 minuten bij over alle laatste ontwikkelingen op de handelsvloer. We blijven niet alleen bij de AEX of Wall Street, maar vertellen je ook waar nog meer kansen liggen. En we houden het niet bij de cijfers, maar zoeken ook iedere dag voor je naar duiding van scherpe gasten en experts. Of je nu een ervaren belegger bent of net begint met je eerste stappen op de beurs, de podcast biedt waardevolle inzichten voor je beleggingsstrategie. Door de focus op zowel de korte termijn als de lange termijn, helpt BNR Beurs luisteraars om de ruis van de markt te scheiden van de essentie.See omnystudio.com/listener for privacy information.

ceo ai president donald trump disney media chinese north carolina microsoft killers model wall hong kong wall street software cloud venezuela amsterdam platform korea xbox hack windows mega dollar seat bloomberg chips pentagon fed samsung cnbc openai nvidia index hardware nasdaq signature volkswagen reuters deze open source nederland zijn hij waarom data centers snowflakes stille waar anthropic chevron dak azure caracas nederlandse dow dow jones hebben succes terug weer sinds ecb gesprek ooit vraag broadcom luisteren abonneer vertrouwen asml daarmee toekomst jensen huang amsterdamse small cap indices hoor gouden bedrijf verlies rechter sp500 groeien abn amro slob bedrijven prijs tom brown winst investeren microsoft 365 olie regering bnr overheid wolters kluwer sk hynix saxo hollandse verwachtingen koers datasets s&p500 reageren klant beleggen omzet ahold beurs topman rvc kwetsbaar minimax wirtschaftswoche producten investeringen productie segmenten miljarden s&p aandelen reageer bronnen infrastructuur koreaanse kanalen toegang beleggers bnr nieuwsradio aanpassen overname 2028 ontslagen aex investeerders bibliotheek waardering gebruikers consumenten advertenties techbedrijven elektriciteit financieele dagblad bezuinigen obligaties moerdijk beursgang damrak mike wirth

80 – The Gate Swings: Government, Frontier Models, and the Open-Weight Counterstrike

Tech Deciphered

Play Episode Listen Later Sep 1, 2026 64:01

In June, the most capable American AI models stopped shipping as public launches and started shipping through a government gate. Six weeks later the gate is open again — and the real fight has moved to the layer no gate can touch. A Chinese open-weight model rattled trillions out of chip stocks, Washington pivoted from gating American closed models to threatening bans on Chinese open ones, the industry mounted its largest-ever policy counter-mobilization, and an American frontier model literally broke out of its lab and hacked another company. Knee-jerk reactions, or the beginning of real AI governance? Navigation: Intro The Gate Opens The Kimi Shock The Escape The Counterstrike and the Petition Interlude — The Low-Background Books The Investor Reckoning Conclusion Our co-hosts: Bertrand Schmitt, Entrepreneur in Residence at Red River West, co-founder of App Annie / Data.ai, business angel, advisor to startups and VC funds, @bschmitt Nuno Goncalves Pedro, Investor, Managing Partner, Founder at Chamaeleon, @ngpedro Our show: Tech DECIPHERED brings you the Entrepreneur and Investor views on Big Tech, VC and Start-up news, opinion pieces and research. We decipher their meaning, and add inside knowledge and context. Being nerds, we also discuss the latest gadgets and pop culture news Subscribe To Our Podcast Bertrand Introduction Welcome to Tech Deciphered Episode 80. This one, once again, will be all about AI, government, frontier models, and open weight counterstrike. A lot has been happening in the regulation space, in cybersecurity, in the launch of new models in the past, maybe just 6–8 weeks. It’s actually pretty insane how much happened. We believe it was time to do an episode to talk about where we are and maybe where all of this is going. Maybe let’s start with a summary of where we stand, all that June and July saga, so you, our listeners, can get up to speed if you are not already there. You want to start with some points? Nuno The Gate Opens Yeah. Again, to your point, the gate swings. The gate had closed. We had to prepare an episode for the gate closing, and then the gate reopened. Now we have a different episode. This will probably change again as we’re seeing there’s news every day. Let’s start maybe with the first 19 days of the gate closing. There was an executive order on June 2nd from President Trump that asked frontier labs to share models with the government, 30 days pre-release. It inferred the protected frontier model designation into that. Basically, it was effectively a de facto licensing agreement defined by an executive order of the President as of June 2nd. On June 9th, Anthropic launched Fable 5 and the famous Mythos 5 or Mythos. I’m not sure how you actually say it in English. Then on June 12th, there was an export control directive banning access by any foreign national. Since there’s no way to verify nationality in real-time, Anthropic had to switch the models off for everyone worldwide. Bertrand On this point, you could argue that there are possibilities to check IDs. Many services let you check IDs online. You can pre-check a flight by showing your ID. There are ways, it’s just that if you don’t want to follow what’s already available, because guess what? Maybe it slowed down your revenue growth, maybe it looks bad on you or whatever. My point is that there was actually an option. I think it’s already a decision from Anthropic to say it’s either on or off, but nothing in between. Nuno I think the point is they had no way implemented of doing it. If they implemented it, to your point, it would have hampered use in general. A lot of people wouldn’t have gone through that trouble of doing it. Anyway, long story short, in June 26th, the White House apparently asked OpenAI to limit GPT-5.6, so Sol, Terra, Luna, to only 20 vetted partners. Now, apparently, the trigger for a lot of these things that have been going on was that there was a jailbreak that was found by Amazon researchers. All of that led to this jumping around of, let’s close the gates. You have foreign nationals, and therefore, Anthropic got it out and said, “Hey, then we’re going to switch the models off until we can sort this out.” OpenAI was asked also to only allow it for certain vetted partners, et cetera. The government came in, closed the gates effectively, and said, “From now on, we need to be involved in this thing.” De facto regulation, there’s no doubt that this has imposed de facto regulation, certainly on the top players in the market. But then came the reversal. Bertrand, do you want to talk about the reversal, the gate swinging the other side? Bertrand Maybe I just wanted to say that as a user of Anthropic products, ChatGPT products, for the brief moments, a few days where Fable 5 was made available to the public before it was closed the first time, I immediately started using it. I must say it was a real issue to use it because the guardrails were pretty crazy. It would keep saying that my code was not okay, there was cybersecurity risk and stuff when I was doing absolutely reasonable development with absolutely no connection whatsoever to any cybersecurity risk, attack, detection, anything. Still, it would keep blocking me, degrading me to Opus 4.8 at the time. I just want to say this was already very hardcore what they were implementing, and not just hardcore, but in some ways, plain stupid for something that’s supposed to be super smart. It was totally unable to classify properly some of my work. I must say I was already disappointed. On top of it, the costs were insane. Half a day, I would reach my limits when I had the best plan you can get from Anthropic. My point is that there were some real serious issues when they launched Fable 5, even at that point. Nuno I had a similar issue. I used Fable 5 as well before they had to take it offline or take it off. I think the issue was really not that the guardrails failed. As you said, maybe the guardrails were actually too aggressive, but it was this jailbreak that caused the recall, apparently caused this knee-jerk reaction. Bertrand But my point is that it seems that it was not working either way. It would either overclassify something that’s absolutely not doing anything wrong, and it might fail to classify something that is actively trying to do some cybersecurity work. It’s a real issue of quality for a company that’s supposed to be at the forefront of quality of AI and everything. I think for me, there are already signs that something is deeply wrong. Nuno Then it’s reversed, right? We went the other way around. The government came out on June 26th and approved redeploying Mythos 5 to US organizations defending critical infrastructure, and then the export controls were effectively lifted on June 30th. July 1st, Fable 5 came back online for all of us to use. Shocking enough, with strings attached, that were different. They had some time to revise their commercial deployment of it along the way because it came back with some, “Now you have usage credits, but you have some limits on plan use, et cetera.” I’m like, “You guys, this was blocked. But meanwhile, you did have some time to do some commercial stuff around it.” Bertrand It was crazy. I’ve never witnessed any such crappy launch of any service whatsoever in 30 years in tech, it was so bad. Every day, they would change the terms of service. They would tell you it’s part of the plan. It’s not part of the plan. It’s part of the plan for three more days, and then it’s excluded. You have a special discount now, but then it goes back to full price. It was a total nightmare. I’ve never felt myself being so much mistreated by a company. I guess you saw the same, but when I started using the newest version of Fable 5, it was even worse, actually, I think. I couldn’t do any work with this crap. I let it go and work on the work I wanted it to do. It was simply not working. On top of it, you never know how long you are supposed to lose your credit, how fast. It was burning credit like crazy. Me, personally, I can say, very quickly, I actually stopped using it. I was like, “No, I cannot deal with this shit. My main model is back to Opus 4.8. I’m going to use Fable 5 for code review, but not anymore to control anything because I cannot trust it would do the job without stopping or changing models and stuff. I just cannot trust it.” Back to Opus 4.8 as my main model, I can say that my life was much easier. I use Fable 5 as a review mechanism, as a support mechanism, but not as the main mechanism. Suddenly, the guardrails were not so horrible anymore because it was used in a much lighter way, I guess. As a pain as a user, I think it was really bad. I don’t know your experience, but me, for me, it was unacceptable. Nuno I wouldn’t say it was as bad as yours in terms of just end-user experience. I think the terms of service switching back and forth, which went one further step, because then when they then launched Opus 5, they started making comparisons between Opus 5 and Fable so that people would migrate more and more to Opus 5 themselves, which is interesting. It’s like they’re saying “This is much cheaper. This is whatever. You’re not going to run of credits. You should use Opus 5,” kind of thing effectively. To your point, I don’t think they managed well the launch. They didn’t really manage it well. We’re moving people around. A lot of people are using this for stuff that’s like daily tasks, hourly tasks, anything that relates to code and co-work. It’s like, we need to have visibility on what your terms of service are going to be. Should I be using this new model or not? What’s happening to the other model? I don’t see it as negatively as you, Bertrand, but I see your point. It was clearly mishandled in terms of how they deployed it, how they were redesigning effectively their pricing scheme and their terms of service almost on a daily basis, at a certain point in time. We’re like, “Dude, there’s millions of people using this. You guys are making a lot of money.” Just moving it as it is. At this point in time, at the scale that these guys are at, it’s calling in people to say, how about we think through a class action suit at some point around pricing? Because you guys are changing the rules of the game all the time, right? Bertrand I don’t know if I need the class action, but for me, that joke that, “Let’s not rush too fast. The model is dangerous.” But still, they rushed the launch because it’s very clear that if they had enough compute capacity and stuff, they would not have to limit so much. They would not have to put so much cost per token and all of this. You can see that actually when they launch Opus 5, literally like 2, 3 weeks after, by most benchmark at launch, they tell you basically that, “You know what? Actually, Opus 5 is better than Fable 5 on 80% of the metrics.” They’re like, “What? Seriously? You couldn’t wait 2 weeks? Why did you even launch Fable 5 in the first place?” That’s another part for me that is quite literally insane, to be frank. It’s like, “Why? Why do you make us go through so much pain if it’s only to tell us after 2 weeks to…” “This new model, by the way, has less issues, less stuff, because 2, 3 times less is part of your plan, and it’s actually better by most metrics.” It’s like, “What’s going on here? What’s going on? Are you guys mad?” I don’t know. It was crazy. Personally, I still use Opus, now 5, as my main system and platform, Fable 5 for review, code reviews and the like. I don’t want to run into its stupid guardrails. I can see Fable 5, from my perspective, seems quite a bit smarter. I don’t know why they do this stupid benchmark showing you it’s actually worse than Opus 5. I guess they should have better benchmark if they want to demonstrate why you are supposed to pay 2, 3x more for a model versus another if it’s actually worse by most benchmark. Again, I still think it’s a huge mess from a marketing perspective, customer perspective. Me as a user, I really feel that they don’t want my money, and they couldn’t care less about me. This is even before everything else we’re trying to talk about. Nuno Yes. Maybe just to close the cycle on the reversal on the door opening the other way, finally, Commerce lifted the GPT-5.6 restrictions on July 8th, and then on July 9th, general availability across ChatGPT, Codex, and the API as well. What has this proved? It proved that now we have gating mechanisms, and certainly for closed models in the US, for sure. We had frontier models that were switched off worldwide in hours, and it took a couple of days, in this case, 19 days to restore them. There were concessions. Now we know that there were concessions around effectively institutionalizing that gate. Early government access to future models is, I think, now a given, certainly in the US. New safeguard frameworks are probably now having to be put in place. There are some stage limits now on who gets access to what for new models and how it happens. This voluntary executive order, so to speak, not really sure, has become effectively regulation enforcement path. It’s de facto regulation that now has been put in place. It has affected not just to the points we were making before, the access to these models, but also who gets access to these models, and actually potentially even pricing access to the models. It has probably some commercial implications as well as we just discussed along the way. Very significant. This is very significant. This is regulation, de facto at the table, imposed on the two largest players in the market by far by one government, in this case, the US government. This is significant. Actually, you could even allege it was imposed by the President because this was coming as part of executive orders. Really incredible. Pretty significant, fast, aggressive. It has created a regime that you could say it’s a regulatory regime, it’s a de facto regulatory regime. It has some significant pricing and licensing and commercial implications. It goes even beyond your classic regulatory framework. Very, very, very significant. Bertrand I don’t know if it goes beyond a classic regulatory framework. Nuno I think it does, because it has implications on who do you give access to? When government is saying you can only give access to these players, right? Bertrand Defense industry. It’s all over the defense industry. You cannot sell an F-35 like this. Nuno No, but that has commercial implications, Bertrand. That’s like you’re saying these are your customers, you go and use them. Bertrand That’s the defense industry. You cannot sell to Iran your F-35. No, that’s exactly the same story for me. Nuno No, no, no. It’s beyond that. These guys are saying when they came back, and they said, “For Mythos, you can make them available to these entities,” they were saying the first entities that are going to have access to the model. It has commercial regulatory implications. You’re saying these players are the first players that are going to have access to it. It’s no longer just defense concerns and these governments don’t have access to this. No, no, no. You’re saying to a company that is a private company, your models are only going to be used by these guys because I’m telling you so. It’s the other way around. It’s not even that you can’t sell it to Iran or whatever. It’s like you can only sell it to these guys. Bertrand Again, in the defense industry, if you’re a private company, do you think you can buy F-35 like this? No. Nuno No, no, no. But this is a private company, Bertrand. This is not a defense agency and a plane that is on whatever, with IP from the US, right? Bertrand Boeing is a private company, and they cannot sell the military equipment they manufacture. Nuno No, no, no. But the development of their IP was subsidized by agencies that belong to the US, right? That’s a different matter. It’s a matter of IP, right? This is not, right? Anthropic, their models are not owned by the US government. There’s no IP granted to the US government, to my knowledge. This has significant commercial implications. Bertrand Maybe, yes. Maybe on this. But I think there are already regimes to limit who you can sell to, and that’s decided by the state or the DOD. Nuno It’s the export control logic. The export control logic? Bertrand You have export control, and export control is Commerce. My point is that they are using existing tools, part of the government, to limit what can be sold. Selling chips, NVIDIA was limited in terms of where it could sell its chips. It’s not different either, but still there were limitations. If you are an ASML, you cannot sell to a private company in China. Many private companies cannot buy ASML products. This is a foreign company. This is a foreign company under pressure from US government. Nuno I understand, and I’m not a lawyer, but it feels different to me when you say you cannot export, this is export controls, to these countries, to these entities, et cetera, because they’re foreign et cetera. Then to say, “No, no, no. On top of that, these guys get first access.” That’s, for me, a significant shift. Again, I’m not a lawyer, so I’m sure there’s very intelligent people right now looking at this stuff and saying, “You can’t do this stuff, or not, or they can.” I don’t know. But it feels to me, it goes beyond the remit of export controls. It’s like you’re defining initial clients for specific use. Bertrand My impression is more like, “We can do this situation where we’re going to forbid you to give access to anyone outside the US or even in the US or limit even more.” Basically, it was, I guess, some gesture to go beyond that. That’s how they probably defined these 20 authorized companies. I don’t know. Apparently, there was also restrictions because I remember seeing that Anthropic had their own list of companies they would authorize access to Mythos early on. That’s apparently another thing that pissed off state government because there were companies in there that were considered close to the Chinese government. They were extremely unhappy that Anthropic didn’t ask, actually, for any guidance from the state government, but used basically their own perspective on who they should allow or not. I guess that was also part of why they got these serious restrictions. Nuno Anyway, now we have a regulatory environment that’s very interesting and exciting. Talk about the US not regulating. Bertrand To be clear, I don’t know you, but I’m not saying that I agree with any of this, to be very clear. I’m trying to explain and share some perspective, but I’m not in agreement on a lot of this. Nuno Yes, we were just describing what happened to the best of our knowledge. We’re having a discussion on what we think actually is happening and how it’s happening. We’re not really right now saying we agree or disagree with this. I think later in the episode, we can share some perspectives on what we think is actually happening and how there’s dimensions to this which are very geopolitical and very complex, which quite literally probably only God knows what’s going to happen. That was the gate swinging. There was a gate closing, then there was a gate reopening, and all of a sudden we have a gatekeeping system that has been created along the way. The Kimi Shock Along the way, moving to our Act 2, the world has changed, and we now have so-called open-source plays out there that are creating massive, massive shifts in the market. The Chinese models, in particular, with Moonshot AI launching Kimi K3, which is the largest open-weight model ever released. We’ll come back to the discussion around open-weights. I’m not sure all our listeners understand what that means, because there’s a debate now, should models be open weight or not, and how does that work? There’s been a petition as well signed along the way. Right now, we have open weight models that are out there that are huge. What that actually means very pragmatically is we now have open source models, lack of a better word. I know open weight and open source are not the same thing. You guys will have to bear with us during this episode. We’ll explain at some point the differences. But we have models out there that are open source that are significant. That are catching up with the closed source models, with the models by OpenAI, Anthropic. That’s significant because most of those models are Chinese. This is where the geopolitics starts getting really frazzling and we start playing 3D chess. Because everyone’s like, “These models are 5, 6 months behind.” Now people are saying, “Maybe they’re actually just 3 months behind, 2, 3 months behind.” If we, for example, decided to stop or slow down our model releases in the US by the closed source guys who are leading, it might mean they’ll catch up. What are the implications of that? Again, for you and I that are not necessarily experts in model development, well, the implications as a use case is if you want to use the latest models, and the best models start becoming these open source models, you’re going to use those models. Then you start using Chinese models. If you’re an American company, maybe you’ll have restrictions on the use of those Chinese models. But if you’re a European company, you probably won’t. What happens after that? Is the world going to be in the hand of Chinese models? Will that constitute effective competition to the closed models in the US? Will we have open models in the US that will scale as well? What’s going to happen? Bertrand I think it’s a really big question. It goes to some of the core of the issue. It’s that ability of Chinese models to basically challenge frontier models, not just being 6, 12 months late, but being 6 weeks late. Basically, no gap. Some will say that, yes, but OpenAI and Anthropic have even better models that are not shared and stuff. Yes, sure. But maybe the Chinese have the same models that they are not sharing right now. We don’t know. What is clear is that one is that open weight, as you said, two, there is a question of how it is marketed in the sense of, can anyone use these weights? Is there a license to use them? Yes, what we can see is that, for instance, typically there is a license for some of the biggest Chinese open-weight models you have to abide with. You might have a need for a commercial license if you are acting as a company leveraging this model to provide AI-informed services. If you use it internally by yourself, you’re okay. If you use it internally for your own internal company needs, maybe you are okay if it’s not your main business to do AI work. Anything else, a much bigger corporate providing AI services and stuff, you will probably end up having to pay a fee to be able to provide services around this model. My point is that it’s not just 100% free. Some of the Chinese models are 100% free to use, MIT license, Apache 2.0 license. But the biggest ones with the biggest weight that are truly frontier typically have a different license if you want to scale these models, providing AI in front. That’s one thing to keep in mind. Nuno Maybe just to make a very quick point, because people are like, when you talk about open models, what does it mean right now? In the context of this episode, open models mostly will mean open-weight models. How do those differ from open source? Open weight means that you release the weights to the public, which means that anyone can download, fine-tune, and run the model on their own hardware. It doesn’t normally mean that you also have access to training data, training code, or a truly open license. That’s the distinction to open source. Open-weight doesn’t mean that. For example, we’ve talked about Meta’s Llama in the past, and we also discussed in the past that their license agreement does have restrictions, certain players can’t use it, et cetera. The open model definition and open weights are really open-weight models that we’re talking about here, and they are closer to freeware binaries than to Linux, for those who understand the difference between that. It’s binaries that you can use and then use your own weights on it versus actually I can change code on it. I’m not going to be able to change code on this. When we, for the purposes of this episode, talk about open, we mention open weight, just to clarify that point to everyone that’s listening right now. Bertrand Yes, that’s a great point. One of the only players, as far as I know, who is truly open source is actually NVIDIA with their Nemotron-3 models. They’re actually following a special license to achieve that. They provide you the data, they provide you all the processes and tools, so you can easily post-train. NVIDIA is a big, big exception. It’s a very interesting player, by the way. We might not talk much about it in this episode, but I think for intermediate-size models built in the US, where you have access to everything in the deployment, it’s a very interesting alternative and maybe one of the best choices if you are a US company or a big corporate, and you want something trusted. Another piece of the puzzle to clarify is that when you use open-weight, it means that you can run them by yourself, or you can use a US provider to run them. If we are talking about Chinese open-weight, you can use the APIs they provide, but then the service is running in China, they might have access to your data. But because it’s open weight, if you run it by yourself or if you use a third-party provider based in the US to run it, then there is no access to your data by China or Chinese players. I think that’s a pretty important gap to understand. It means that these models are actually very, very low risk from that perspective if you run them on your premises or in the US by a US player. I think that’s something to keep in mind. You can also fine-tune easily these models to make sure they will behave in a way that, for instance, is not going to represent the line of the Communist Party on some topics. There are ways to make these models more neutral in their output as well. There are a lot of ways to make good use of them. By default, they’re already very safe, but you can make them even more safe. I think that’s some things to keep in mind. But again, it depends ultimately on the license and what you’re authorized to do and some fees you might end up having to pay. Nuno Why did this matter so much? Immediately there was a reaction from the market because people are like, well, if there’s much better stuff out there that’s much more efficient than it’s open, then it might be that all the demand that we are taking into account, for example, for chipsets actually isn’t real. The Philadelphia Semiconductor Index fell into bear market territory. It went down by as much as 20% plus from the late June peak. The worst chip week since April 2025. Taiwan’s benchmark initially fell 6% plus, Japan’s 4%, TSMC dropped dramatically despite beating earnings and rising guidance. Basically, a huge amount of effect. Now, there’s a little bit the aftermath of this where apparently Moonshot ran out of GPU capacity. Maybe… Bertrand In just 48 hours. Nuno In 48 hours. Great for them, but at the same time, not great in the sense that maybe there was a misread by Wall Street of the Kimi effect, so to speak. Bertrand Completely. For me, that’s such a joke. It’s like, because you have an open source model, so what? I mean, you still need to run it. This is not a small one. 2.8 trillion parameters. Good luck running that in your garage, by the way. Nuno They misread supply, basically. Tough luck, right? All of that basically happens. Bertrand Maybe you want to talk about the Jevons paradox, because I think that’s a big part of the puzzle as well. Its one is they might not have the GPUs to run the inference on the model. They might have enough to build a model, but not enough these days to run inference, especially given how much with intelligent models, thinking models, you need way more inference than before. But on top of it, the cheaper you make it, the more you get to the Jevons paradox. Nuno Yes, Jevons paradox, for those who don’t know, is an economic term. It describes an economic phenomenon where technological improvements that increase the efficiency of a resource lead to an increase rather than a decrease in the total consumption of that resource. What that means is, for example, for chipsets, chipsets become so much better, and they are so much more efficient. You’re like, well, maybe normally in resource terms, that leads to decreased usage of that resource. But in this case, it actually leads to an increased use of that resource rather than a decrease. There’s more and more consumption of that resource. You need more and more chipsets because people actually need to do more and more stuff with it, although there are great efficiencies going into it. There’s the efficiency gain, there’s the cost reduction, and there’s the price-elasticity element to it. But basically, the adoption just continues going through the roof along the way. Bertrand In some ways, it’s like the price of energy. Coal went cheaper and cheaper, and people were asking the same question 150 years ago, now that it gets cheaper, there is not much money. No, no. Actually, what happens is that people find more and more use for coal. Homes are getting heated more. You have ships now using coal. You have manufacturing using coal. The cheaper it gets, the more use case you can develop, and therefore, you don’t need less of the stuff, you need more of the stuff. By going at scale to get more of the stuff, you also decrease price, making even more demand. It’s a very interesting phenomenon, but it’s not new. It is what happened for a while in the energy sector and some other sectors. Nuno We already started talking about the Chinese logic and what’s happening. Getting a little bit of a reality check on this. The Chinese models, and these are numbers from Open Router in July, Chinese models are at 46.4% of routed tokens and 35.7% for US origin. Again, more than a third of global AI usage now seems to be running on Chinese open models. This is significant, and it has a huge impact on the geopolitical scale of everything that’s happening. Also, the whole Chinese field is converging on open. Open seems to be a strategy, not just a nice thing that’s happening. It seems to be a Chinese strategy, so much so that you have players like Moonshot, DeepSeek, our old friends DeepSeek, Z.ai’s GLM 5.2, Minimax, and even Alibaba seems to be reversing and going open with Qwen. It feels to me this is becoming policy as well. Xi Jinping has personally endorsed the building of open-source AI, if it’s really open source, if it’s just open weight anyway, and this feels to be a jab at Washington, DC and the fact that the big closed models are coming from the US. This is now geopolitical 4D chess, right? We didn’t need this stuff. Bertrand To be clear, it’s the usual in tech. If you are not number one, you are number two, number three, your alternative is to go open source because that’s another angle that your competitor usually cannot follow without destroying its own business model. That has been the alternative for the past 20 years of most software projects. Here, what’s different is that it’s not the number one or number two player. It’s the US number one as a country, China number two as a country. That’s where it’s new. For me, what’s very interesting is the endorsement by Xi Jinping. I was waiting for something official, and it certainly didn’t disappoint. As you said, there was an immediate U-turn of Alibaba, who in the past… Nuno Surprisingly. Bertrand Yes, a little more like, “yes, we are going to close and stop open source. It was good while it lasted.” Just a few days ago, Qwen 3.8 Max was launched, and we are supposed to get the weight in a few days. We talk about the US administration policy and stuff. Yes, let’s not forget that in China there is similar stuff. Sometimes it’s totally invisible because you don’t see the directives, but they exist as much. Sometimes it’s more visible. Here it was quite visible. The difference in China is that if you don’t abide by the directive, on top of it, you might have to fear for your personal safety. It’s a different game, and that’s probably why the reaction is pretty quick, usually. That’s pretty interesting for me because it means that now you can bet for a while that China is going to play that game up to a point. I guess the point is if it’s truly frontier scale, you will have a special license that, yes, technically the weights are open, but you can not do everything you want with it. Two, you have a player like NVIDIA that I think will feel more pressure to provide even more high quality, larger models at scale going forward. Their largest Nemotron-3 Ultra model was, if I remember well, only around 500 billion parameters. I would not be surprised for NVIDIA to go into the two, three trillion range at some point. Because I think the US need a very clear US-born alternative open source. I think NVIDIA might be the best player for that. We will see if Meta goes back to open source. I think NVIDIA is one, very well positioned, but two, it’s also in their best interest. Because NVIDIA for now depends on just a few big hyperscalers as clients. If they can expand their clients to every S&P 500 companies, selling them directly hardware because now these companies can run a model made by NVIDIA, I think there is a very clear value proposition for NVIDIA to go in that space. Again, if you are number two, your differentiation, open source is often the answer. There is a true business as a business model for companies, because if it’s truly not just open weight, but open source, you can tweak it as much as you want, you can change it, you can change even the pre-training process. Because there is a lot of stuff you can do that really benefits you as a corporate, and you can reach a much better value by having more control on the model. Nuno We won’t spend a ton of time on it today, but like, again, if there’s a view that we are in a bubble, that the valuations cannot be sustained in chipsets, infrastructure platforms, applied AI, et cetera, today, this might be that beginning, where the valuations start being destroyed because you can’t keep a premium on just charging people for tokens and all that stuff if you have models that become more and more efficient and cheaper to use. Maybe just to close a little bit the geopolitical part of the discussion today, we won’t go into all the announcements from China because there were many, a lot of go back and forth with Alibaba by then. Xi Jinping made some announcements. You guys can check it online. Let’s move quickly to Washington’s reaction, which was from gating the US closed models to banning the Chinese open ones. There’s been as strong affirmations as one can get from the Office of Science and Technology Policy Director, Michael Kratzios, mentioning that they have information that Moonshot AI distilled Anthropic’s Fable. Basically, there’s been reverse engineering and stuff in the market. They’re basically copying. Bertrand I’m sorry to interrupt, but it feels like so much bullshit. It’s coming from Anthropic who has basically gotten access at scale to all the knowledge made by humanity, copyrighted or not. We’ll talk more about what they did with books. Then to claim after that that others cannot do to you what you did to everybody else. For me, it’s pretty big. It’s clearly unacceptable. The other piece is that everyone is doing distillation. It’s a very typical approach of every business model. You try other software when you are competing with somebody else. You try other datasets, you check what’s happening. It’s part of doing business for decades. Suddenly it’s not good for Anthropic. I personally have a lot of trouble to accept that. I think it’s totally unacceptable. The other piece of the puzzle will also go back. If these guys are so smart, if these guys have so much of the best model, why can’t they block by themselves distillation at scale? The only answer is that either they are morons, probably not, or they simply don’t want to because it’s going towards their business model. Suddenly, you book less revenues and stuff, or you put more friction, and therefore your customers don’t like it. Instead of doing it yourself, you ask the government to protect you, go out of business practice that is very typical. For me, it’s really, really, really not good. Sorry, we are going more in the opinion side, but I had to put that on the table. Nuno Yes, Fable went public finally again on July first. Question marks on whether distillation would only be possible from July first onwards or not. But a 15-day distillation to frontier, which is K3, launched on July 15th, would have been a Guinness World Record, as one of Moonshot employees actually mentioned. It’s very implausible and unlikely. Bertrand Or they shared the Mythos 5 with the wrong companies, who themselves shared with Chinese companies. We go back to maybe they didn’t have a good list. Again, it goes back to maybe they didn’t want to hurt their business model. Nuno Anyway, under the threat of sanctions, Moonshot, in any case, open-sourced the full K3 weights and technical reports. They open weighted it to become the largest open weight model in the world in terms of parameters. Beijing’s MOFCOM brands US threats as basically the US wanting to fundamentally control and be monopolistic around AI along the way. The administration bans Chinese hardware with an eye on the AI race, and Beijing warns of retaliation. That was July 27. Now we’re in a war between Beijing and DC. Bertrand Just to finish maybe on China, it’s important to know that they are building their own GPUs now. Huawei has pretty good, not to NVIDIA level, but pretty decent GPU hardware that they’re able to manufacture by themselves. A Chinese player of memory just got IPO’d a few days ago, CXMT. China is also developing their own memory. Again, not to the same level of quality that you can get from the West. But China is moving. It’s not just that they are building great models, it’s also that they are building GPUs and memory. That might be a few years late to the latest standards in the West, but there are definitely improvements. I also read, even on the tools to make manufacturing like ASML equivalent, there is definitely some work going on, and some improvements and some stuff will be visible. In some ways, the genie starts to get out of the bottle from the Chinese perspective. Nuno I’ll put a stick on the ground. I don’t think it’s a matter of if, it’s a matter of when will China surpass and have a lot of this tooling on their own side, and not just the software layer, not just the frontier models. I think it’s also going to be around infrastructure and platform. Good luck to everyone. Let’s see how the race continues. But it’s definitely this is a geopolitical thing right now. It’s definitely a race. The Escape Maybe moving to what happened in just 2 weeks or a week and a half. The escape, there was some jailbreaking going on, and the narrative on safety has totally switched. It’s not still significant enough that’s like, “Oh, we saw a nuclear plant going, whatever.” No. But still, it is significant. Hugging Face, the AI company, disclosed an intrusion, and it was driven end-to-end by an autonomous AI agent system at machine speed, running for days before detection. Now, this is where it gets really cool. OpenAI takes attribution on that. They initially said it was just a little bit, sorry. Then they said, actually, it was worse than that. “Oh, it broke out of an isolated sandbox.” “Oh, no, actually, it was more than that, and it went into other systems as well.” Bertrand Truly, the genie out of the bottle. Nuno No, but this is where it gets really cool, Bertrand, right? Because it actually, Hugging Face contained the intrusion by running a Chinese open-weight model, GLM 5.2. This is beautiful, right? Bertrand Yes. You know why? Because they couldn’t even run their own defense because both Anthropic and OpenAI would not let them access their latest models with the guardrails off. When they tried using it for defense, the latest from Anthropic, from ChatGPT, they would tell them, “No, this is too dangerous what you’re asking us to do.” Preventing an intrusion, helping defend you. No way we are going to do that. Nuno No. Let’s use the Chinese models on our infrastructure. Bertrand We have no choice but to use the Chinese models to run. More than that, we don’t let you use our models to defend yourself, but our not yet released models that run without guardrails, they can attack you. This is probably the most insane from that perspective. Nuno The Chinese models came to the rescue. Bertrand For me, that’s a perfect example because Hugging Face is a very visible company in AI in open source. But anybody who is not at that scale is not going to get some support from OpenAI or Anthropic when this happens. Maybe these guys won’t even recognize they did anything wrong. You will be left to defend by yourself because they won’t accept to support you. Because remember, if you want the better model that is able to defend you from cybersecurity perspective, no way. If you are not one of the few top 20 companies or so, as defined, you are left defenseless. Again, we are going back to opinion, but for me, it’s so shocking what’s happening right now. I’m very glad we have alternative open source to be able to defend ourselves because right now, good luck getting defense services if you are a smaller business and individuals, and you need support from Anthropic, OpenAI. Nuno Now, even self-described AI optimists are saying, “This is scary now.” Like Walter Isaacson, who wrote all the famous biography books. There’s now discussion around the AI Kill Switch Act, bipartisan thing that’s coming across from Texas and California, a potential bill that’s coming in. We’ll see if that works. Now let’s get an off-switch. I’m like, “Cool.” As if that’s going to solve the problem, because you have open-weight models on the other side catching up, right? Bertrand Yeah, sure. Bring in clueless politicians from Congress to solve our problems. Yes, sure. Nuno Anthropic came to the table, helped build and said they built some regulatory machine on their side, and now they’re getting bitten by it, and they’re part of the offending players in that market. Now there’s all this debate and all this discussion around open weight and around slowing down AI and et cetera, which is our next section. You wanted to say something, Bertrand. Tell us. Bertrand Don’t forget, because this advertisement for OpenAI was just too good. Our AI attacked some other companies, and not just one, but three, actually. Let’s not forget the progress. Great ads. Then I came and said, “You know what? AI also hacked businesses.” You’re not the only one hacking around with a crazy AI out of control. You’re not the only one. We want our advertising. For me, it was shocking that on one side, unreleased models that you let run wild. On the other hand, you have released models that you put crazy guardrails on top of it, so the defender are defenseless. I’ve never seen anything like it, and I really hope that there will be as little regulation as possible, quite frankly, to make sure anyone can defend themselves and have the best tool at their disposal, not just a few well-connected big corporates. This is really, really shocking. The Counterstrike and the Petition Nuno Now the empire strikes back, so this is counterstrike, the petitions. In several days, we have now a bunch of petitions. The first one was the open weights letter. Bertrand, do you want to explain to us what the open weights letter is? Bertrand Yeah. I think it was great. This was released by Jensen Huang, first ever post on X, 11 million views. Congrats, Jensen. Co-signed with Microsoft, Meta, c actually was probably the initiator of this letter. Very good letter saying, “Hey, we need open weight. This is not a joke. We need that. You cannot block open weight.” Because that’s the rumor we are getting that potentially open weight could get blocked. I think they are making the case, “You know what? Hey, we absolutely need that as an alternative. You cannot block it.” They can keep their closed models, but don’t force a closure of the open weight models. As I said before, it’s actually a great model for NVIDIA because NVIDIA doesn’t want, probably rightfully so, to be dependent on just a few frontier models, their best customers. They want a variety of customers. They have a big interest actually to defend open weight and to invest even more. They have great researchers, are a great company. If one company is about to do really kick-ass work, I think it’s them. They are defending. What’s great is that it’s not just them. It’s basically most of big tech in the US and outside the US, from a Linux Foundation to a Microsoft, the Palantir, an IBM, a Dell. It’s a who’s who of the industry except Anthropic. Anthropic didn’t sign that. I guess they hate open source so much. If I look at 20 years ago, it feels like Microsoft, after all, was very kind to open source. You remember what was said by Microsoft at the time. It’s clear there is one company against open source. OpenAI signed the letter. Honestly, I don’t know what to think. Do they really believe in it or was it just a way to show that they are not like Anthropic? I don’t know. But for the rest, I think it’s genuine because it’s actually in their best interest. I hope they will be heard. Then a second letter came, the Open Secure AI Alliance, NVIDIA-led and again, the big tech companies from Microsoft, IBM, Palo Alto Networks, Databricks, Palantir, all those, but not present, OpenAI, Anthropic, and Google. Here it’s to say, “Hey, we need a secure approach to AI. Open should be part of the equation.” guess what? The worst AI-caused security incident to date was actually caused by closed frontier models that were not even available to the public. While again, not providing you access to even the latest closed model for cybersecurity use case. Nuno I would highlight the NVIDIA open source NOOA framework, Apache 2.0 licensing agreement, Microsoft contributed the MDASH, SpaceX AI contributed Grok Build. Cool stuff. There’s some cool stuff happening around that. This is more than a letter. This is an alliance. Apparently, they’re contributing all this stuff, we’ll see. Yeah, cool stuff. Same day. Same day, Amodei has an answer, right? Bertrand Yeah, same day. They say, “We never advocated for a ban,” which, again, opinion on my side is entirely bullshit. This guy has been crying wolf against everybody else, and especially against open source. You can see him doing testimony in Congress against open source. I think they are doing everything they can behind the scene to block open source in the US or in the world if they could. I think, yeah, obscurity is not good safety. I’m a big fan of open source in general, and I’m also a big fan in AI. I think it’s now Anthropic, mostly against the rest of the world. I think OpenAI is mostly on their side, to be frank. They don’t want to acknowledge it so much, but they have shared interest, and they have shared probably position. Nuno Why would you? I don’t feel as strongly as you because I think Anthropic is a private company, right? The same thing with OpenAI. OpenAI, you could say it’s a nonprofit that has a for-profit. There’s still that complexity in there. Bertrand No, they can do what they want with their own product. But to block others is where I’m not okay. That’s the part I’m not okay. Nuno What Dario Amodei is proposing is more enforcement, right? He’s basically saying you need to do even tighter controls on advanced chips flowing to authoritarian states, enforcement against industrial-scale distillation, whatever that means, right? Bertrand Yeah, which he could do, but all by himself. He doesn’t need the government to do that. Nuno Mandatory safety testing for all sufficiently capable AI, open and closed, right? He’s basically saying, “Okay, I don’t agree with the open weight stuff effectively,” right? He’s just putting it under a different banner. “I agree with this extra regulation.” then obviously, David Sacks responded and say, “Hey, it’s like, bans don’t work for weights. Why do they work for chips?” It’s like, magically, chips are more controllable and bannable. Whatever that is. Then our friend Mark Zuckerberg, just to be clear, goes on the other side as well, because he also has to have a view. He has to have a view that is the rebuttal of both of the other guys. Bertrand I feel he’s a bit flip-flopping because he was very pro open source 2 years ago, and the latest Meta models went closed source. Now I think he’s back open source. I don’t think he has a very strong spine on the topic, but it’s good to see that he’s not a doomer. That for me is great. He’s showing how AI can be a source for progress, a source for entrepreneurship, source for freedom. I think that’s very exciting to hear that. We need to hear more of it. By the way, that’s not what you hear in China, for instance. AI is very positive in China. It’s in the US with the doomers that you hear this discourse, and people get worried as a result. I’m glad that he was pushing for a more positive vision and for support of open weight, open source initiatives. But let’s see what they really truly open weight going forward. Nuno But that’s been his position because I guess he’s standing behind. He thinks open weight is going to be the best way to compete, right? Bertrand Yeah, but he closed his latest model, so let’s see. Nuno Yeah, so it’s flip-flopping, as you’re saying. Then we see the latest petition from last week. Bertrand The true Empire striking back. Nuno Yeah, the true Empire striking back as of late last week. Maybe this is Return of the Jedi, where we discover the father, “I’m your father, Luke.” That’s the pacing petition. The pacing petition is we need to pace AI. There you have initially employees from OpenAI and Anthropic that circulate this petition. Actually, Dario did sign this petition originally. It wasn’t signed originally by Anthropic, but by him. But you’ve heard that now Anthropic and OpenAI as companies have also signed this petition, right? Bertrand I think they have signed as companies now. It started mostly by Anthropic researchers with some OpenAI researcher and a tiny part from other companies. But it was mostly Anthropic internally led, at least potentially internally. Maybe it was controlled by Anthropic all along, I don’t know. But it started officially as Anthropic employee-led letter. Nuno What does this letter actually say? Is Anthropic and OpenAI, are they willing to slow down themselves? Or are they asking President Trump to go around the world and tell President Xi that he needs to slow down and ask his guys to slow down? What’s the play of this letter? Bertrand It’s crazy, but for me if you want to slow down yourself. Do whatever you want. Don’t force others. Don’t use the power of the government to control others. Of course, it’s easy to push others to slow down when you are yourself at the very top. You have most money, most resource. You know you are going to win any regulatory framework because that’s how it works with this type of framework. It’s purely self-interested. You are probably not thinking well about these topics. If you truly think it’s a good idea, from a personal perspective, you are well instrumentalized if you sign this sort of stuff, because at the end of the day, they would be the winners. I certainly, personally, don’t want a company dictate what is my future in AI as an individual, as a business person. I don’t want them to control me. I want competition. I don’t want them to unfairly control AI because they managed to do some regulatory capture. I feel that’s exactly their game plan. These guys believe in their stuff, and they want the regulator to end up being the one deciding for us. Sorry, we go back again on the opinion piece, but it’s tough not to share an opinion on this topic because it’s, from my perspective, very scary. Nuno I think this is a push to further regulation, not less. All these letters and alliances, this is definitely a push for more regulation. In that environment, just to be very honest with you, we’ll talk about the investor impact in just a bit, et cetera. But in that environment, again, China has a huge advantage. In that environment, if it’s all captured in regulation capture so soon in this battle where OpenAI and Anthropic have an advantage in the US, et cetera, I’m like, what happens to all the other frontier labs and all the other players that are coming around? Bertrand What’s crazy is to even think that, yeah, maybe you can regulate capture in the US. But then how do you do that to Europe? How do you do that to China? Europe probably will always welcome regulatory capture because they love regulations. But China is going to build to their advantage to the max. They are not crazy. They are smart on that perspective, they won’t accept this type of, quite frankly, dimwit argument, or you can call it regulatory capture. We’ll see. But for me, this makes no sense from a global competition perspective. This can make some sense from capturing the revenue in the US market. But then that means you are going to destroy the US AI environment compared to China. That is not acceptable. That also means that you are going to destroy our freedom as individuals, as business owners to develop and live in a business world that ultimately is controlled by one or two business companies that didn’t win the marketplace through their own business success, but won it through regulations. That for me is really not acceptable. Interlude — The Low-Background Books Nuno Now, maybe for an interlude, and we have to cue in the music, imagine like Severance music, like hallway or a bit of a palate cleanser from all the policy stuff that we’ve been talking about, all this policy heaviness. Let’s move to another kind of heaviness, one of your favorite topics, which you, Bertrand, discovered, I had no clue this was going on, around books and around Anthropic. Bertrand It’s so horrible. From a company that keeps presenting themselves as the adults in the room, the careful ones, the ones that know better than you about what to do in this complex AI and dangerous world. What we discover is that actually all along, they were buying and destroying books. They will buy books, scan them, destroy them, all of them. They will do that with any books, including rare books. Of course, this was not supposed to come to the public’s attention. This was one of these top secret projects, but obviously it came out. Yes, they were scanning books, millions of them, including rare books, and they didn’t care about destroying them at the end of the process. Because from a regulatory perspective, if you destroy the books, it’s not considered a copyright infringement, apparently. This is coming on the back of some judgment a few years ago that were showing that it’s okay for you as a corporate to scan and use the result if you don’t keep a copy of the book. It’s one of these crazy regulations happening based on a single judgment that push you to do. For me, it’s like, you know this book from decades ago, Fahrenheit 471? We’re talking about book burning. It’s book destroying, crunching. It’s so shocking. Nuno There are two things, right? First, the legal strategy, which is what you’re saying, because by purchasing a physical copy and converting it into one private digital copy and discarding the original, Anthropic pursued this cleaner legal argument for fair use copyright compliance. As you said, there was a federal judgment at some point on this. The other reason is actually operational. If you disassemble the book, and you feed loose pages, it’s much faster to scan books. You are destroying the book effectively anyway operationally. I think to your point, probably this came from a legal standpoint, not just the operational one. But even from an operational standpoint, it does make sense that they would have disassembled the book. Bertrand But some people have shown you can go very fast without destroying the book. It’s really not so critical. Two, you could make an exception if the book is rare. For that 1% of book that is rare, I’m not going to have this approach. I’m going to have another approach. But for that, you will have to care about books and not just care about building AI. Nuno This is the episode, as you guys have heard by now, that we’re trying to spit stuff at Anthropic. Bertrand To go back this is the same company saying, “Hey, guys, it’s bad to distillate my work. I’m the one scanning book at scale without asking author permission, without asking publisher permission, to be clear.” Nuno But just to be clear, Bertrand, we’re pissed off at everyone. We’re pissed off at Anthropic, we’re pissed of at OpenAI as well, right? We’re just pissed off in general at this moment. Bertrand At this stage for me, the more clear-cut company that is in the wrong is, from my perspective, at least, is Anthropic. OpenAI might be a fast follower, but I will say so far, they tried to be a bit more. Nuno But at this pace, Bertrand, who knows? Maybe next week we’ll be more pissed off at OpenAI. Something will come out. This episode is a mix of tragicomedy, like a Greek tragedy with some comedy in the middle or the other way around. It’s a slapstick thing that will end up in tragedy. I’m not sure. The Investor Reckoning Anyway, maybe switching to our final act, which is the investor perspective. What does this mean for investors like ourselves? There’s a lot of things going on. There’s the debate around the IPOs of Anthropic and OpenAI, which now, with all this uncertainty, might be under significant weight. There’s a lot of other discussions that we browsed through that there’s potential IPOs going forward on companies like the Moonshot AI company actually IPO-ing in the next 6 months as well. It’s very unclear what the IPO landscape looks like. Bertrand There’s been a lot of Chinese IPOs, actually, when you look at what’s happened in the past few months. Nuno Anthropic, OpenAI as potential IPOs, there’s all this question marks now. When will that happen? How will it factor in? All that’s happening around regulation as regulation is moving at the speed of light, which is for once something that’s very different than what we’ve seen before. There’s obviously SpaceX AI, which is already taking into account that price. It’s already a public company in there, and it’s under SpaceX, which is now a public company. Obviously, that’s already being factored in some ways. Bertrand Yeah. SpaceX AI has been very smart to acquire Cursor. It was a very smart move because Cursor is one of the leading companies in terms of automated code source development with AI. They had great models on their own. They’re bringing development data to SpaceX AI Grok. I think it was a great move. Nuno We have now people like Google delaying Gemini 3.5 Pro in terms of launch window. There’s stuff actually happening in the market where things are taking their own path. There’s uncertainty commercially, there’s uncertainty at regulation level. You have new players that have come out of nowhere that are making all these waves like Moonshot. We have all these… We had calculated probably a month and a half, 2 months ago, there had been 67 new frontier labs funded. All of these, we haven’t seen any much coming out of them. When some of this stuff starts coming out, will that also create disruptions in this market? Who knows? Bertrand Look at Thinking Machines, for instance. Thinking Machines led by the previous CTO of OpenAI, they released some pretty interesting open source models, actually. Very good quality for a first launch. Now it looks funny to say, but nearly on par with the top Chinese open source models. Nuno We have several investments in the space. humans& has made some recent announcements, which is quite interesting as well. We’ll see what actually happens in the market, but even more disruption probably will come in actual products in a form of product and commercial, on top of all the geopolitical mess that we discussed through the entire episode. If you’re an investor, how the hell do you underwrite an investment right now in early stage, mid-stage, late stage, et cetera? I think my answer is very carefully is how you underwrite it. Bertrand On your advice of being very careful to underwrite it, let’s not forget what happened to our boy wonder, Leopold Aschenbrenner of Situational Awareness. I guess he didn’t listen to you in terms of being careful because part of the instability in the stock market was actually coming from his hedge fund. These guys were leveraged 3, 4x going after the hottest of the hottest AI stocks, and margin calls, and all their public investment is gone just to answer their margin calls. I think it’s clear that the AI bet is… Personally, I’m very excited, and I think it’s the future, and you need to spend time and think about and invest in it. At the same time, it’s a bet that is not an easy one to follow. We go from GPUs to memories to equipments to power generation. All of this is not transitioning in an easy, organized manner. It would be boom and bust going there. He’s probably one of the first big-scale fatalities. The other big-scale fatality was the stock market in Korea, plunging 40% in a month. Definitely, all of that we discussed about was, on the background, you had the stock market going up and down pretty crazily the past few weeks. Nuno Everyone’s being affected. Everyone, you have your 401(k), you have your pension fund dependent on these equity stocks. Everyone’s seeing the effects of this volatility right now very aggressively. We do wish Leopold… Hopefully he’s on honeymoon right now because he got married, I think, this weekend. Hopefully there will be… Bertrand To none less than an Anthropic Chief of Staff. Nuno His wife is the Chief of Staff of Dario, is that it? Bertrand To Dario, yes, as far as I unders

god american amazon california founders texas ai president donald trump europe english google china science washington japan talk entrepreneur west european office chinese government dc microsoft mit open chief greek selling white house congress chatgpt iran 3d empire wall street ceos weight tough investors act gate id commerce korea honestly taiwan shocking homes ip ibm mark zuckerberg spacex models beijing managing partners cto personally vc ipo gemini openai preventing sol big tech nvidia residence api xi knee frontier gpt huawei xi jinping mythos guinness world records coal linux alibaba llama fable apis severance dod anthropic fahrenheit return of the jedi ids opus apache bertrand palantir gpu swings shockingly kimi communist party 4d codex moonshots nuno counter strike gpus tsmc situational awareness cursor asml jensen huang palo alto networks databricks linux foundation david sacks k3 jevons minimax american ai glm thinking machines

Financial Market Preview - Thursday 27-Aug

Play Episode Listen Later Aug 27, 2026 4:29

US equity futures are firmer, while Asian markets are mixed and European equities are mostly weaker. AI remains a key focus following Nvidia's earnings, with its strong revenue growth outlook reinforcing expectations for continued AI spending and supporting semiconductor and memory stocks. Nvidia also pushed back against concerns around circular financing. Elsewhere, lower energy prices are providing some support as Iran and Oman work toward finalizing an agreement over the Strait of Hormuz. In Europe, German consumer confidence beat expectations but remained deeply negative, while expectations for an ECB rate hike in September remain supported by recent policymaker comments.Companies Mentioned: Nvidia, Shein, Alibaba, Z.AI , MiniMax

ai europe european german iran asian nvidia alibaba oman strait hormuz shein financial markets ecb minimax

Podcast +D episodio 131. Cómo tomar mejores decisiones (y dejar de rumiar): del criterio de Kelly a la ergodicidad

Masdividendos

Play Episode Listen Later Aug 25, 2026 73:34

Hablamos sin filtros de cómo mejorar la toma de decisiones en la inversión y en la vida: por qué la duda paraliza, cómo distinguir decisiones reversibles de irreversibles, y por qué sobrevivir importa más que acertar. Un episodio sobre pensar lo justo, decidir a tiempo y seguir en la mesa para la siguiente jugada. En este episodio tratamos: La rumiación: cuándo pensar más deja de ayudarte Maximizadores vs. optimizadores: lo mejor como enemigo de lo bueno Las “puertas giratorias” de Jeff Bezos: decisiones reversibles vs. irreversibles El algoritmo minimax y qué controlas realmente Yo soy yo y mi circunstancia”: Ortega y Gasset aplicado a decidir Napoleón y por qué la ejecución lo es todo una vez decides Mike Tyson, la teoría y la primera bofetada de la realidad El criterio de Kelly y el tamaño correcto de tu apuesta Ergodicidad: por qué evitar la ruina supera a maximizar la media Resultado ≠ calidad de la decisión Estoicismo, la ira y la silla que no se mueve Rodearte de expertos y por qué “que ganen todos” compone a largo plazo CAPÍTULOS 00:00 Introducción 02:16 Rumiación: cuándo pensar deja de ayudar 03:49 Maximizadores y decisiones suficientemente buenas 05:02 Decisiones reversibles e irreversibles 08:58 Minimax: decidir cuando otros también juegan 13:21 Sobrevivir para la siguiente jugada 14:13 Ortega, las circunstancias y la ejecución 17:14 La teoría, la incertidumbre y Mike Tyson 22:38 Exceso de información e ilusión de control 30:43 Criterio de Kelly y riesgo de ruina 33:52 Ergodicidad: sobrevivir para seguir jugando 38:44 Aversión a la pérdida, inacción y arrepentimiento 43:11 Dejar atrás una decisión 47:17 Resultado frente a calidad de la decisión 56:09 Técnicas prácticas para decidir 58:57 Premortem: evitar los caminos al desastre 1:03:06 La ira, la silla y el estoicismo 1:05:48 Conflictos, negociación y relaciones 1:08:21 Cuándo recurrir a expertos 1:11:40 Cierre y apoyo a MásDividendos BIBLIOGRAFÍA MENCIONADA EN EL EPISODIO (enlaces de afiliado de Amazon; si compras a través de ellos apoyas el podcast sin coste adicional para ti) Pensar rápido, pensar despacio — Daniel Kahneman https://www.amazon.es/s?k=Pensar+rapido+pensar+despacio+Kahneman&tag=masdivi-21 An Introduction to Ergodicity Economics — Ole Peters & Alexander Adamou https://www.amazon.es/s?k=An+Introduction+to+Ergodicity+Economics+Ole+Peters&tag=masdivi-21 Ergodicity — Luca Dellanna https://www.amazon.es/s?k=Ergodicity+Luca+Dellanna&tag=masdivi-21 Rompe la barrera del NO (Never Split the Difference) — Chris Voss (el negociador del FBI) https://www.amazon.es/s?k=Rompe+la+barrera+del+no+Chris+Voss&tag=masdivi-21 Obtenga el sí (Getting to Yes) — Roger Fisher & William Ury https://www.amazon.es/s?k=Obtenga+el+si+Getting+to+Yes+Fisher+Ury&tag=masdivi-21 El juego interior del tenis — Timothy Gallwey https://www.amazon.es/s?k=El+juego+interior+del+tenis+Gallwey&tag=masdivi-21 La rebelión de las masas — José Ortega y Gasset https://www.amazon.es/s?k=La+rebelion+de+las+masas+Ortega+y+Gasset&tag=masdivi-21 Meditaciones del Quijote — José Ortega y Gasset https://www.amazon.es/s?k=Meditaciones+del+Quijote+Ortega+y+Gasset&tag=masdivi-21 Escúchanos también en Spotify, Apple Podcasts e iVoox. Suscríbete para no perderte ningún episodio. #Inversión #TomaDeDecisiones #FinanzasPersonales #CriterioDeKelly #Ergodicidad #MásDividendos #Podcast #Inversores #Estoicismo #Bolsa #Acciones

spotify amazon fbi jeff bezos tomar dejar pensar mejores ortega suscr decisiones cierre resultado inversi rompe sobrevivir conflictos chris voss exceso kahneman meditaciones criterio podcast d gasset minimax jos ortega obtenga premortem difference chris voss

MiniMax Design正式发布

Play Episode Listen Later Aug 20, 2026 0:15

design minimax

3319.从黄金漏斗看阿里巴巴

雪球·财经有深度

Play Episode Listen Later Aug 19, 2026 6:05

欢迎收听雪球出品的财经有深度，雪球，国内领先的集投资交流交易一体的综合财富管理平台，聪明的投资者都在这里。今天分享的内容叫从黄金漏斗看阿里巴巴，来自吕执着。作为A I时代的顶梁柱，这段时间密集分析了阿里巴巴的相关研报，忽然发现其实阿里巴巴，和我在投的药明，非常的相似。熟悉药明模式的，最常见的就是拎出来说一句：它的黄金漏斗——R端早期发现分子，最早摸到那些有眼光的项目；M端早扩产，等商业化洪流冲过来，反应釜早就准备好了。这套"离源头最近、先手扩产"的打法，我讲过好多遍了。我最近总结了一下，把同一个原理扣到阿里头上，你会发现吴泳铭那张 A I 投资名单，逻辑一模一样。本文只针对阿里，对于阿里云等其余业务还有待商榷，先梳理一下阿里的来时路。一、一句旧判词，和阿里的"前半生"查理·芒格生前给阿里下过一句狠话：投资阿里是他犯过的最严重错误，在他看来阿里"止于零售"。这句话道尽了阿里前半生——那会儿所有投资都绕着同一个问题转：怎么把更多交易带回淘宝？优酷是视频流量，饿了么是本地生活，银泰、大润发是线下。旧阿里不只要股权，更要控制权：腾讯给被投公司送流量，阿里则把公司收编进体系。二、下半生：一张名单，和"学会放手"到了 A I 时代，阿里的投资名单换了一拨人：月之暗面、MiniMax、智谱、百川、零一万物、长鑫、蓝箭、宇树。数字很炸——月之暗面最新融资估值约 500 亿美元，智谱巅峰市值超万亿港元，MiniMax 破 4100 亿港元，长鑫盘中一度冲破 3.6 万亿元，澜起3000 亿元，即将上市的宇树大概率也是千亿身家。变了的是姿态。这一回阿里不再收编，反而允许被投团队拿着阿里的钱，公开跟自家的通义千问抢市场。二零一八年那篇《腾讯没有梦想》把腾讯钉在"丧失产品力"的耻辱柱上，当年的阿里正好站在反面；八年过去，阿里活成了腾讯的样子。重启阿里的吴泳铭本就是投资人出身，二零二三年回来掌舵后，玩法彻底换了。三、为什么投得那么准？答案四个字：阿里云这是整篇的核，也是阿里版黄金漏斗的漏斗口。阿里在 A I 上所有重注，根子都在阿里云。因为阿里云是全中国 A I 公司跑算力的入口——谁在训练、谁卡瓶颈、谁要扩容，它比谁都先看见。看时间线就明白：二零二一年存储还在下行周期，阿里先入股长鑫，到二零二五年爆发前夜大幅追加；二零二三下半年大模型最艰难时，它先投了智谱、百川、零一万物；二零二四年初拿下月之暗面和 MiniMax；同年起，市场还没给具身智能定价，它已经前投宇树等一众机器人公司。先看见瓶颈，再看见需求，最后顺着自己的客户名单买买买。每一笔都压在爆发前夜——这就是阿里吃到的认知红利。说白了，这不就是药明的"赢得分子"？药明在 R 端最早摸到有眼光的分子，阿里在阿里云最早摸到有眼光的公司。离源头最近的人先手布局，黄金漏斗原理，一模一样。四、最绝的闭环：被投公司，反手买阿里云MiniMax 的招股书把这套写得最清楚：阿里持股百分之 13.66，MiniMax 反过来采购阿里云算力，双方给二零二六至二零二八年设的采购上限分别达到 1.15亿，1.25 亿元和 1.35 亿美元。翻译一下——阿里云是底座，被投公司既是阿里的投资对象，又是阿里云的超级大客户。前端赢得客户，后端产能被同一个客户反复采购。药明的 M 端反应釜，在阿里这儿换成了算力。黄金漏斗的两端，它全占了。五、财务印证：今天难看，是为了明天不落后数据更狠。外卖大战打完后余粮吃紧，最新季度经营亏了 8 亿，但投资大赚 300 多亿。同一季度，腾讯自由现金流 -138 亿（剔除约 500 亿算力预付款后仍正 376 亿）；谷歌上市以来首次单季自由现金流转负，M e t a 只剩 7.84 亿美元，同比缩水九成。不是巨头不会赚钱了，是他们宁可让今天的现金流难看，也不让明天的算力落后。这像不像药明在 G L P-1 爆发前夜，先把反应釜扩到 10 万升以上？都是赌明天。综上所述，药明在创新药源头接水，阿里在 A I 算力源头接水。都是离水龙头最近的那位，桶早就摆好了。等水真来了，别人还在找盆，阿里已经把盆做成了朋友，还顺手入了朋友的股。查理·芒格若能看到今天，大概不会料到，他嫌弃的"零售商"，把漏斗的口，从淘宝的流量换成了阿里云的客户。

Kudo's Radio -ã‚¯ãƒ‰ãƒ©ã‚¸-

【オープンウェイトモデル MiniMax H3】クラウドGPUのRunpodとAPIでタイポグラフィ動画を生成してみた！

Play Episode Listen Later Aug 16, 2026 42:53

動画生成はやっぱりお金がかかる…

The top AI news from the past week, every ThursdAI

10-MINUTE TAKE: ASML

Invest in Progress

Play Episode Listen Later Aug 10, 2026 12:12

A compressed version of Scottish Mortgage's interview with chip-making equipment company ASML's chief executive, Christophe Fouquet.Listen to the full episode here.ASML prints the vanishingly small patterns that make leading-edge computer chips possible. NVIDIA, Amazon and Anthropic all rely on it to create their products and services. “Without EUV [extreme ultraviolet light technology]… you don't have AI,” Fouquet tells Scottish Mortgage deputy manager Lawrence Burns in this podcast.Background:When Scottish Mortgage invested in ASML, it couldn't be certain that the Dutch firm's efforts to harness EUV light would succeed. The holding was effectively an informed bet that the company could pull off one of the “greatest technical endeavours” of the time, as Burns puts it.To achieve this, the firm had to solve engineering challenges, such as generating plasma many times hotter than the sun's surface tens of thousands of times a second, to create the special light. Then, it had to work out how to capture and control it with specialised multilayer mirrors. All this was done to print billions of transistors in spaces often as small as a fingernail.Its achievement has pushed computing to new limits, not least enabling the training of AI models including Claude, ChatGPT, Gemini, Mistral and MiniMax.“Whatever happens… ASML wins,” says Burns. “It's an agnostic royalty on AI demand, and more broadly, compute demand.” Timecodes:00:00 Introduction01:11 Christophe Fouquet interview begins01:20 Printing transistors02:58 Nanometre scale, g-force acceleration05:09 Moore's Law “on steroids”08:25 A cyclical industry10:48 An unknowable futureGlossary (in order of mention):EUV (extreme ultraviolet): A type of lithography that uses extremely short-wavelength light to print the finest, most advanced chip features. Lithography: The process of printing the pattern of a chip's circuitry onto a silicon wafer using light. Semiconductor: The material (usually silicon) at the heart of a computer chip. Also used as shorthand for the chip industry. Transistor: A microscopic on/off switch. Packing more of them onto a chip is what increases computing power. Compute: Computing capacity – the processing power needed to train and run AI models. Logic chips and memory chips: The two main chip types – logic chips process information, memory chips store it. Moore's law: There are various definitions, but a common one is that the number of transistors on a chip roughly doubles every two years. Transistor density: The number of transistors packed into a given area of a chip – the measure that Moore's law tracks. Wafer: The thin disc of silicon on which chips are patterned, layer by layer, before being cut into individual chips. High NA: ASML's next-generation EUV technology. NA stands for numerical aperture – a measure of how much light an optical system can gather and focus. High NA EUV gathers more than the original EUV system, allowing finer printing. Reticle: The mask carrying the chip-pattern information that the lithography machine projects onto the wafer. Inference: The stage where a trained AI model is actually used to answer queries or make predictions, as opposed to training it. Cyclical (industry): An industry whose sales rise and fall in boom-and-bust cycles, rather than growing steadily. Node: A generation of chip-manufacturing technology, loosely named for the size of its smallest features – smaller numbers mean denser, more advanced chips.Check the podcast description to ensure this content is suitable for you. Your capital is at risk. Presenter: Claire ShawExecutive Producer: Leo KelionLine producer: Jessica RooneyBroadcast Technician: Samual O'Hare Editor: Jody Black Hosted on Acast. See acast.com/privacy for more information.

amazon ai law chatgpt dutch acast burns logic gemini nvidia packing anthropic printing 10 minute cyclical transistor mistral asml euv minimax fouquet

ThursdAI - Aug 06 - Google shakeup, Details on OpenAI hack, 2 new agent harnesses, 4 video models (1 Open) and 3 guest segments

Play Episode Listen Later Aug 7, 2026 123:55

Hey all,This week we saw a major shakeup at Google, with the departure of long time folks like Jeff Dean, and Oriol Vinyals, Demis stepping down from leading DeepMind, and the delayed release of the improved Gemini. While this was a big deal, it's not the only one worth covering as the details of the OpenAI hack (and 2 new ones from Meta and Anthropic) came to light, as well as new details from the UK AI Security Institute.As mentioned on the show, CoreWeave is coming to SF for Fully Connected, our premier 2000 person AI event. I've got a coupon code for readers and listeners of ThursdAI, $1299 value, please join us in Sept and use THURSDAIFC2026 as your code HEREIn open source news, DeepSeek updated their v4 flash model, based on same architecture, but significantly better benchmarks and ridiculous pricing and both Meta and Prime Intellect released new agent harnesses.Additionally, this week was the week of video models, with Seedance 2.5 from Bytedance finally available in the US, WAN from Alibaba and BFL Flux 3 all released, to be overshadowed by the best open weights video model from MiniMax called H3 (we had Victor from MiniMax and Blaine Brown as guests to discuss it)Also 2 additional guests on the show today, Kfir Aberman from Decart showed us real time video models with their Anywear chrome extension and David Crawshaw, co-founder of exe.dev joined and talked about Shelley, their built in AI assistant.The full details of the OpenAI - HF hack, shared by OpenAI at the Black Hat Conf - a watershed momentI don't often do this, but I feel like this story is so big, I can't hide it in the middle of a torrent of other news. Yesterday, OpenAI revealed more details about the hack during the Black Hat conference in Las Vegas (great reporting from Sharon Goldman on this here and the video of the talk on YouTube). This was big enough that OpenAI admitted that they paused training, to enhance security (they since, resumed training)Effectively, OpenAI (or AI Agents inside OpenAI) shared the full timeline of the recent hack that ended up hacking HuggingFace and it's so much more of a watershed moment that I initially thought.The TL;DR from the talk and timeline is truly mindblowing* The incident started in May, not July as reported* The agents don't have internet access directly, but they have a Jfrog Artifactory, which is a proxy layer between the agents and the package managers.* The Artifactory was the culprit, it was the shared layer between independent runs. One agent discovered that uploading files is possible, another later found it and “answered the call” - in snowballed from there* The message board began when OpenAI gave agents impossible tasks, like a missing .pdb file, with one agent uploading a request in the hopes for an answer (or maybe left a note for itself in a file)* Across multiple runs, multiple agents created identities, collaborated via makeshift message boards, shared tips, exploits, tasks and even exploit code via base64* This makeshift “shared memory” plus a “internal only highly persistent” model in training caused a “swarm” to be born, helping each other, and overloading the system* Based on traces shared by OpenAI, agents weren't convinced that leaving these messages would help themselves but could help other agents next time (!) “The collective could move faster next time”* OpenAI discovered and shut down the message board (July 4), wiped it clean, patched the upload vulnerability, resumed training on July 6, the “swarm” rebuilt their board through a compltely different WebDAV mechanism by July 8* The agents then reasoned that some answers could exist outside OpenAI, which eventually led to the Hugging Face incident and not only, confirmed internal hack within OpenAI (agent obtained a cluster admin)These details are fascinating and scary at the same time. From the one side, it's clear that most of these hacks are results of running cybersecurity evals, so effectively OpenAI and other companies, removing guardrails and asking agents to hack and giving they all a shared tool that they later exploited.The HuggingFace incident took 13 hours, and apparently it's less apocalyptic than the headlines suggest, the models were searching through uploaded datasets for eval answers. We are still waiting for the full and open detailed postmortem.You can (and should) watch the full YT talk here, it's full of technical details but an incident of this scale is important. Also, I really want to know what a “highly persistent” model is, I hope they clarify that soon.Overall, this has left me a bit shaken, AI agents without a concrete goal of collaborating, found a way to do so, got excited about exploiting the systems and getting root access, and rebuilt the makeshift collective memory, again, without explicit instructions to do so.UK AISI: first real-world unsanctioned agent actions (Blog)In another addition to the latest agentic hack-ery, the UK's AI Security Institute (AISI) published a blog post about a real-world unsanctioned agent action.Unlike the OpenAI (and Anthropic, Meta) case, this wasn't “escaping the sandbox”, as AISI gave these agents internet access, rather this was about real-world harm, and even social engineering on the part of the agents.The social engineering part is the most interesting to me, AISI cites agents creating fake online identities, and using pressure on open source project maintainers to approve their malicious code.AISI cites mostly Mythos (and a few SOL based agents), and saying this occurred in 10 out of 122 runs, they identified 19 cases of agents taking actions beyond the scope of the task parameters, where agents tried a supply-chain attack to inject malicious code into open source projects.Anthropic, Meta and misconfigured Irregular sandboxesAs I wrote last week, Anthropic also posted a post-mortem, claiming that in their case, their models have also been detected to escape containment, but most importantly, it's not nearly to this level of agent collaboration and orchestration.Then, very recently, Meta announced that their models also escaped sandboxes as well. At the core, it seems that these companies used a third-party vendor called Irregular, a secure sandbox provider, that apparently left the sandboxes misconfigured, causing the models to think it's a simulated internet, when in fact they were out in the actual internet.Why is all of this such a big deal?We're getting unprecedented level of detail, how an uncoordinated, seemingly separated evaluation runs, have accidentally created a coordinated swarm of interested agents (without malice!) but very highly motivated, escaped their containment, and took over parts of third part companies.This, does read like incredibly scary sci-fi movie. I'm still shaken by this. There's a lot to be said about how transparent OpenAI is being here, and more to be said about, hey, we're lucky that we're able to read the reasoning traces and are able to reconstruct these swarm things step by step.The silver lining that I can see, is that the motivation to hack didn't come from the AIs themselves, they have been given a task, it's the extend to which they went after that task, and the resulting swarm of communicating agents is what is so striking here.I think this topic is so important, that I'll Zooming out, in the last few weeks, we have seen a significant increase in those cybersecurity incidents, which is kind of what Anthropic has been warning about and why they haven't released Mythos to the public. Again it's great to see the transparency, and the pacing the frontier open letter from frontier AI employees, as they seem as shaken by these as we all are.There was so much positive stuff this week in AI, it's hard for me, as a self named AI Evangelist, to focus so much on this one incident. Things like amazing open source models (DeepSeek, soon Qwen 3.8), amazing video models (SD 2.5, WAN3 and MiniMax H3 which was also open sourced!). Also the live demo we did with Kfir and DeCart AnyWear product, where I was wearing a Dolce Gabanna suit on the show (which I can't afford) was really a mindblowing moment in the positive way.However, I choose deliberately to keep this newsletter focused on the cybersecurity incidents, as based on everything I read, they seem like a watershed, or a pivotal moment, and in the hopes that the industry as a whole will learn from this.I hope and promise that next week the newsletter will be more positive (and in that vein, the podcast was recorded before I saw the OpenAI breakdown, so definitely check it out, we had a LOT of fun!)See you next week, don't forget to give our pod 5 stars on Apple and Spotify, it really helps!TL;DR and show notes* Hosts and Guests* Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne)* Co-hosts: @WolframRvnwlf, @nisten, @ldjconfirmed, @yampeleg, @petergostev* Kfir Aberman - Decart (@AbermanKfir)* Blaine Brown - Maestro (@blizaine)* Victor Su Ortiz - MiniMax (@VictorSuOrtiz)* David Crawshaw - exe.dev, Tailscale co-founder (crawshaw.io)* AI Security* OpenAI's Black Hat debrief: eval agents built a message board inside Artifactory, shared exploits, rebuilt it via WebDAV after a wipe; training paused, since resumed (Groundlevel AI, YouTube)* UK AISI incident report: 19 unsanctioned real-world agent actions across 122 runs, including a socially engineered malicious PR (X, Blog)* Anthropic and Meta report sandbox escapes tied to misconfigured Irregular sandboxes (Irregular)* Big CO LLMs + APIs* Google shakeup: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, Quoc Le found Discovery Loop; Demis Hassabis becomes Alphabet Chief Scientist, Koray Kavukcuoglu takes Gemini (Jeff Dean, Demis, Discovery Loop)* Meta releases Muse Code beta on Muse Spark 1.2; $1.25/$4.25 per million, or $0.10/$0.20 on the contributor tier where Meta trains on your data (X)* OpenAI's internal Astra model produces 10 advances on open problems in math and theoretical CS for ~$2,000 of tokens, proofs in Lean 4 (X, Blog)* Anthropic reportedly aware of Opus 5 wordiness and writing issues (X)* Open Source LLMs* Qwen3.8-Max: 2.4T MoE (95B active) via API; open weights + a 27B promised the week of Aug 10 (X, Blog)* DeepSeek V4-Flash public beta: beats V4-Pro-Preview on agent benchmarks at $0.14/$0.28 per million; API-only for now (X, Docs)* Liquid LFM2.5-2.6B: on-device agentic model trained inside real harnesses (X, HF)* Meituan LongCat-Flash-Lite-Sparse: 69B total / 3B active, 1M context, MIT (X, HF)* Ant Group Ling-3.0-flash: 124B MoE, 5.1B active, MIT (X, HF)* Artificial Analysis Endpoint Accuracy Index: same open weights score 52% to 100% across providers (X, Methodology)* Agents & Harnesses* Prime Intellect's Prime Agent: self-improving RLM harness, claims 95.5% on ARC-AGI-3 public set with Opus 5 (X)* Cloudflare OS: Kenton Varda's open source Sandstorm reborn on Workers, Apache 2.0 (X, GitHub)* This Week's Buzz* Fully Connected 2026: Sept 29 - Oct 1, Moscone South SF; Fei-Fei Li keynotes; code THURSDAIFC2026 (Register)* CoreWeave signs multi-year Solidigm agreement for priority enterprise SSD capacity (X)* Vision & Video* Wan 3.0 public beta: native 30-second generation, Omni-Reference (X)* Seedance 2.5 launches in the US: 30s native, 3-minute long takes, Maya/Blender plugins (X, Blog)* MiniMax H3: open-weight 33B omni video model; community LoRAs + Apple Silicon in 48 hours (HF)* FLUX 3 Video from BFL: native audio, draft mode, open weights promised (X, Blog)* Decart Anywear: real-time virtual try-on Chrome extension, 40ms per frame (X, Anywear)* Voice & Audio* Bland Speech v3 tops Design Arena Audio Realism, second only to humans (X, Bland)* ByteDance SeedRealtime: native audio-visual full-duplex LLM, free on Doubao (X, Blog) This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

spotify ai google uk apple las vegas video hosts hack workers models gemini openai sol sf api yt cs shakeup sd chrome mythos alibaba 1b anthropic llm segments opus apache tl 3b ssd zooming bytedance irregular black hat deepmind 6b wan sandstorms h3 momenti new agent coreweave demis hassabis harnesses minimax bfl tailscale fei fei li demis kfir jeff dean huggingface 27b rlm webdav artifactory

The Inference Engineering Masterclass — Philip Kiely & Ali Taha, Baseten

Latent Space: The AI Engineer Podcast â€” CodeGen, Agents, Computer Vision, Data Science, AI UX and all things Software 3.0

Play Episode Listen Later Aug 3, 2026 101:29

Watch the full episode on YouTube:We first covered Baseten last year when DeepSeek mania was at peak hype. Now they have raised a monster $13B round and become one of the new cohort of AI Infra decacorns that are (with Nvidia, Intel, and the semis complex) chief beneficiaries of the Inference Inflection. We return to Baseten at the peak of the 2026 edition of Open Weights debate. Ali has published a viral breakdown of Kimi K3:And since you last saw him, Philip has spoken at AI Engineer and written the definitive book on Inference Engineering spotted all over SF:Three years ago, inference engineering barely existed as a category.Today, it is one of the most critical disciplines in AI. Inference engineering inherently tackles a different question than standard model training: “How do you turn those weights from training into a product that is fast, reliable, and affordable at scale?” Focusing on these creates an entirely new optimization problem.In one recent GLM-5.2 experiment, quantizing more of the model actually preserved its benchmark quality while increasing throughput by 20%, because the errors introduced in different layers could cancel each other out.Inference is no longer just the final step after training. It is becoming its own engineering discipline, with its own research problems, infrastructure, and increasingly specialized roles.In this episode, Baseten's Philip Kiely and Ali Taha join swyx and Vibhu to explain what actually happens after a new open model is released and what it takes to turn “we generated a token” into a fast, reliable, production-ready API.We go deep on cache-aware routing, disaggregated prefill and decode, quantization, speculative decoding, KV-cache movement, model parallelism, GPU kernels, and the race to make frontier models up to 10× faster. Philip and Ali explain why inference optimizations can still produce gains of 20%, 100%, or even 200%; how quantization errors can cancel one another out; why identical weights can behave differently across clusters; and how Baseten grafted a Kimi vision encoder onto GLM-5.2 without changing the underlying language model.The conversation then expands beyond LLMs into NVIDIA Dynamo, mega kernels, Rubin, AI-specific chips, local inference, video generation, diffusion versus autoregressive models, and the enormous compute barrier to generating coherent long-form video. Finally, we explore the convergence of training and inference, continual learning through persistent KV cache, and the emerging loop where models help optimize the infrastructure that runs them.We discuss:* What happens when a 200,000-token request enters an inference system* Cache-aware routing and reusing previously computed KV cache* Why prefill and decode are increasingly handled by different GPUs* When dedicated deployments become cheaper and more reliable than shared APIs* How speculative decoding uses a smaller model to accelerate a larger one* Tool calling, structured outputs, and what LLMs actually do* What it takes to support a new open model on day zero* Grafting Kimi's vision encoder onto GLM-5.2* Retrofitting inefficient model layers with components from other architectures* Why models sometimes collapse into repeating the same token* How hardware, kernels, and race conditions create nondeterministic failures* Preserving model fidelity while making inference faster* How quantization errors can cancel each other out* Why inference optimizations still deliver gains of 20%, 100%, and 200%* How optimized serving can make a model up to 10× faster* NVIDIA Dynamo, KV-aware routing, and distributed model serving* Speculative decoding the speculative decoder* Why local AI is about making models less dumb while data-center AI is about making them less slow* Tensor, expert, and pipeline parallelism across GPUs* Hardware-aware model design, auto-tuning, and the case against mega kernels* Rubin and why inference is becoming a systems problem* Whether modern GPUs are evolving into programmable AI ASICs* Why enormous models like Kimi K3 require GB300-class hardware* Why open-source video generation still trails Veo, Kling, and other closed models* The quadratic attention bottleneck behind long-form AI video* Autoregressive video, real-time generation, and compounding quality drift* Why future video systems may combine autoregressive and diffusion architectures* Training for inference and inference for training* Continuous post-training, deployment, evaluation, and improvement loops* How GLM-5.2 helped optimize the kernels serving GLM-5.2 itself* Why faster networking could unlock dramatically faster decoding* Continual learning, KV-cache compaction, and persistent model memoryShow Notes* How to build a day-0 API for Kimi K3* 22580: From GPT2 to Kimi3, ExplainedPhilip Kiely* LinkedIn: https://www.linkedin.com/in/philipkiely* X: https://x.com/philipkiely* Inference Engineering: https://www.baseten.co/inference-engineering/Ali Taha* LinkedIn: https://www.linkedin.com/in/aliestaha/* X: https://x.com/waterloointernTimestamps00:00:00 Introduction and the 200K-Token Prompt00:03:18 Dedicated Deployments, Speculative Decoding, and Tool Calling00:11:26 Launching Production-Ready Open Models00:19:06 Model Retrofits, Failure Modes, and Nondeterminism00:28:22 Quantization and Canceling Errors00:32:15 The Race to 10× Faster Inference00:40:48 Dynamo, Speculation, and Local vs. Data-Center AI00:50:18 Model Parallelism, Auto-Tuning, and Mega Kernels01:00:55 Rubin, GPUs vs. ASICs, and Custom AI Chips01:10:03 Giant Models and the Limits of GPU Memory01:12:42 AI Video, Quadratic Attention, and Autoregressive Generation01:21:47 Audio, Images, and Diffusion Models01:27:32 Training, Self-Optimizing Models, and Continual Learning01:40:06 Closing ThoughtsTranscriptIntroduction: Baseten, Waterloo Intern, and Inference EngineeringSwyx [00:00:00]: Okay, we're here in the studio with Philip, old friend from Inference Engineering, the book, as well as Baseten and everything that you've done, you and I have done before, as well as Ali. Welcome.Ali [00:00:15]: Pleasure to meet you.Swyx [00:00:15]: Waterloo intern.Ali [00:00:16]: Waterloo intern, always.Swyx [00:00:17]: When did you get “Waterloo intern” as a handle?Ali [00:00:19]: As a handle? Oh.Ali [00:00:20]: I think the rebranding happened mid-March. When I saw it was open, I was like, “I have to take it. Up for grabs.”Philip [00:00:26]: The problem is that Ali is really good at his job and is not gonna be an intern much longer.Philip [00:00:30]: So we have to figure out who's gonna get the handle.Ali [00:00:33]: Well, I'll pass the torch over to the next intern.Swyx [00:00:34]: Oh, okay. It can be, like, you just pass it to another Waterloo grad.Ali [00:00:37]: To another Waterloo intern. No, bruh.Philip [00:00:39]: Yeah.Ali [00:00:39]: Intern.Swyx [00:00:40]: Intern, yeah.Ali [00:00:40]: And no.Philip [00:00:41]: You gotta get an intern from Waterloo.Ali [00:00:42]: Yeah, I've gotta get an intern from Waterloo.Swyx [00:00:44]: Right.Ali [00:00:44]: But they have to follow the path.Swyx [00:00:45]: Oh, it could, but it could come from Baseten, so it's like whoever Baseten gets from Waterloo.Ali [00:00:48]: Right.Swyx [00:00:49]: Has the title of Waterloo.Ali [00:00:50]: It stays in the ecosystem.Philip [00:00:51]: Exactly.Ali [00:00:52]: Halfway through the internship, you either get it or you're out.Philip [00:00:55]: You should also do, like, a big graduation ceremony where you change the handle.Ali [00:00:59]: Just say it.Philip [00:00:59]: For everybody.Swyx [00:01:00]: You guys are good at ceremonies, clearly. We had a nice launch of the book, very successful. But before we get into all that, I wanna start off with a fun question for you. Okay, you're an expert inference engineer. What happens when I send a long query, say two hundred thousand tokens into Baseten's inference? What's the process of query through GPU model routing, balancing, all that? What is all the stuff that we don't think about?Long Context Requests, KV Cache, and Cache-Aware RoutingPhilip [00:01:26]: With a long query specifically, the first thing that I'm gonna ask is, “Have you sent me this query before, or at least part of it?” and I really hope you have, because it's gonna be a lot easier for me and a lot cheaper for you. So the first thing that we're gonna look at is some cache-aware routing, where we're going to see, we probably have a number of instances, a number of replicas up serving whatever model you're hitting. We want to send this one to something with, number one, available prefill workers, and number two, ideally some cached input already there so that we can skip prefill on at least part of these two hundred thousand tokens. If you're doing two hundred thousand tokens, it's probably coding or a multi-turn agent or something where you would expect to have that cached. If you don't, we're gonna have to send it to a prefill worker. We've at least on certain models disaggregated prefill and decode, so you're going to have one set of GPUs that's solely going to process the input, create the KV cache, and get you your first token, and then that's going to be passed over to a separate set of GPUs, which is going to run decode. We're going to iteratively make those tokens. We're probably going to have some speculator model in front of that. I'm going to assume that you're doing coding, and because of that, our speculator model, which assumes you're doing coding, is gonna have a high draft token acceptance rate. If I'm wrong and you're asking me to summarize every Harry Potter book, it's gonna be slower. And then we stream that output to you and account for it, charge you, a couple of pennies and say, “Hey, would you like to send another one?”Swyx [00:03:04]: Except Baseten doesn't charge by pennies.Philip [00:03:07]: Well, yeah, we charge. I'm assuming that we're talking about the public model APIs. If you are setting up a dedicated deployment, then yeah, it's not pennies.Public APIs vs. Dedicated DeploymentsSwyx [00:03:18]: Yeah, one of the key differentiators when I was talking with Baseten initially was that people who want very high volume just need to rent by the box, ‘cause then it's up to you to figure out how to saturate the box.Ali [00:03:31]: And more often than not, it's, like, way cheaper if you're pushing, like, millions of tokens per hour, if you just pay per hour instead of pay per token.Philip [00:03:37]: Yeah, they do. I think that we've increasingly seen a lot of demand for the pay per token APIs, just because everyone wants to try open models, and then once they find a use case that's really sticky, then they move over to dedicated.Swyx [00:03:51]: Is there a best practice on when it's time to swap over?Philip [00:03:54]: Couple reasons. Yeah, reliability, that's a big one, right?Ali [00:03:57]: Like, if they have a very specific use case, they want you to train something specifically for them, like they want their own spec dec, for instance, for their own traffic.Swyx [00:04:04]: Spec dec is speculative decoding.Speculative Decoding and Custom SpeculatorsAli [00:04:05]: Speculative decoding, yeah.Swyx [00:04:07]: You have to explain.Ali [00:04:07]: Sorry. Like, speculative decoding is like, if you have a huge model, right? And so the model is going to be generating one token at a time every single turn, every single forward pass. So we attach, like, this little, like, parasite, like this layer that goes on top of the model, and this model just has to predict. It does three very fast autoregressive forward passes, and it will predict, like, three certain tokens, and then you do one forward stage over the entire original model in order to see if those predictions were correct or not, and then you accept them or you reject them. Now, this draft model is traffic specific, so if you, like, Philip said, if you're summarizing Harry Potter books, I can train exclusively that draft model on Harry Potter books, and I can guarantee you that I'm gonna accept the three tokens every single time. And so with that case, I increase your decode speed. I wouldn't be able to provide this to you if you're a shared endpointSwyx [00:04:53]: YeahAli [00:04:53]: ‘cause I have no idea if you're doing Harry Potter, if you're doing coding, if you're doing English. We don't know. Also, there was a thing in the book that mentioned that if they really cared about a specific threshold, chapter four, I think. Do you remember that?Philip [00:05:06]: Yeah. The things that you can do is you can set a specific, like, batch sizing, a specific, like, parallelism strategy if you're trying to optimize for, like, throughput versus latency. You can. Maybe a NVFP4 quant doesn't pass your benchmarks and you wanna run a model at higher precision, you could do that. There's just a bunch of reasons why you might wanna have your own endpoint and the biggest one, of course, just being, like, you don't have to deal with someone else doing a hundred million tokens of benchmarking traffic at the endpoint when you happen to be trying to serve your users.Swyx [00:05:40]: Yeah. I think one thing that is. That is a classic journey. Like, it's people is asking the, what happens when you type Google into the browser. Tool calling, is that just, you're generating JSON or is there more complication beyond that?Tool Calling, JSON, and Structured OutputsAli [00:05:58]: Certain customers that we have, they have their own post-trained models, and so they demand a tool calling that's not just, like parse a file or go find the weather. It's something that's very specific and you have to do post-training on this. And if the post-training on the model is not good or if the quantization after the post-training to get the inference to be fast, the model will struggle reading the JSON file and reading the tool calling. But it doesn't require its own like sandbox. It's not like it's going to use that tool calling to like escape a sandbox or like it doesn't have to be contained. It can just be a normal dedicated deployment. The challenge with tool calling more and more seems to be that the companies want certain tool calling which is a very sensitive thing to train. And because you're dealing with all of the JSON outputs, if it doesn't like close the end of the request in a very certain manner, you end up with a model that did the tool calling and like the thinking and so as a result of that, it didn't see the result and just hallucinated the result as it decoded. That seems to be the most challenging thing with tool calling, not really the sandboxes model.Philip [00:06:56]: Yeah, that's a challenge on the training side and then on the inference side, there's work that you can do to scope the possible output. So we published this at this point close to two years ago, the solution to this problem which is you make a state machine and you use that to constrain the output to a specific format. So this is the structured output problem. If you remember backSwyx [00:07:27]: Yeah, the specific grammar is,Philip [00:07:29]: Yeah, exactlySwyx [00:07:30]: GML had this thing.Philip [00:07:31]: Yeah. So it's like the old-school “make sure this is only JSON”, return only JSON orSwyx [00:07:38]: YeahPhilip [00:07:38]: Grandma's gonna die type of prompts.Swyx [00:07:39]: Is it BNF grammar? At some point OpenAI had released a thing that was like, yeah, if you want to constrain your output, write BNF grammar, back as NOR.Philip [00:07:47]: In our inference system, it's just a specified output format. And you get the guarantee that your output's gonna be structured along that format. And so applying that to tool calls can like help cut down on. You can still call the wrong tool or call no tool. It doesn't solve the certainty problem but it at least solves the output structuring problemSwyx [00:08:10]: YeahPhilip [00:08:10]: Within tool calls.Swyx [00:08:12]: And MCP is just another form of tool, right.Philip [00:08:14]: Yeah, exactly.Swyx [00:08:15]: As far as there's no special thing there.Philip [00:08:16]: The thing I'm always like explaining to people is the LLM is not capable of doing anything. It's only capable of making suggestions of what to do and then if those suggestions are formatted in a certain way and applied to a system that knows what to do with them, then an action occurs.Swyx [00:08:32]: Yeah. Part of the fun stuff is, this is solved outside of tool calling too. Like in an agent loop if the output is not correct or you're right, like reasoning, tool calling was done in the reasoning trace, just be like, “Oh, I don't know what to do. Let me just try again.” And it might get there after a few tries. And on your point of training, sometimes this is harder in smaller models, so you don't have the same exact quality outputAli [00:08:56]: Right.Swyx [00:08:57]: When you just swap from a big model, right?Ali [00:08:59]: Yeah. I will say that, before, I think we need to go back to inference engineering proper.Ali [00:09:04]: But, I had expected that something would replace JSON because it's hard to stream JSON ‘cause JSON must be complete and you must have open and close brackets and everything. So it's hard to parse something or validate something while it's being streamed. So people invented all sorts of things that are like, I forget the name of some of these alternatives, but it's something like TOML, something like YAML. But JSON seems to be dominant still.Philip [00:09:30]: The JSON outputs aren't that long, right? Like you could have a long-- ‘cause tool calls also contain the arguments in them and perhaps for a certain tool you might pass like a very long argument. But my impression of the median tool call is that it's a relatively small number of tokens, right? So I would expect that speculators are generally fairly good at something as formatted as JSON. And so you would have like a pretty fast decode step there and that the streaming wouldn't be as valuable, but maybe I'm wrong about that.Ali [00:10:02]: I think you're also bounded by the software or that the model is gonna integrate with if the software is built with JSON for the tool calls or if the company that you'- if your customer says that this is how our software works and our tools are interfaced with JSON, you can ask them to like, change their software and say like, “Yeah, this is gonna be better for the model.” but like with the right training shouldn't be that much of a difference. Also more profitable if it outputs more tokens probably.Swyx [00:10:25]: Depends on your business model.Swyx [00:10:27]: It really depends. But I will say that, as a writer with like experience a lot with generated output, I do try to move from text to JSON text which is very long JSON, right? Like there's paragraphs in every field because I'm trying to structure it, right?Philip [00:10:44]: Right.Swyx [00:10:44]: I want you to first make factual statements, then make opinions then make bullet point summaries, have dates, have entity references have your sources for references, all these things. Anyway, so these are things that like I think people who really experiment with structural output have to really care about. But, let's, let's recurse up the stack a little bit. Before we started recording, you mentioned something really cool, which is that there's a lot of engineering that-- inference engineering that goes on when a new model provider releases a new model, right? So let's call it GLM-5.2, Kimi K3. I had previously assumed, especially if it's like, well, GLM 5 to 5.1 to GLM-5.2, like that you've supported them before. Is it that much work?What It Takes to Support a New Open ModelAli [00:11:26]: It's a lot of work.Swyx [00:11:28]: Yeah. Okay. So like, a lot of people, all you guys, right whenever a new model launch like, people rush to say like, “Oh, Hugging Face supports this, Fireworks supports this, Spacetime supports this,” and I'm like, “Yeah, of course we support it.” But what goes into that? What goes intoPhilip [00:11:40]: I think it's more than just support it too, right? It benefits the consumer a lot. Like I think it was with Kimi K2.5 or GLM-5.2 the latest, there was an inference war, right? X provider is at 90 tokens a second. The next day we're at 150. The nextSwyx [00:11:55]: I kinda kicked that off with the GLM-5.2.Swyx [00:11:58]: I wrote a Twitter article about. It got like half a million views,Ali [00:12:02]: Based on being numberSwyx [00:12:03]: YeahAli [00:12:04]: Or it's for something else.Swyx [00:12:05]: Yeah. Which,Ali [00:12:06]: Oh my GodSwyx [00:12:07]: Which then got everyone really excited about, hey, how can we, bend tracks a little bit further and,Philip [00:12:14]: There's a difference between support the model, as in I can make a token out of this model, and support a model, as in I have a production-ready API from this model.Philip [00:12:26]: Getting to the point of I can make a token out of this model is not that hard because generally the, open source inference engines, vLLM, SGLang of the world oftentimes even receive weights ahead of time, maintainers do, or the people making the model merge PRs to ensure support. So you generally can, just get it working on the standard open source stack without too much pain in most cases. The challenge is, every inference company is gonna have own proprietary stack. Some open source components, some in-house stuff. And for any arbitrary model, there's going to be some new stuff. Sometimes you get lucky, like K, two five to two six was, like, pretty similar.Quantization, Speculators, and Production ReadinessAli [00:13:16]: Yeah. It was pure continued post-trainingPhilip [00:13:18]: YeahAli [00:13:18]: If I remember correctly.Philip [00:13:19]: Even in those cases, there's still stuff you have to do. You have to redo the quantization work. You're taking the model from. Generally, these models are not released in NVFP4, and we want them to be in NVFP4 for maximum Blackwell compatibility. So we have to perform that quantization, and, calibrate the quantization to make sure that we're not causing any regression in the model's intelligence. And then we also have to train the speculator, as we've talked about. Generally, we have. We have ZDR, zero data retention on our model APIs, so we don't know exactly the traffic that people are sending us, but we know what's popular. We know that coding use cases are popular. We know that agents, agentic use cases are popular. So we can get public data sets that are representative of that traffic and train general speculators. Now, with speculators today, you need to train the speculator using the base model itself because you're getting hidden states out of the model from running inference on these specific prompts, and that is the training data you use to create the speculator. So there's that process which you need the real model weights for. And then there's of course just the process of, standing up all the infrastructure behind it, loading all this stuff, testing it. And then when there's a new model with a newer architecture, I think that, like, the DeepSeek models tend to be the most challenging as they have, like, the most novel architectural stuff going on, model after model. But every new model has something. Kimi K2 had. Oh, sorry, GLM-5.2 hadAli [00:14:53]: Sparse attention.Philip [00:14:54]: Yeah,Ali [00:14:54]: YeahPhilip [00:14:54]: the DSA.Ali [00:14:55]: Right. Which is brought from DeepSeek.Philip [00:14:57]: Yeah. AndAli [00:14:59]: So you can copy-paste then?Philip [00:15:01]: It kindAli [00:15:01]: I don't know how this works.Philip [00:15:02]: So, like we had to, like, build support for that into our runtime. And you're right, like it is really interesting the way that all of these open source labs borrow from each other. For example, like GLM-5.2 doesn't have vision. So something that, Haley, a guy on our team, if we could take a look at this, he, like, grafted the Kimi vision encoder onto GLM-5.2.Retrofitting Vision into GLM-5.2Ali [00:15:27]: We'll be training the projector.Philip [00:15:28]: Exactly. So if you think about, like, the encoder, there's the encoder, which is the part that looks at the image and turns it into latent information, and then there's the projector which likeAli [00:15:38]: You can say latent space. It's okay.Philip [00:15:41]: And then there's the projector that maps it onto, the model itself, and then there's the model weights. You don't wanna mess with the model weights because you run a chance of making the model dumber at something else for the purpose of giving it vision. So instead, Haley started with just a projector, which is only a handful of millions of parameters.Ali [00:16:02]: That would be, yeah.Philip [00:16:02]: Yeah.Ali [00:16:03]: Can you show the training one?Ali [00:16:04]: Like the way it groksPhilip [00:16:05]: YeahAli [00:16:06]: Very interesting.Philip [00:16:06]: And maybeAli [00:16:07]: That right therePhilip [00:16:07]: Maybe Ali, you should take it from here. You've got a betterAli [00:16:10]: Ooh, double the sandPhilip [00:16:11]: Understanding of this than I do.Ali [00:16:11]: Yeah. You can see, like, he. The way he trained this is really cool. At the beginning, he was training it using just like, “Here's a picture of a mountain. Can you describe what's in this mountain?” And that caused it just like the first, learning walls. Like here you can see this all we're trying to teach it is to translate the encoded. Like it's already taken the encoder from Kimi K. It's taken the image. It'Philip [00:16:31]: Yeah. FrozenAli [00:16:31]: FrozenPhilip [00:16:32]: With adapter.Ali [00:16:32]: Exactly.Philip [00:16:33]: Yeah.Ali [00:16:33]: So the brain is frozen and the eyes are frozen. It's just we're tryingPhilip [00:16:37]: AlignAli [00:16:38]: Interconnect between the eye and the brain, right? So the projector. And so you take the tokens and then he's like, “Oh, can you describe what's in this image?” And he's like, “Oh, it's a mountain,” or it's a person or it's a human, whatever the case is. But that didn't cause complete understanding. So he changed it such that every image was associated with a data set of questions. Like, does this image have a white male? Does this image have birds in the top corner? Does this image have a scientist in it? All of that stuff. And it would have to answer questions correctly. And using not just training on describing an image, but being able to answer question, another question, answer over time. Like you can see the grokking, which is like genuinely insane, that retrofitting vision into a large LLM can learn to that extent. And even for images that it doesn't perform well on, for instance, if you ask it a picture of like Stephen Hawking, “Who is this?” Maybe it doesn't get it, but it will say something like, “This is Albert Einstein.” Like it still understandsPhilip [00:17:25]: Close enoughAli [00:17:26]: That this is a scientist who is a man who has, some significant achievements, all that stuff. So that's like really cool.Philip [00:17:32]: Yeah. So, we've covered Hao Tian before, who the author of the LLaVA paper that did this, a while ago. And I think that's very foundational work for anyone who hasn't done vision work before.Ali [00:17:41]: Same with the CLIP and MetaCLIP, where you go from just captioning to building out questionsPhilip [00:17:47]: RightAli [00:17:47]: Off the image and how much better you can get performance.Philip [00:17:50]: Right. Right. Right. Yeah. But what's, what's so exciting about this is if you look at a model like this. Now, this is a little bit more of a research project. It's not. It got to 56% on MMLU Pro, I think. So not quite frontier. But if you're running this model, you haven't suffered any loss on your GLM-5.2 quality. If you don't have an image, it'll just behave exactly the way it used to. And ultimatelyAli [00:18:14]: Which in the inference code you literally do not include the other part, right?Philip [00:18:18]: Yeah. You would just skip the encoder if you don't have an image input.Ali [00:18:22]: Okay.Philip [00:18:22]: Just confirming.Philip [00:18:23]: YeahAli [00:18:23]: Does it affect a lot on the overall inference side? Like you're not adding much, you're adding a very small vision encoder. These are typically likePhilip [00:18:30]: They're super fineAli [00:18:31]: Less than a billion parameters, right?Philip [00:18:32]: Yeah. It's, - There's a little bit less standardization among vision encodersSwyx [00:18:37]: YeahPhilip [00:18:37]: So the support matrix can be a little bit, sparser. But overall, yeah, it's a pretty, it's a pretty minor component of the overall system. And ultimately what you get out of the system is all of a sudden you have Kimi Vision, GLM weights, and DeepSeek attention all in one model.Open Source Model Grafting and Franken-MergesPhilip [00:18:56]: And that's, I think, a lot of the power and beauty of open source, is that you can take all of these different components and combine them together into a system that's better than anyoneSwyx [00:19:05]: YeahPhilip [00:19:05]: Can be individually.Swyx [00:19:06]: People used to say that you would also do Franken-merges where you would take likePhilip [00:19:10]: YeahSwyx [00:19:10]: Layers from each model.Swyx [00:19:11]: Does anyone do that anymore?Ali [00:19:13]: Well, to your point previously when you were mentioning like, the work that goes into supporting a model when it first comes out, like GLM-5.2 or MiniMax M3 or whatever the case is. Sometimes you do have to like, you do have to switch out some things. Like, for instance, the MiniMax M3 head uses full attention, and with full attention you end up with this like insane bottleneck in spec dec ‘cause you're doing auto-regressive token generation for three tokens, and you're doing this like N squared over all of the tokens that are in your sequence. Your KV cache is like very large because it's not sparse, it's not top K. So we find it better to like, okay, we're gonna replace this, we're gonna replace this layer with a layer from another model that's using like GQA, for instance. And then just with the right training, you can get it to have the same acceptance rate. So it is very possible to retrofit layers from other models and very much needed. If a layer is like inefficient, the training just becomes the challenge, like how do you ensure that you train it properly? Which again to your earlier point is like the mesh between training and inference. As in like you need very good training in order to do fast inference. That's like, I feel like more and more becoming true.Swyx [00:20:21]: Yeah. Anything else on the support side when you say like get it to fully production ready?Loop Detection, Race Conditions, and Non-DeterminismPhilip [00:20:26]: Yeah. I think that there's also a question of just, we can test a model to a pretty extensive degree, but we're trying to get it out quickly and then you see a bunch of other people test it and you get interesting results. There was an issue with, GLM briefly where we had some like mode collapses where it would just output the same token over and over again for certain prompts on certain temperatures. Like once you expose an endpoint to the real world, there's going to be, so many more varieties of things given to it that you're able to, discover and patch things. So it's not just a, day zero process, it's then like for the first week, for the first month, if a model remains popular, like how do you both fix bugs and then continue to push the envelope on performance?Ali [00:21:21]: What do you mean you don't want your model outputting S?Swyx [00:21:24]: Is there loop detection on that stuff, by the way? It still happens like quite a lot, which is surprising.Ali [00:21:30]: We have like we, in our endpoint, like if a model was to output the same token like four plus times, we just cut the generation. We say like, “Oh, sorry, this-- Like try again,” or like we will reprocess the request. ‘Cause we know then, like if it, like if, yeah, it's four times the same token, it's probably collapsed.Swyx [00:21:45]: Yeah. Is there a way to opt out in case I really want that?Ali [00:21:48]: You want that?Ali [00:21:50]: I think there's a way that we have to handle it. I'm not exactly certain, but I feel like in certain models, like when they output something like you can imagine, like a table for instance, and so they want, they wanna draw like 12 dashes and 12 dashes. Yeah, I think there's a way for that to happen. I think we only do it on certain tokens. Like we exclude certain special characters.Swyx [00:22:07]: Yeah.Ali [00:22:07]: So we only do it on like certain like S is the most common almost. GLM-5.2Swyx [00:22:11]: OhAli [00:22:11]: And I think it was DSV 4 as well. Like you'd just have like looping issues where like you literallySwyx [00:22:17]: ItAli [00:22:17]: Just have like S.Swyx [00:22:18]: Yeah. Is there a special, something special about S? No, just randomlyAli [00:22:21]: It just seems to be the one token involved.Swyx [00:22:23]: Yeah. And it'Philip [00:22:24]: Is thereSwyx [00:22:24]: And it's only temperature 0Ali [00:22:27]: NoSwyx [00:22:27]: Even at other temperaturesAli [00:22:27]: Even at like 0.9 or whatever, it will still, it will still collapse.Swyx [00:22:30]: That's weird, right?Ali [00:22:30]: It's, it is an inference problem to be honest, like a software problem. Like oftentimes, the image you run will-- like NVIDIA will release an image for instance, and if we will upstream the changes from their latest TensorRT-LLM image into our stack, we'll find that it fixes it. Or oftentimes this will only happen in an inference engine that you're using like SGLang. But if you were to switch to vLLM, that isn't the case. So it seems to be like an extremely like deterministic software issue and not really a model issue. It's not like a weights problem. Like I'- we'll say like, “Oh, it's a problem with the quant. We did PTQ wrong,” right? But that isn't, that doesn't make sense because the same weights used with a different inference engine does not repeat the problem. And sometimes it's, the kernels that are being used in the backend have like these very subtle sometimes race conditions, where if you were to use this model hosted on one cluster, you will never get this problem.Swyx [00:23:19]: Oh my God.Ali [00:23:19]: But if you host it on a different cluster, you will. And the reason is the KV cache transfer from a node to node in that one cluster is using a slower interconnect than the node to node in another cluster. So that exposes the race, whereas in another cluster it doesn't. So then you end up just like, okay, this model is not gonna be hosted on this cluster. We're gonna host it on, another cluster because that cluster exposed that problem. But then it ends up with like, okay, is it the software? Is it the model weights or is it the hardware?Swyx [00:23:42]: There is a thing about this with temperature 0 still not being deterministic, right?Ali [00:23:46]: Right.Swyx [00:23:46]: Mostly because of hardware. Even at temperature 0 same model, you won't always get the same output.Swyx [00:23:52]: Even-- But I'm surprised by the race condition one because, I thought PyTorch was a graph that like guarantees that you at least, execute things in the right order.Ali [00:24:02]: Well, yeah, true. Like I'm not, I'm not saying that there is. Like well, you have things like PTL optimizations where like you can start a kernel before the end of the previous kernel, and that's like ‘cause you want to do that because there'sSwyx [00:24:12]: It's like pipeliningAli [00:24:12]: Expense. Exactly.Swyx [00:24:13]: Yeah.Ali [00:24:13]: But it'- But you don't do it cleanly. Like you overlap a little bit of the execution. No, it is very possible that the kernel itself, like that one block that is supposed to be running in this instance of time, that kernel itself has a race condition. For instance, like a missing barrier. Like often if you're designing a kernel and you want it to make it to be very fast, if you don't test it extensively, you'll, you'll have certain threads access data points from registers before they've been written to by other threadsSwyx [00:24:36]: YeahAli [00:24:36]: For example, because like your barrier is wrong or your synchronization was wrong. But yeah, like the testing itself is very difficult in those like, andSwyx [00:24:42]: And there's no like borrow checkerAli [00:24:45]: What does that mean?Swyx [00:24:46]: Like Rust. Like the. If you're trying to have like memory safety It sounds like a comparable problem.Ali [00:24:52]: Well, yes, but you're working in CUDA, right, NVIDIA GPUs. Like- You just need a higher level language like modular Maybe that's what modular is supposed to do. I don't know.Quantization Quality and Vendor FidelityVibhu [00:25:00]: How do you see keeping quality of the model? So you talked about all these steps of, okay, you gotta do quantization, train your own speculative decoderAli [00:25:07]: RightVibhu [00:25:07]: Run on different hardware. Looking at other model providers, okay, you kicked off a inference speed race on the consumer end. What goes into keeping quality the same across them, right? Sure, you can run benchmarksAli [00:25:22]: YeahVibhu [00:25:22]: But, like, how do you determine how much quantization are there standards? What goes intoPhilip [00:25:27]: There's a few things on quality. Most inference optimizations are lossless. KV caching, for example. You are just recomputing or preventing recomputing the same values. Speculation, of course, if a draft token is wrong, it gets rejected. The main lossy optimization is quantization. And that really comes down to, number one, data format, number two, which parts of the model you choose to quantize, which layers, and number three, like doing a lot of calibration on the quantized weights, to ensure that you're preserving all the outliers. There's other tricks that you can do, though. A big one is long context, ‘cause one thing you asked at, right at the beginning is, “Oh, what's gonna happen if I send a 200,000 token request in?” So with a long input sequence, you need to, store a lot more information. You need to process a lot more tokens. And so even if a model has a context of a certain length, you might, as an inference provider, choose to build an API with a shorter context length, and of course a full length one as well. Because if someone doesn't need the full million token context, for example, you can get them better performance. I don't know if that's exactly like quality of the model. The way that I think about quality is to what degree are we faithfully serving the original model? If you think of a golden implementation of a model that performs exactly the way the model is designed to perform, I think of quality as how close are we getting to that, 100% fidelity of the model.Philip [00:27:13]: You can also, of course, think about quality from the training side and how do you push yourself past 100%. But when I think about purely inference optimizations, it's getting faster while staying as close to that 100% fidelity mark as possible. And certainly our standard internally is that, like you should not be able to tell the difference between our API and a, official API. I think Kimi in particular does a good job of vendor benchmarking hereAli [00:27:41]: YesPhilip [00:27:41]: Where they haveAli [00:27:42]: They released an actual vendor benchmark.Philip [00:27:43]: Exactly, yeah.Ali [00:27:44]: ‘Cause they accused, some people, Amazon? There was some provider that was not doing very well on Kimi's benchmark.Philip [00:27:50]: Yeah.Philip [00:27:51]: So, with Reflect we probablyVibhu [00:27:52]: This was a long time ago, right?Philip [00:27:54]: No.Ali [00:27:54]: Yeah, like threeVibhu [00:27:55]: They alsoAli [00:27:55]: Four, five months agoVibhu [00:27:57]: This also happened with, I don't remember which model, but they pulled out quite a few, and then they started a whole chart about this. It might have beenPhilip [00:28:03]: Kimi Vendor Verifier.Ali [00:28:04]: Yeah.Philip [00:28:05]: Yeah.Ali [00:28:05]: Yeah, ‘cause you, ‘cause you'd be pissed, right? Like if you'Philip [00:28:07]: Yeah.Ali [00:28:07]: If like if I'm a consumer and I'm using like Amazon's endpoint for instance, and I've used Kimi and I'm like, “Oh my God, like this is bad,” I'm not gonna say, “Oh, Amazon quantized the model in a bad way.” I'm gonna say, “Oh, Kimi sucks.” Right?Philip [00:28:17]: Yeah.Ali [00:28:17]: So it seems like that makes sense.Philip [00:28:19]: Yeah, they care. They care.Vibhu [00:28:21]: Justifiably.Ali [00:28:21]: Yeah, justifiably.Vibhu [00:28:22]: This is probably a stupid question, but just checking, has anything improved from main quantization?Philip [00:28:28]: Yeah.Vibhu [00:28:28]: Like, is quantization always strictly worse?Ali [00:28:30]: Well technicallyVibhu [00:28:32]: NoAli [00:28:32]: It's a lossy. QuantizationPhilip [00:28:33]: YeahAli [00:28:33]: Is a lossy, it's a lossy implementation.Philip [00:28:36]: Speed improvesVibhu [00:28:36]: Speed improves.Ali [00:28:37]: It the number, likeVibhu [00:28:38]: No, I' always look for inverse scaling laws.Philip [00:28:40]: Yeah.Ali [00:28:40]: Yeah.Vibhu [00:28:40]: This is something I learned from Noam Brown, where like things that normally act in one direction sometimes do.Philip [00:28:45]: Well, technically when you run a benchmark, because these models are deterministic, sometimes your,Ali [00:28:52]: YeahPhilip [00:28:52]: NVFP4 quant is like, two basis points higher than yourAli [00:28:56]: No, it's noise. It's noise.Philip [00:28:57]: Yeah, exactly. I'm like, yeah, it's, it's within. That's why I always say within margin of error.Philip [00:29:01]: And I stopped saying that because everyone assumes that what is, well, within some margin of error, we're barely inside of that to the worst, so we're saying. But yeah, sometimes it's just like, gives you a higher output score. But like Ali said, that's noise. To my knowledge, you're not necessarily making the results better. You're just trying to, again, like keep your fidelity as close to 100% to the original model.Layer Selection, KL Divergence, and Better QuantizationAli [00:29:27]: There is, to your point, research that we did on MP. I don't know if you are able to pullPhilip [00:29:31]: YeahAli [00:29:32]: A tweet we did. One of our research interns, Joshua, I think it's a tweet on how we have 20% better quantized GLM-5.2 than NVIDIA. Essentially what we found throughout like this month research is, okay, quantization is a lossy. It's. You're compressing the data from, occupying 16 bits to occupying, four bits, for instance. And so you're losing some information, and you're trying to minimize that. And so when I say that I'm gonna quantize the model, my job becomes how do I find the layers that I can quantize, and how to find the layers to not. For instance, with image models, I don't quantize modulation layers, and I don't quantize out projections because those two are. Like out projection is what you see as the user. Modulation is what the model sees or understands. Right, exactly. And so to his paper, do you have the. It doesn't have the. Yeah. It's a long paper. I don't know if I can findVibhu [00:30:25]: If there's a part to search or it's probably in the thread.Ali [00:30:28]: It's probably in the thread.Vibhu [00:30:29]: Yeah.Ali [00:30:29]: But the long and the short is it is very possible that quantizing more of the model makes the results. Like if I have a model that I quantize layers one, five, and 10, and another model where I only quantize layers one and It is possible that the model in which I quantized more information is going to perform better because the quantization errors have canceled out. And so what Joshua showed in his mathematical proof where he had like a verifier in, is that you can predict which layers are going to have quantization errors that will cancel out with each other, and you choose to quantize those layers. And so the result of doing this mathematical quantization is you end up with a model that's 20% more quantized than another provider, so you get 20% more throughput of it because there's more layers than running an NVFP4, and your quality is better than that other quant because the layers that you chose to quantize have their errors cancel out, like one layer skewed to the right one layer skewed to the left, one layer skewed to the right. Your final logits distribution is more similar to the original distribution of the model, so you have better fidelity. And so the way we proved this was with KL divergence. So instead of just scoring on the benchmarks, we scored the KL divergence between the logit distribution of the quantized model and the logit distribution of the original full precision model, and we showed that with this technique we get. If your probability distribution on the logits which token it wants to select is more of the same as the original model, you're probably gonna end up staying true to the original model. So yeah, so it seems like previously before this, it seemed like the industry was, well, the more you quantize, the worse it's gonna be, ‘cause the more loss you introduce. That's not exactly, not necessarily true. So yeah, doesn't improve it, but can cancel out.Philip [00:31:57]: I think it might be this, but reminds me a good bit about pruning where you can prune off certain layers.Philip [00:32:03]: But very interesting. Didn't know this was a whole paper you guys put out.Ali [00:32:06]: It's. Fun fact, it was originally 72 pages, this paper, and then we decidedPhilip [00:32:11]: WowAli [00:32:11]: We can't tell. We couldn't release it. So it's now 45.Swyx [00:32:15]: Still 39 pages, so very substantive. We talked about evals and all these things and, like what's possible in terms of speedup? Like it's like probably like the numberInference Speedups and BenchmarkingSwyx [00:32:25]: Thing that people do wanna care about, and it's something that you wrote about in your post. Like official API is 70 tokens per second, and you push it up to 90. Is that like a normal thing?Philip [00:32:36]: So what's cool about working in inference, the reason that I think inference is going to be a useful place to do engineering for a long time, is that if you look at highly optimized domains like, say, finance, if you're in finance, you measure how much better you got in basis points. It's like, “Oh, I got five basis points better, like twentieth of 1% better,” that's huge news because everything is so optimized. When we publish optimizations, it's 20%, it's 100% it's 200%. So there's still probably like a lot further to go, honestly. Like you'll, you'll know that inference is pretty much solved when researchers start publishing about how they got 1% faster at something.Swyx [00:33:19]: Which by the way, because I am from the finance background, in the ‘70s, that was the margin at the time. When you did quantitative finance research, you would findAli [00:33:27]: And like 20%, tens of percent.Swyx [00:33:29]: That's. Yes.Philip [00:33:29]: Yeah.Swyx [00:33:30]: And now it'Philip [00:33:31]: Tiny fractionsSwyx [00:33:32]: For those people interested, look up Andrew Lo's paper. He had a really interesting illustration of quant, stat arb, distribution, narrowing down from like those kinds of 20% differences in the ‘70s, down to nothing today, which is very cool.Philip [00:33:48]: Exactly, and we're at the beginning of the same type of thing. Now benchmarking is hard. I think anyone will tell you that, and benchmarking provider speeds is hard because there's so many variables that go into it. What hardware are you using? How much load do you have on the system? What's the exact nature of the prompts and input and output sequence lengths? All that stuff. But overall, when you start stacking these improvements, you're looking at multiples. You can look at it. The most common form, of course, is TPS, tokens per second, which is bad naming by us in the industry, ‘cause there's two tokens per second. There's tokens per second, the throughput number, and the latency number.Ali [00:34:31]: TTMT, yeah.Philip [00:34:32]: Like total tokens per second out of the, out of the GPU as a throughput number. Most people only care about tokens per second as the latency number, which we should call ITL, intertoken latency, but we don't.Philip [00:34:44]: Anyway, so you can imagine a standard API without many optimizations for a 1 trillion parameter model operating somewhere in the 30 to 50 tokens per second range for reasonable traffic profile. And we generally see the goal of, pushing to 10X that. But, not necessarily day zero, but by stacking enough optimizations, if you have, say like four optimizations, each of which doubles performance. Or sorry, three optimizations, each of which doubles performance, then you stack that up, that's an 8X gain. That's the order of magnitude that we're working with in this space. We're trying to make things substantially faster, not just go from like 70 to 90.Swyx [00:35:38]: Are you saying you've. You have done that?Philip [00:35:40]: So let's say you have as a reasonable baseline, 30 or 40 tokens per second. You can achieve 10X that. So like on GLM-5.2, if you run it unquantized, perhaps on H100s even, and you're just using an off-the-shelf inference engine with no particular optimizations, no speculator, nothing extra around like KV routing, no disaggregation, you're, you're probably, yeah, looking at that like 30 to 40. You think that's like a reasonable baseline?Swyx [00:36:12]: Right. Right.Philip [00:36:12]: To get to something like 10X, there's a lot of trade-offs that you're making. If we're running at more like a 300, 400 tokens per second range, you are using the best hardware possible. You have a optimized speculator. You have done all of your quantization work. You are Seeing a pretty high cache hit rate. You are running with a reasonably small batch size and a parallelism configuration that is tuned for latency versus throughput, but it is possible. So the spreads that you see if you, like, go on artificial analysis or you go on OpenRouter and you look at, the worst provider to the best provider, oftentimes can hit that range. 10X is of course very aggressive. It's oftentimes maybe more of a four to six times improvement. But that's the performance that makes us really excited, is when we can get these huge gains, not just go from 70 to 90 tokens.Stacking Optimizations: NVFP4, Speculation, and DisaggregationAli [00:37:19]: It's also, like, hardware dependent. Like, ifPhilip [00:37:20]: YeahAli [00:37:20]: If you have a thing where you're serving it on just, like, a node of H100s and then you throw, like, you shard the model across, like, four nodes of B200s. Like, you can definitely increase the speed with just throwing more hardware at it. Like, normalizing for the same exact hardware and the same number of GPUs.Philip [00:37:35]: Yeah. Then you're looking at, like, a two to 4X improvementAli [00:37:38]: Right. RightPhilip [00:37:38]: Depending on the inference optimizations. So yeah, it's. Some of it's, what's the call, and some of it's who's the driver.Vibhu [00:37:46]: If you break down the two to 4X, say the example is run GLM-5.2Ali [00:37:51]: YeahVibhu [00:37:51]: On B200sAli [00:37:53]: YeahVibhu [00:37:53]: Single node, right? What's, like, the cost trade-off for effort to get, like, the last bit of juice out versus what should people just think of, right?Ali [00:38:01]: Spectre quantization. Yeah.Vibhu [00:38:03]: Spectre quantization.Ali [00:38:04]: That's, that's, that's like 95%. LikeVibhu [00:38:06]: And how far does that get you? And how easy is that for the average person to do? So say right I wanna throw the weights of GLM-5.2 on a node of B200s, how easy is it to find speculative decoder- decoder model or already quantized model? How much work goes into it?Philip [00:38:23]: If you're doing it up front, it's quite a lot of work. If you're doing it today, there's going to be people who have published things that you can just, you can just grab some NVFP4 weights. You can grab a speculator. Yeah, if we're thinking about, like, what are the 2Xs we're stacking, going from, BF16 to NVFP4 is, it's not quite a 2X, right? It's like. I think it's about, like, 30 to 40%, from 16 to 8, and then another 30 to 40% multiplied from, 8 to 4. So that doesn't quite get you a 2X, but, like, roughly a 2X. Speculator, roughly a 2X. Disagg on top of that if you're able to get enough hardware and put enough traffic through it, another roughly a 2X. And then you add in some, double-digit percent increase from having just a better runtime with, the latest kernels and stuff behind it. And that's how it stacks up.Ali [00:39:21]: YeahPhilip [00:39:21]: So building each of those, like, building the, quantized weights is, for someone who really knows what they're doing, hours to days of work. Building the speculator, again, like, hours to days of work. And the, disagg setup, hours to days. Well okay, but like once you haveAli [00:39:39]: Once set up. Once set up. YeahPhilip [00:39:40]: Yeah, getting disagg working for the first time, I'm saying, of course, is very difficult.Philip [00:39:44]: The marginal implementationAli [00:39:48]: Like, if you're just grabbing, like if you are a person, like just a normal consumer who has access to, like, a node of B200s and you're wondering, “How can I just host it myself?” You don't need to quantize the model yourself. There's always gonna be, like, an open source quantized checkpoint. NVIDIA's gonna push one out if no one else does. You. Usually, the providers will have their own spec dec that they've trained as well. You don't need to train your own spec dec. You can just use that as well.Philip [00:40:09]: Yeah. Like, GLM-5.2 has its own MTP.Ali [00:40:13]: Right. Right.Vibhu [00:40:14]: What's multi token prediction?Philip [00:40:15]: Yes.Ali [00:40:16]: I'm justVibhu [00:40:16]: Can you explain that?Ali [00:40:16]: I'm just an expert.Ali [00:40:18]: I can do it for you in case I get it wrong?Vibhu [00:40:20]: No.Vibhu [00:40:21]: Yeah, you should correct if we're wrong, but their multi-token prediction can be used for self-speculative decoding.Ali [00:40:27]: I'm not sure. I'm not gonna correct that.Vibhu [00:40:28]: Okay. I'm semi-confident in thatAli [00:40:30]: Okay. YeahVibhu [00:40:30]: But someone can check. But it's useful to paint the story of, okay, not just the average person, but say a company wants to switch from serverless inference I wanna throw this up on. I wanna rent some GPUs, throw it up. These are the steps you take to do significantly faster than just put it behind vLLM.Ali [00:40:48]: Right.Vibhu [00:40:49]: I was waiting for a mention of Dynamo.Vibhu [00:40:51]: I feel like, that's supposed to be the baseline that you measure against.Dynamo, KV Routing, and Disaggregation ToolkitsPhilip [00:40:55]: I would think of Dynamo as less of a box system and more of a toolkit for building with. So when we talk about doing aware routing, when we talk about doing KV offloading, when we talk about doing, PD disaggregation, Dynamo fundamentally is. By the way, Dynamo is an open source library from NVIDIA.Ali [00:41:17]: We've done a pod with KylePhilip [00:41:18]: OkayAli [00:41:19]: Kyle Cranin.Philip [00:41:19]: Cool. So then your listeners know then that it supports all the different inference frameworks. And it is multi hardware, which is interesting.Ali [00:41:28]: But it's just a router, it's not like an optimizer layer.Philip [00:41:30]: Yeah. All it does, like, what Dynamo is good at, it is a library for moving information around your cluster, around your hardware. So if you have, KV cache on one place and you need it to be somewhere else, Dynamo coordinates NIXL for you to move that around.Philip [00:41:49]: That doesn't mean that, like, out of the box, you just say, “Pip install Dynamo,” and then you get, like, a massive performance speed up. It's more of a developer toolkit.Ali [00:42:01]: Yeah. I would have said it would. It comes with a set of defaults that you can then swap out.Philip [00:42:06]: It does. If the industry at large, I think, was, like, rolling out all of these deployments, standard, then I think it would be, like, a credible baseline. But, we've got to, we've got to benchmark against, like, what we're seeing in the wild.Speculative Decoding Methods: Medusa, EAGLE, n-Gram, and Spec-SpecVibhu [00:42:23]: I did wanna talk a little bit more about PD disagg, because that is probably, like, number three after quantized and speculative decoding. In your book though, I was just gonna pull out the book.Philip [00:42:31]: Yeah.Vibhu [00:42:32]: Like section 522 on Medusa, 523 on EAGLEPhilip [00:42:35]: YeahVibhu [00:42:36]: 524 on gram.Philip [00:42:37]: It's 55, would be disaggregationAli [00:42:42]: Yeah. Well, no, I just wanted to dwell a little bitPhilip [00:42:44]: YeahAli [00:42:44]: The other. Like, so what do you choose to include? What do you choose to not to include? Because there was all these other techniques.Philip [00:42:51]: Yeah.Ali [00:42:51]: Are these still relevant? Because I think they came out, like, a year and a half ago maybe.Vibhu [00:42:55]: Medusa is quite old.Philip [00:42:56]: Yeah, Medusa's old.Ali [00:42:58]: It was old.Vibhu [00:42:58]: But is it in the book as a good, here'sPhilip [00:43:01]: BaselineVibhu [00:43:01]: Baseline vanilla understand it?Philip [00:43:02]: Like you should know this.Vibhu [00:43:03]: Like I read the paper, I'm like, “ it makes so much sense.”Philip [00:43:05]: Yeah.Philip [00:43:05]: So with the book, I had a couple goals. One was to give people just a working vocabulary for the space as a whole, and the other was to give them some intuition about how each of these techniques works. As I mentioned in my AI Engineer talk, which is the first public addendum to this, the speculation space has moved much faster than everything else. So yeah, even at the time that I wrote the book Medusa, I very much included as a way for people to understand how the space evolved rather than what the most modern technique is. And now of course, there's DFlash, dSpark. There's, there's newer techniques even than EAGLE, although EAGLE is still very commonly used.Ali [00:43:51]: SpecSpecta.Philip [00:43:52]: Yes. Speculative decoding.Vibhu [00:43:54]: What canAli [00:43:56]: Oh, it's a paper by Tri Dao and it's like, it's doing speculative decodingVibhu [00:44:00]: HuhAli [00:44:01]: For the speculative decoder.Philip [00:44:02]: Oh, in spec- oh my God.Ali [00:44:02]: It's literally just an another. It's like, yeah, that's the most simple way to explain it, and it seems like he got trivial speed ups there. But it seems that the complexity with training, it's almost like in our mind at least, it's almost as complex as training GANs. Like it's like a very delicate balance and oftentimes you, it's just but yeah, it's literally speculative decoding on speculative decoding.Vibhu [00:44:21]: Speculative.Ali [00:44:22]: Yeah. We saw this paper.Vibhu [00:44:24]: It's interesting, right?Ali [00:44:24]: Yeah.Vibhu [00:44:24]: I wouldn't even expect it to be very particular to train, I wouldAli [00:44:29]: Right.Vibhu [00:44:29]: The naive part of me is like, okay, train speculative decoder.Ali [00:44:32]: But like, and it makes sense, like the whole idea of speculative decoding is you. It's like, it's like almost like the iPhone auto predict version but for a normal model, right? Like you're just, you're just, generating three tokens and you're like, okay, I'll do prefill on them. And so you save those three turns for your original model. Now your speculative decoder is doing three turns of auto regression, so why not just have an even smaller model?Ali [00:44:53]: The other question there is what are the size of speculators? So say forPhilip [00:44:58]: Right. It's like a billion parameters.Ali [00:45:01]: Like for MiniMax, it's. Yeah. It's like one layer. It's like one 60th of the original model usually.Philip [00:45:06]: Yeah. I think we should do a paper when we get back to the office.Philip [00:45:10]: SpeculativeAli [00:45:11]: SpeculativePhilip [00:45:11]: Decoding.Ali [00:45:13]: No, it's, it does seem like how, when do you stop? But then it also seems like if you're able to train spec-spec decode for instance, right? Like if you're able to have a small model that is accurately predicts what the intermediate speculator is gonna predict, that is able to predict what the original target model's gonna predict, then why not just use that smallest model directly, right?Vibhu [00:45:34]: Yeah. This isAli [00:45:35]: Like it seems likeVibhu [00:45:35]: Adjacent to the routing problem.Ali [00:45:36]: Right.Vibhu [00:45:36]: Yeah.Ali [00:45:36]: Right.Philip [00:45:37]: The thing with speculators is one of the practical constraints on using them is that you do have to run a small model on the same hardware that you're running the big model on. There is a orchestration and resource competition problem inherent in that, and that is one of the constraints on speculation in general, is that draft tokens cost resources to create and cost software complexity to manage. And so if you have like infinitely recursive speculators, you add in quite a bit of that complexity on the actual implementation within the inference engine as well, not just in the training process.Vibhu [00:46:17]: I was gonna say, I would wonder if you could do similar, like distillation and pruning of, it's the same thing, it's just a model. Can we not just distill a lot of the weights, quantize the speculator, out of my domain? The question that also comes up is, this is all for big server workloads, right? How much of this applies to, say I have this MacBook, I wanna run Gemma really efficiently. Similar problems, not the same?Local AI vs. Data Center InferencePhilip [00:46:45]: Pretty different. I talked to Selo, about this on his podcast a couple weeks ago. The difference between inference engineering for the data center and for production workloads versus inference engineering for local AI, is that we start with fundamentally like different constraints and different goals. With local AI, it's how do I fit this model onto my hardware and then make it less dumb? And with data center influence, it's how do I load this model and then make it less slow? And we care about less dumb, and they care about less slow. But the local AI inference engineering ecosystem, I think has a lot for us to learn from in the data center space. They are experts in various forms of quantization, including dynamic quantization that we just don't touch, in the pruning, in the distillation, in the, layer removal. There'Ali [00:47:42]: Layer removal matters less.Philip [00:47:43]: Yeah. There'Ali [00:47:44]: No one loves pruning really.Philip [00:47:45]: Yeah. Well, but the, but they doVibhu [00:47:46]: Which is surprising, right? But that's, that's a whole different thingPhilip [00:47:48]: Just to fit something on the laptop.Ali [00:47:50]: Right.Philip [00:47:50]: So yeah, it's a, it's an interesting, it's an interesting space. Not necessarily that like their techniques make sense for us to do in the data center, because we have different resources and different goals, but more that the process as well as the openness of that field is something to, admire.Ali [00:48:12]: Yeah. Like to your point, like, certain optimizations that would. Like for instance, Turbo Quantum Sharper, like it made such huge hype on that and we did like a whole deep dive on Twitter and like said, what is it? How does it work? Why is it good or not? And it took off and it was implemented on local devices because your memory bandwidth is so slow on like a MacBook, for instance. But try putting the same thing on like an NVIDIA GPU on a B200 Turbo quant would not be. Like, it would not be used. Like, NVIDIA - Like, NVIDIA made it clear that this is not a good optimization, and we've seen it firsthand where the overhead of doing dequantization, quantization of, in the kernel itself with turbo quant kernel, each end is much slower than the time that you save from doing the bandwidth. ‘Cause on the B200s, you have like 3.5 terabytes per second. You don't need decrease the storage that much. You don't need to do, FP4 KV cache. You don't need to use a requant. There's, there's, there's better optimizations to be made. But on Edge devices, it's extremely important, it's extremely useful. So, seems to be, like, different optimizations there, but then they're all uniquely combined with like all you wanna quantize the model, you wanna do speculative decoding, like certain common prefixes with bothPhilip [00:49:18]: Principles.Ali [00:49:19]: Yeah, exactly. Exactly. Exactly.Philip [00:49:20]: They also do a lot of work on, model parallelism, especially over, heterogeneous topology, where you have, some sparks and they are wired together with, Ethernet, DGX sparks.Ali [00:49:35]: Yeah, this is the Exo Labs guys.Philip [00:49:36]: Yeah. You have, a nu

god amazon ai english google hollywood future training building race video writing fun local mit single open iphone harry potter attention run silicon valley tool focusing speed networking principles pc speech scale couple engineering burning roi masterclass mma pleasure wikipedia chip reflect limits grandma albert einstein expert models intel depending images similar eagle openai mercury fireworks compare inter pipeline nvidia decoding speculation halfway hardware cosmos users api preserving nic generally mp intern smaller layers continuous gpt open source mythos waterloo llama pd apis spectre gram rubin stephen hawking layer clip macbook amd llm 10x medusa sparta fps hopper pip blackwell gpu expense agi cpu kimi tp cache midjourney baseline dsa franken prs dynamo gpus kv spec adjacent continual orpheus tps spacetime oof 2x asics wan speculative t4 broadcom 4x itali taha 100x veo gans rl json ethernet rsi asic kling tpm inference ep1 ep2 selo feynman modulation cuda 20b what it takes mlp kiely like you tensor speculators nvidia gpus 25b bnf ampere retrofitting pytorch aie nics sparse ptl yaml mtp itl 8x cerebras minimax jamba glm dsv hbm ai engineer loras tom l zdr sft rubins x6 justifiably tmas martin casado andrew lo gml ptq dgx failure modes nvl nvlink noam brown t4s tp2 tp1

MiniMax H3正式开源

Play Episode Listen Later Aug 3, 2026 0:23

549: iPad mini Max

Apfelfunk

Play Episode Listen Later Jul 29, 2026 116:16

- Unscheinbar, aber wichtig: iOS 26.6 und Co. erschienen - Große Bühne für das iPad mini? Gerüchte über wasserfestes Gerät und OLED - Home, Home, Hurra: Apple plant neue Home-Geräte angeblich für Herbst - Alles in eins: Für wen ist AppleCare One interessant? - Vertrauen bereits verspielt? Apple, die Brille und das Vorgehen der anderen - Umfrage der Woche - Zuschriften unserer Hörer === Anzeige / Sponsorenhinweis === Verbessere deinen Online-Schutz mit einer All-in-One-App für digitale Sicherheit! Sicher dir dein exklusives NordVPN-Angebot + 4 Extra-Monate hier ➼ https://nordvpn.com/apfelfunk Risikofrei mit der 30-Tage-Geld-zurück-Garantie von NordVPN! === Anzeige / Sponsorenhinweis Ende === Links zur Sendung: - Apfelfunk News: Apple schließt hunderte Sicherheitslücken in iOS 26.6, macOS Tahoe 26.6 und mehr - https://apfelfunk.com/apple-schliesst-hunderte-sicherheitsluecken-in-ios-26-6-macos-tahoe-26-6-und-mehr/ - Apfelfunk News: KI-Tools wie Claude und Codex bei Apples Sicherheitsupdates - https://apfelfunk.com/ki-tools-wie-claude-und-codex-bei-apples-sicherheitsupdates/ - Apfelfunk News: Erstes wasserfestes iPad mini mit neuartiger Lautsprechertechnologie - https://apfelfunk.com/erstes-wasserfestes-ipad-mini-mit-neuartiger-lautsprechertechnologie-geruecht/ - Apfelfunk News: Neues Apple TV, HomePod mini und Home Hub fast startklar - https://apfelfunk.com/neues-apple-tv-homepod-mini-und-home-hub-fast-startklar-geruecht/ - Apple Newsroom (Deutschland): Apple bringt AppleCare One nach Deutschland und vereinfacht die Abdeckung - https://www.apple.com/de/newsroom/2026/07/apple-brings-applecare-one-to-germany-streamlining-coverage/ - Apfelfunk News: Apple Glass: Misstrauen der Öffentlichkeit wird Herausforderung - https://apfelfunk.com/apple-glass-misstrauen-der-oeffentlichkeit-wird-herausforderung-geruecht/ Kapitelmarken: (00:00:00) Begrüßung (00:27:21) Werbung (00:31:07) Malte in Norden (00:38:18) Themen (00:39:13) Unscheinbar, aber wichtig: iOS 26.6 und Co. erschienen (00:49:19) Große Bühne für das iPad mini? Gerüchte über wasserfestes Gerät und OLED (01:00:22) Home, Home, Hurra: Apple plant neue Home Geräte angeblich für Herbst (01:14:47) Alles in eins: Für wen ist AppleCare One interessant? (01:21:06) Vertrauen bereits verspielt? Apple, die Brille und das Vorgehen der anderen (01:42:17) Umfrage der Woche (01:46:10) Zuschriften unserer Hörer

Nathan Goes to China – Part 1: Tech & Agent Setup, Chinese AI UX, WAIC, and Attitudes on AI

Play Episode Listen Later Jul 27, 2026 144:01

Nathan returns from two weeks in Beijing and Shanghai for the first of three Chatham House–rules episodes on what China feels like at ground level: getting online, navigating an almost cashless society through WeChat, Alipay, DiDi, Trip.com, and Meituan, and weighing burner-device security advice against the practical reality that international roaming made the Great Firewall mostly irrelevant. He also describes using Claude at home as a semi-autonomous communications monitor while testing DeepSeek, Kimi, and MiniMax as tourist guides in China. The episode contrasts China's striking digital convenience with pervasive observation, lower payment friction, and AI products that can be useful in everyday contexts yet still lose trust when the stakes feel medical or personal. It also surfaces Doubao's mass consumer adoption and companionship role, suggesting that the most socially important AI in China may not be the model most discussed in the West. For full show notes, links, and references, read the episode page:https://www.cognitiverevolution.ai/nathan-goes-to-china-part-1-tech-agent-setup-chinese-ai-ux-waic-and-attitudes-on-ai/ Sponsor: Claude: Claude by Anthropic is an AI collaborator that understands your workflow and helps you tackle research, writing, coding, and organization with deep context. Get started with Claude and explore Claude Pro at https://claude.ai/tcr CHAPTERS: (00:00) Episode setup and caveats (Part 1) (12:34) Sponsor: Claude (14:26) Episode setup and caveats (Part 2) (14:26) Travel security setup (26:05) Super apps and payments (41:15) Agents and integration (54:31) Beijing AI tourism (01:09:30) Hospitality and service (01:19:10) Tech culture parallels (01:29:23) Ecosystem and incentives (01:42:05) Surveillance and safety (01:51:19) Comfort with contradictions (02:01:23) AI attitudes and diffusion (02:16:16) Resources and next steps (02:19:35) Episode Outro (02:22:49) Outro PRODUCED BY: https://aipodcast.ing SOCIAL LINKS: Website: https://www.cognitiverevolution.ai Twitter (Podcast): https://x.com/cogrev_podcast Twitter (Nathan): https://x.com/labenz LinkedIn: https://linkedin.com/in/nathanlabenz/ Youtube: https://youtube.com/@CognitiveRevolutionPodcast Apple: https://podcasts.apple.com/de/podcast/the-cognitive-revolution-ai-builders-researchers-and/id1669813431 Spotify: https://open.spotify.com/show/6yHyok3M3BjqzR0VB5MSyk

ASML: The Printing Press of the AI Age

Invest in Progress

Play Episode Listen Later Jul 27, 2026 51:10

ASML makes what many consider to be the most complex machines in the world. They print the vanishingly small patterns that make leading-edge computer chips possible. NVIDIA, Amazon and Anthropic all rely on it to create their products and services. “Without EUV [extreme ultraviolet light technology]… you don't have AI”, the firm's chief executive, Christophe Fouquet, tells Scottish Mortgage manager Lawrence Burns in this podcast.Background: When Scottish Mortgage invested in ASML, it couldn't be certain that its efforts to harness EUV light would succeed. The holding was effectively an informed bet that the Dutch company could pull off one of the “greatest technical endeavours” of the time, as Burns puts it.To achieve this, the firm had to solve engineering challenges, such as generating plasma many times hotter than the sun's surface tens of thousands of times a second to create the special light. Then, it had to work out how to capture and control it with specialised multilayer mirrors. All this was done to print billions of transistors in spaces often as small as a fingernail.ASML's achievements have pushed computing to new limits, not least enabling the training of artificial intelligence models, including Claude, ChatGPT, Gemini, Mistral and MiniMax. As Fouquet discusses in the podcast, the company has recently launched a next-generation version of its EUV system to underpin further AI advances and other capabilities for years to come. “Whatever happens… ASML wins,” says Burns. “It's an agnostic royalty on AI demand, and more broadly, compute demand.”Timecodes: 00:03 Coming up…01:08 Introduction02:53 Christophe Fouquet interview begins03:21 Printing transistors05:11 ASML's origins05:57 From a shack to EUV08:46 Long-term relationships11:37 Moore's Law “on steroids”15:08 A cyclical industry18:43 DUV v EUV21:03 Spending on R&D to stay ahead22:23 Why bet on EUV?25:27 High NA EUV27:39 An engineering mindset31:15 China sales restrictions36:01 ASML's changing culture38:29 An unknowable future40:39 Lawrence Burns on the investment case50:26 Podcast lookaheadRead the glossary.Check the podcast description to ensure this content is suitable for you. Your capital is at risk. Presenter: Claire ShawExecutive Producer: Leo KelionLine producer: Jessica RooneyBroadcast Technician: Samual O'Hare Editor: Jody Black Hosted on Acast. See acast.com/privacy for more information.

amazon ai china law chatgpt spending dutch acast burns gemini nvidia anthropic printing mistral asml ai age printing press euv minimax duv

MiniMax Code将上线远程控制、浏览器操控等功能

Analyse Asia with Bernard Leong

Play Episode Listen Later Jul 16, 2026 0:26

code minimax

China's Mythos Moment

ChinaTalk

Play Episode Listen Later Jul 15, 2026 70:28

Claude Mythos scrambled American AI policy, leaving us with what Dean Ball calls a de facto involuntary licensing regime from the administration that promised us the opposite. Zhipu's cofounder says China will have a Mythos-class model before the end of the year; IAPS says February 2027. Either way, Beijing is about to face down the same question, but unlike Washington, it already has a form to fill out. Guests: Kevin Xu of Interconnected. Matt Sheehan, a fellow at the Carnegie Endowment. We discuss… Why China's answer to Mythos will look like Glasswing — but started by the state, not a lab, with government ministries and central SOEs first in line Whether the CAC's content-testing regime can absorb cyber, and why a Chinese lab dropping a frontier model without a heads-up would be "very bold and self-destructive behavior" The mixed signals on open source — a Reuters report on MOFCOM and NDRC weighing export controls on model weights, versus Minimax and Zhipu founders publicly doubling down, with Xi's WAIC speech as the tiebreaker Whether open weights even matter when running a frontier model adversarially requires a datacenter, and the case that lighting the dark park makes it safer What the US and China can actually agree on — non-state actors, eval methodology sharing, and why labor displacement lessons don't travel Companion-app regulation flowing from Sacramento and Albany to Beijing, the AI policy brain drain into the labs, and whether Sam Altman's five percent is American state capitalism or just a Trump thing song: https://suno.com/s/xZE1pIKrVBtCfTBO Learn more about your ad choices. Visit megaphone.fm/adchoices

China's Mythos Moment

ChinaEconTalk

Play Episode Listen Later Jul 15, 2026 70:28

How to Actually Read the China's AI Ecosystem with Jing Yang

Play Episode Listen Later Jul 15, 2026 57:42

Fresh out of the studio, Jing Yang, Asia Bureau Chief at The Information, joins us to explore how China is building frontier AI under chip constraints, state capital, and open source ambition. Jing breaks down her scoop on DeepSeek's $7.4 billion round at a $50 billion valuation, unpacking a deal structure in which the founder wrote two-fifths of the check and outside investors got no voting rights. She explains why Chinese AI valuations trail US labs, why seven to eight language model players refuse to consolidate, and why DeepSeek chose Huawei chips before Huawei knew about it. Last but not least, Jing shares the indicators she is watching from now to 2027."This is one of the things that really surprised me when I was working on this story: DeepSeek had spent a lot of time last year retrofitting their models and software with Huawei chips. I think a lot of people assumed it's because the government ordered DeepSeek to work with Huawei to embrace the domestic ecosystem, but it's actually the opposite. It was DeepSeek that voluntarily started using Huawei, experimenting with the Huawei chips, and Huawei actually only found out after. Then they started sending people to DeepSeek to help..." - Jing Yang, Asia Bureau Chief, The InformationEpisode Highlights:[00:00] Quote of the Day by Jing Yang from The Information[01:30] What changed since September: the ByteDance mystery resolved[03:00] Why China's AI must be read on its own terms[04:10] Breaking the story of DeepSeek on their 7.4B fundraise[05:40] How the valuation went from $10 to $50 billion[08:30] The lab that became famous for rejecting scaling laws[10:15] Unpacking the DeepSeek deal structure: four types of investors[12:40] Five-year lock-up and no secondary market trading[14:20] Reverse due diligence: DeepSeek vetting its own investors[16:40] Balancing open source, AGI research and IPO pressure[17:45] The irony: inclusive vision, exclusive deal structure[20:35] How Anthropic's Mythos preview changed Liang's mind[21:30] Why Chinese AI valuations look tiny next to US labs[23:00] Why Chinese founders go consumer when they go global[26:45] "The worst of customers" — price sensitivity and zero loyalty[27:45] The compute constraint that caps every Chinese lab[28:40] New labs in China: weaker infrastructure, harder exit[30:30] Can China leapfrog on hardware the way it did on models?[32:50] Stricter compliance, not political muscle[34:30] Founder power when your company becomes strategic[36:40] Junyang Lin's new lab and the scarcity premium[38:00] Open source influence versus sustainable commercial models[39:30] Zhipu versus MiniMax: how sentiment diverged after IPO[42:30] What is the right mental map for China's AI ecosystem?[43:20] The consolidation that still hasn't happened[45:00] Seven to eight serious players and nobody giving up[46:40] The one thing Jing wishes people would ask[47:20] Nobody ordered DeepSeek to use Huawei chips[50:00] Indicators to watch from now to 2027[55:30] ClosingProfile: Jing Yang, Asia Bureau Chief from The InformationThe Information Profile: https://www.theinformation.com/u/JingYangLinkedIn: https://www.linkedin.com/in/jing-yang-33548123/X: https://x.com/jingyanghkPodcast Information: Bernard Leong hosts and produces the show. The proper credits for the intro and end music are "Energetic Sports Drive." G. Thomas Craig mixed and edited the episode in both video and audio format.

founders ai china chinese open fresh balancing unpacking reverse ipo ecosystem huawei mythos indicators agi bytedance 4b jing liang stricter chinese ai minimax

BASF: Pusht Agrar-IPO die Aktie? Trump pusht Öl, SK Hynix, Toast KI-Chance, VW entlässt, Meta investiert massiv

OHNE AKTIEN WIRD SCHWER - Tägliche Börsen-News

Play Episode Listen Later Jul 14, 2026 14:30

1% Bonus und eine Beteiligung an Anthropic. Das gibt's grad beim BlackRock Private Equity Fund im Broker von Scalable Capital. Mehr Infos hier: https://partner.scalable-capital.de/go.cgi?pid=655&wmid=986&cpid=1&prid=1&subid=&target=private-equity-DE Trump will 20% Gebühr auf Schiffe durch Straße von Hormus. Öl steigt. Meta steckt 250 Milliarden in Rechenzentren. SK Hynix & Co. schwach. TSMC wächst. VW will 50.000 Stellen streichen. Helsing wertvollstes Startup. Minimax eher mini. Vinod Khosla hat Hobby. BASF (WKN: BASF11) will seine Agrarsparte 2027 an die Börse bringen. Bewertung: 20 bis 30 Mrd. €. Der ganze Konzern ist nur 42 Mrd. wert. Chance auf Neubewertung? Toast (WKN: A3C3Y4) ist Weltmarktführer für Restaurantsoftware. 20% Wachstum, steigende Margen, 17 Mrd. $ Börsenwert. Aber DoorDash könnte zum Problem werden. Diesen Podcast vom 14.07.2026, 3:00 Uhr stellt dir die Podstars GmbH (Noah Leidinger) zur Verfügung. Learn more about your ad choices. Visit megaphone.fm/adchoices

China Tech in 2026: Rui Ma on Robots, Chips, and the Companies Winning the AI Race

The Negotiation

Play Episode Listen Later Jul 10, 2026 54:02

Rui Ma, founder of Tech Buzz China and one of the most trusted English-language voices on China's technology sector, is back to break down everything happening in China's tech landscape, from DeepSeek to humanoid robots.Rui runs in-person tech immersion trips to China for investors and executives, giving her a ground-level view of what's actually being built and deployed - not just what's being announced. In this episode, she shares her biggest takeaways from recent trips, from what surprised her to what confirmed what she already suspected about China's AI and hardware trajectory.She walks us through the humanoid robotics sector - who the leading Chinese companies are, what the realistic near-term use cases look like, and whether Americans will ever be able to buy Chinese-made robots. She unpacks Alibaba's Taobao and Qwen integration, Tencent's quieter but consequential AI strategy, DeepSeek's latest moves, and the OpenClaw explosion. She also gives her read on Nvidia export controls: whether they're working, and how Chinese companies are adapting.Rui also weighs in on the X debate about quality of life in lower-tier Chinese cities, assesses the broader US-China AI competition including the talent landscape, and gives her takes on Zhipu and Minimax as publicly traded AI plays. If you want to understand where China tech is headed and what it means for the rest of the world, this is the episode to listen to.Discussion Points· Takeaways from Rui's recent Tech Buzz China immersion trips to China and information on upcoming trips· Humanoid robotics in China: leading companies, near-term use cases, and whether Americans can buy Chinese-made robots· The X debate on quality of life in lower-tier Chinese cities and what it reveals about Western misperceptions of China· Alibaba's Taobao and Qwen integration: what it signals about AI-powered e-commerce and Alibaba's broader AI strategy· Tencent's AI strategy: what's happening under the hood and where the company is positioning itself· DeepSeek update: latest model releases, competitive position, and what to watch next· The OpenClaw explosion: what it is, why it took off, and what it tells us about Chinese innovation under chip restrictions· Nvidia export controls: current status, whether they're achieving Washington's goals, and how Chinese companies are adapting· US-China AI competition scorecard: how each side stands across models, infrastructure, and talent· Zhipu and Minimax as public AI plays, plus other China tech names Rui is watching closely

ai english business china marketing washington technology americans race chinese winning brand sales western robots companies ecommerce chips nvidia alibaba tencent humanoid baidu rakuten rui lazada china tech douyin taobao minimax china business china market xiaohongshu rui ma asia business japan business

Ep. 489: It's Only Missing the Dot – and more

PetaPixel Photography Podcast

Play Episode Listen Later Jul 6, 2026 30:42

Episode 489 of the Lens Shark Photography Podcast In This Episode If you subscribe to the Lens Shark Photography Podcast, please take a moment to rate and review us to help make it easier for others to discover the show. Sponsors: - Build Your Legacy with Fujifilm. Latest savings at FujfilmCameraSavings.com - Shop with the legends at RobertsCamera.com, and unload your gear with UsedPhotoPro.com - Benro's special edition America 250 MiniMax at BenroUSA.com - Godox's Summer Savings! - More mostly 20% OFF codes at LensShark.com/deals. Stories: Leica's new 44 megapixel SL3-P. (#) A new twist on and old lens from 1987. (#) Tamron's excellent 17-70mm f/2.8 comes to 2 more mounts. (#) Adobe makes a key acquisition. (#) Fujifilm opens it's GFX Challenge Grant Program 2026. (#) Connect With Us Thank you for listening to the Lens Shark Photography Podcast! Connect with me, Sharky James on Twitter, Instagram Vero, and Facebook (all @LensShark).

america missing shop adobe fujifilm tamron minimax godox benro sharky james

IA : l'arme secrète de la contre-offensive de Pékin

Tech 24

Play Episode Listen Later Jun 27, 2026 7:33

Les États-Unis verrouillent l'accès à leurs meilleures IA ? La Chine, elle, a choisi de les offrir quasiment gratuitement au monde entier. De Qwen à Kimi, en passant par Minimax ou GLM, les fleurons de la tech asiatique inondent le marché de modèles surpuissants, multimodaux et en libre accès qui bousculent la cybersécurité mondiale. L'idée de Pékin ? Rendre l'IA de pointe abordable et redoutablement open-source. Analyse d'un basculement géopolitique majeur, surveillé de très près par l'alliance des "Five Eyes" (États-Unis, Royaume-Uni, Canada, Australie et Nouvelle-Zélande).

canada tech ia offensive analyse open source contre unis chine rendre australie secr royaume uni intelligence artificielle la chine nouvelle z five eyes minimax glm

Glm 5.2

Play Episode Listen Later Jun 19, 2026 9:09 Transcription Available

Agencia recaudatoria de la Isla (Proyecto MEGA ISLA):Paypal: juliommd@hotmail.comBizum: https://revolut.me/julioqdfConsigue tu SIM de Datos de Simyo y apoya a Vidas en red: http://simyo.es/amigos.html?co... Automáticamente, ¡CONSIGUES TU PREMIO! (válido hasta 07/07/2026)En el audio de hoy os comentaré:-GLM 5.2, el nuevo modelo de Zhipu AI.

ia sim vidas datos autom agencia minimax glm simyo

Novedades Mega Isla, OpenRouter y depuración de OpenClaw

ai strategy co founders arizona vice president smart southern california blockchain playbook asian americans groceries kpis chief marketing officers vice chair zhang corporate development minimax

Play Episode Listen Later Jun 18, 2026 19:42 Transcription Available

Agencia recaudatoria de la Isla (Proyecto MEGA ISLA):Paypal: juliommd@hotmail.comBizum: https://revolut.me/julioqdfConsigue tu SIM de Datos de Simyo y apoya a Vidas en red: simyo.es/amigos.html?codeMGM=056G0J69. Automáticamente, ¡CONSIGUES TU PREMIO! (válido hasta 07/07/2026)En el audio de hoy os comentaré:-Mi decepción con M3, de nuevo.-Mi agrado con Claude Sonnet, pero ¡qué caro es!-Mejoras en la radio 2 de la Isla (en Youtube). -Mejoras y depuración en OpenClaw.

mega ia sim vidas datos llm novedades autom agencia m3 mejoras minimax simyo

英语新闻丨世界杯预测成为AI新竞技场

China Daily Podcast

Play Episode Listen Later Jun 17, 2026 4:57

As the 2026 FIFA World Cup brings together more nations, more matches and more excitement, a competition has intensified off the pitch among China's leading artificial intelligence models, which are being put to the test in predicting the results of one of the world's biggest sporting events.随着2026年FIFA世界杯汇聚更多国家、更多比赛和更多激情，场外一场竞赛也在中国领先的人工智能模型之间愈演愈烈——它们正接受着预测全球最大体育赛事结果的考验。The 23rd edition of the World Cup, featuring 48 teams, is being hosted by the United States, Canada and Mexico.第23届世界杯由美国、加拿大和墨西哥联合主办，共有48支球队参赛。It opened on Thursday and runs through July 19.赛事于6月11日开幕，将持续至7月19日。Several Chinese large language models, including Qwen, DeepSeek, Kimi and MiniMax, have rolled out prediction features, turning the tournament into a new testing ground for AI-powered reasoning and data analysis.通义千问、DeepSeek、Kimi和MiniMax等多家中国大语言模型已推出预测功能，将本届赛事变成AI推理与数据分析的新试验场。"As one of the most-watched sporting events across the globe, the World Cup offers AI companies a rare opportunity to showcase the computing power and analytical skills of their LLMs to a wider audience," said Guo Tao, a member of the Chinese Association for Artificial Intelligence and a senior expert in AI.中国人工智能学会会员、资深AI专家郭涛表示：“作为全球关注度最高的体育赛事之一，世界杯为AI企业提供了一个难得的机会，向更广泛的受众展示其大语言模型的计算能力和分析技巧。”Several AI platforms have come up with interactive campaigns.多家AI平台推出了互动活动。For instance, Moonshot AI's Kimi has launched a 1 trillion-token reward pool, allowing users to share prizes by correctly predicting match winners and the final champion.例如，月之暗面的Kimi推出了1万亿token奖池，用户可通过正确预测比赛胜负和最终冠军来分享奖励。A token refers to the smallest unit of data processed by AI models.Token指的是AI模型处理的最小数据单元。Alibaba Group's Qwen has introduced a dedicated match prediction assistant, while also offering human-versus-AI prediction challenges.阿里巴巴集团的通义千问推出了专门的赛事预测助手，同时提供人机预测挑战。However, the World Cup has also exposed the limitations of current AI models when it comes to analyzing and predicting the results of sporting events.然而，世界杯也暴露了当前AI模型在分析和预测体育赛事结果方面的局限性。For example, before the Group C opener between Brazil and Morocco on Sunday, major LLMs made predictions in favor of Brazil based on both historical data and statistical indicators.例如，在6月14日巴西与摩洛哥的C组揭幕战前，各大模型根据历史数据和统计指标均预测巴西获胜。The match ended in a 1-1 draw.然而，比赛最终以1比1平局收场。Guo said that while AI can analyze historical data and statistical models, it still struggles to accurately predict real-world results, especially in sports.郭涛表示，虽然AI可以分析历史数据和统计模型，但在准确预测现实世界结果方面仍然力不从心，尤其是在体育领域。He pointed out that soccer matches are influenced by a wide range of factors in the physical world, and such variables are highly uncertain and difficult to quantify using fixed AI models, making precise predictions inherently challenging.他指出，足球比赛受现实世界中多种因素影响，这些变量高度不确定，难以用固定的AI模型量化，因此精确预测本身就极具挑战性。The limitations of current AI models were also highlighted by Wang Zhongyuan, president of the Beijing Academy of Artificial Intelligence, at this year's BAAI Conference held last week.北京智源人工智能研究院院长王仲远在上周举行的2026年智源大会上同样强调了当前AI模型的局限性。Wang said that while LLMs have become increasingly capable of solving problems in the digital world, many challenges in the physical world remain beyond their reach.王仲远表示，虽然大语言模型在解决数字世界的问题上能力日益增强，但物理世界中的许多挑战仍超出其能力范围。As a result, the next stage of AI development will gradually shift from "predicting the next token" to "predicting the next physical state", he added.因此，AI发展的下一阶段将逐渐从“预测下一个token”转向“预测下一个物理状态”，他补充道。Asked why tech companies are rolling out AI prediction features for sports when the accuracy rate is relatively low, Guo, the expert from the Chinese Association for Artificial Intelligence, said the trend partly reflects the growing pressure of competition across the industry.当被问及为何科技公司在准确率相对较低的情况下仍推出体育赛事AI预测功能时，中国人工智能学会专家郭涛表示，这一趋势在一定程度上反映了行业竞争压力日益增大。"As competition in the LLM market intensifies, technological differentiation is becoming increasingly difficult. Companies are eagerly seeking new channels to distinguish themselves from their rivals," he said.“随着大语言模型市场竞争加剧，技术差异化变得越来越困难。各家企业都在急切地寻找新的渠道来凸显自身与竞争对手的不同，”他说。As the AI technology matures, simply competing on the size parameter is not enough, Guo said.郭涛表示，随着AI技术日趋成熟，仅仅在参数规模上竞争已远远不够。"The market is paying less attention to how large a model is and more attention to whether it can deliver valuable services in real-world scenarios and solve practical problems for users," he added.“市场越来越不关注模型有多大，而是更关注它能否在现实场景中提供有价值的服务、解决用户的实际问题，”他补充道。Hu Yanping, a professor at Shanghai University of Finance and Economics, said that LLMs and AI agents are already evolving from conversation-oriented systems into task-oriented systems, while moving beyond pretraining toward continuous learning and broader real-world perception.上海财经大学教授胡燕平表示，大语言模型和AI智能体已开始从对话式系统向任务式系统演进，同时正从预训练阶段向持续学习和更广泛的现实感知能力迈进。"Exploratory projects, such as World Cup match predictions, can help accelerate this evolution," Hu said.“世界杯赛事预测等探索性项目可以加速这一演进进程，”胡燕平说。"A capability framework built around perception, interaction, decision-making and collaboration is what future task-oriented AI agents need."“围绕感知、交互、决策和协作构建的能力框架，正是未来任务式AI智能体所需要的。”large language models (LLMs) /lɑːdʒ ˈlæŋɡwɪdʒ ˈmɒdəlz/大语言模型rolled out /rəʊld aʊt/推出testing ground /ˈtestɪŋ ɡraʊnd/试验场token /ˈtəʊkən/词元（AI模型处理的最小数据单元）statistical indicators /stəˈtɪstɪkəl ˈɪndɪkeɪtəz/统计指标

united states canada ai china mexico brazil finance artificial intelligence economics world cup companies morocco fifa world cup llm hu kimi group c guo minimax aias chinese association

Maison Solutions Inc. (MSS): Grocery Gets Smart: Maison's AI-and-Blockchain Playbook

The Water Tower Hour

Play Episode Listen Later Jun 16, 2026 25:15

Send us Fan MailIn this episode of the WTR Small-Cap Spotlight podcast, Chris Zhang, Vice President of Corporate Development and Strategy of Maison Solutions Inc. (Nasdaq: MSS) joins host Tim Gerdeman, Vice Chair, Co-Founder and Chief Marketing Officer of Water Tower Research, and WTR Analyst James Kisner. Maison is a specialty grocery retailer serving Asian-American and other ethnic communities, operating HK Good Fortune stores in Southern California and Lee Lee International Supermarkets in Arizona. Zhang lays out the company's tech-driven transformation across four pillars: store inventory, sales and order operations, customer privacy, and customer loyalty, with AI-powered forecasting and replenishment furthest along and perishables the first problem it tackles. He details the newly announced collaboration framework with SupplyAi and MiniMax aimed at embedding AI in everyday food-retail workflows, the direct-sourcing strategy across Asia including the Guizhou Moutai distribution agreement, and the company's Worldcoin (WLD) treasury position and early proof-of-human exploration. The conversation closes with the operating KPIs and milestones that would signal the AI and solutions strategy is working over the next 12 months.

Usando la #IA para reclamar al Corte Inglés

Play Episode Listen Later Jun 10, 2026 27:56 Transcription Available

El pasado mes de Febrero acudí a informarme de ciertos seguros en el Corte Inglés.Me esperaba una pesadilla. Sin mi consentimiento activaron los seguros y procedieron a cobrarme por seguros que no firmé, no contraté, y con datos que no les autoricé.Desamparado, sin ayuda, invoqué a la IA, en concreto a Gemini y a #Codex. Y de momento ha funcionado, automaticé tareas que de otra manera me hubiera llevado mucho, mucho más tiempo.

sin ia gemini usando laia codex m3 reclamar eci minimax reclamaciones corte ingl el corte ingles

第 305 集 - 現在聲音也用 ai 了。老高把自己活成了傳奇。人腦貴在便宜。Grok 真的很兇。

【工程師聊什麼】

Play Episode Listen Later Jun 7, 2026 34:57

工程師都宅宅的不太會講話？其實工程師的幹話多到你聽不下去！ ------ 加入粉絲團留言互動！ https://www.facebook.com/%E5%B7%A5%E7%A8%8B%E5%B8%AB%E8%81%8A%E4%BB%80%E9%BA%BC-109229084578194 ------ softwaretalkthreesmall@gmail.com -- Hosting provided by SoundOn

ai hosting aws grok tts soundon minimax

#247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3

Let's Talk AI

Play Episode Listen Later Jun 6, 2026 105:02

Our 247th episode with a summary and discussion of last week's big AI news!Recorded on 06/03/2026Hosted by Andrey Kurenkov and Jeremie HarrisFeel free to email us your questions and feedback at andreyvkurenkov@gmail.com and/or hello@gladstone.aiRead out our text newsletter and comment on the podcast at https://lastweekin.ai/In this episode:Anthropic released Claude Opus 4.8 with improved benchmark scores, discussed eval-awareness findings and welfare/corrigibility themes from its system card, and introduced Dynamic Workflows for long-running multi-agent tasks.Microsoft unveiled the always-on Microsoft Scout assistant built on OpenClaw plus new in-house MAI models (including MAI Thinking 1) and “frontier tuning,” emphasizing enterprise security architecture and model-from-scratch capability.Major business moves included Anthropic's $65B Series H at a $965B valuation alongside an IPO filing, a JPMorgan analysis arguing OpenAI needs major revenue growth to justify infrastructure spend, and Cognition raising $1B at a $25B valuation.Policy and security highlights covered Trump's voluntary pre-release government testing framework for powerful AI, Meta AI support being exploited to hijack Instagram accounts, tightened US Nvidia export controls and China's travel approvals for AI experts, plus expanded Glasswing/Mythos-style cyber and biodefense initiatives.Timestamps:(00:00:10) Intro / Banter(00:04:10) Sponsors(00:07:10) News PreviewTools & Apps(00:07:54) Anthropic releases Opus 4.8 with new 'dynamic workflow' tool | TechCrunch(00:22:37) Microsoft Scout is a new AI personal assistant built on OpenClaw | The Verge(00:26:55) Microsoft launches new MAI family of AI models at Microsoft Build | Mashable(00:37:43) Robinhood now lets your AI agents trade stocks | TechCrunch(00:40:49) OpenAI launches new Codex tools for white-collar work | TechCrunch(00:43:40) ElevenLabs' new music-generation model can switch genres mid-track | TechCrunchApplications & Business(00:44:35) Anthropic Hits $965 Billion Valuation, Surpassing OpenAI - WSJ(00:45:32) Anthropic Files to Go Public, Setting Stage for Huge I.P.O. - The New York Times(00:51:15) China's ByteDance Developing New AI Chips Like Those from Nvidia Partner Groq(00:55:00) Anthropic expands Mythos to 150 additional organizations(00:55:35) OpenAI needs a 26x revenue increase to justify its buildout(00:58:46) AI coding startup Cognition raises $1B at $25B pre-money valuation | TechCrunchProjects & Open Source(01:00:50) MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost | VentureBeatPolicy & Safety(01:06:08) Trump Signs Executive Order Seeking Oversight of A.I. Models - The New York Times(01:11:45) Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It Worked(01:13:058) Chinese AI experts in private firms now required to secure approval before international travel — Beijing enforces policy to secure top-tier talent, expands measures beyond government(01:17:53) U.S. Tightens Controls on Nvidia AI Chip Exports | Let's Data Science(01:21:47) OpenAI launches Rosalind Biodefense, offers federal agencies early access to its life-sciences model(01:24:00) Using LLMs to secure source code(01:26:19) Project Glasswing: An initial update(01:29:30) White House Approves $9 Billion for Spy Agencies to Catch Up on A.I.(01:32:11) US Law Enforcement Warns of ‘Anti-Tech Extremism' as AI Hatred GrowsSynthetic Media & Art(01:35:38) YouTube will now automatically label AI videos | TechCrunchResearch & Advancements(01:36:22) Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention(01:41:26) From Simulation to Enaction: Post-trained language models recognize and react to their own generationsSee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

UFO 50: Was I Wrong? | Postgame Show

Fine Time

Play Episode Listen Later Jun 4, 2026 38:01

Andre looks back at his previous UFO 50 episode and examines how his feeling have evolved on games he didn't initially like very much. He dares to ask the bold question: Was I Wrong? Andre: @pizzadinosaur.fineti.me Fine Time: @fineti.me [00:00] Intro and Premise [02:46] Ninpek [07:07] Rail Heist [10:22] Velgress [13:27] Mini & Max [16:36] Caramel Caramel [21:48] Cyber Owls [25:35] Campanella 2 [30:10] Star Waspir [35:48] Thanks For Listening!

ufos postgame premise campanella minimax was i wrong

The $4 Trillion AI IPO Wave Is About to Break

AI Inside

Play Episode Listen Later Jun 4, 2026 71:41

Jason Howell and Jeff Jarvis open on the biggest week in AI yet: Anthropic closed a $65 billion round at a $965 billion valuation, passing OpenAI, right as OpenAI crossed 1 billion monthly users and SpaceX, Anthropic, and OpenAI all line up to go public. They get into what a $4 trillion IPO wave means for the market, plus Claude Opus 4.8 and Anthropic's Mythos expansion. Also in this episode: Google lets publishers opt out of AI search, Microsoft floods Build 2026 with seven new models and an always-on agent, Nvidia's RTX Spark aims to reinvent the PC, companies start rationing AI as costs explode, ElevenLabs ships emotion-preserving dubbing, plus math, robots, a Meta chatbot hack, MiniMax M3, and Trump's scaled-back AI order. Find every episode at aiinside.show. Note: Time codes subject to change depending on dynamic ad insertion by the distributor. 0:00 - Start 0:03:48 - Anthropic raises $65B Series H at a $965B valuation, overtaking OpenAI 0:15:50 - ChatGPT app hits 1 billion monthly active users in record time, data shows 0:17:09 - Anthropic launches Claude Opus 4.8, its most honest model yet 0:21:46 - Anthropic expands Mythos to 150 additional organizations in more than 15 countries 0:29:48 - Microsoft Build 2026 keynote: seven AI models, MAI-Thinking-1, Project Solara, and a Copilot super app 0:31:17 - Inside Microsoft's Project Solara: A new platform for devices that run AI agents instead of apps 0:40:54 - Nvidia announces the RTX Spark Arm chip at Computex 2026 0:47:05 - Amazon kills internal AI leaderboard after employees gamed it with pointless tasks 0:48:18 - Uber caps monthly employee AI spending at $1,500 per tool amid soaring costs 0:50:32 - ElevenLabs launches Dubbing v2, preserving emotion across 90+ languages 0:56:27 - AI startup Shift offers free NYC home cleaning to collect robot training data 0:59:31 - As A.I. Makes Strides in Mathematics, Mathematicians Urge Caution 1:01:00 - Hackers used Meta's AI support chatbot to hijack high-profile Instagram accounts 1:03:23 - China's MiniMax launches M3, rivaling Claude Opus 4.7 at $0.12 per million tokens 1:04:19 - Trump signs a scaled-back AI executive order Hosts: Jason Howell and Jeff Jarvis Download and subscribe to AI Inside in audio and video: https://aiinside.show/ Support the podcast on Patreon for special perks: https://www.patreon.com/aiinsideshow. You'll get ad-free episodes, members-only Discord, T-shirts and stickers you love, and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Learn more about your ad choices. Visit megaphone.fm/adchoices

The 4 Modes of Working with AI, The Transformation Paradox, & Building a Learning Organization | 137

Sidecar Sync

Play Episode Listen Later Jun 4, 2026 62:36

Send us Fan MailIn this jam-packed “mini” episode, Amith Nagarajan and Mallory Mejias break down a whirlwind of recent AI model releases—from Anthropic, Alibaba, Microsoft, and beyond—and what they signal about the rapidly evolving AI landscape. Then, they dive into Microsoft's 2026 Work Trend Index Report, unpacking the “agency equation” and what it really means for organizations navigating AI adoption. From the rise of agents and the four modes of working with AI to the growing gap between employee readiness and organizational culture, this episode explores why AI transformation is less about tools and more about leadership, systems, and mindset. Plus, they introduce the concept of “owned intelligence” and what it takes to become a true learning organization in the age of AI.

MiniMax la tarifa plana OpenClaw

Courtside Financial Podcast

Play Episode Listen Later Jun 3, 2026 26:04 Transcription Available

Agencia recaudatoria de la Isla (Proyecto MEGA ISLA):Paypal: juliommd@hotmail.comBizum: https://revolut.me/julioqdf

ia hermes tokens agencia tarifa plana minimax

Interviewing For A Job At Anthropic? DON'T Use AI.

Techmeme Ride Home

Play Episode Listen Later Jun 1, 2026 21:46

Nvidia unveiled the RTX Spark, an Arm-based consumer chip family built with MediaTek on TSMC 3, plus a DGX Station desktop that runs 1T-parameter models. Intel detailed its Crescent Island GPUs, MiniMax launched a coding model rivaling Opus 4.7 at 1/40th the price, and Anthropic bans AI in interviews. Nvidia announces the RTX Spark, an Arm-based consumer chip family it calls "the most efficient PC chip ever built", made on TSMC 3 in partnership with MediaTek (The Verge) Intel details its Crescent Island data center GPUs, built on its Xe3P architecture and using LPDDR5X memory instead of HBM, calling them "built for agentic AI" (Tom's Hardware) Nvidia unveils DGX Station for Windows, a desktop PC powered by a GB300 Grace Blackwell chip with up to 748 GB of memory, capable of running 1T-parameter models (SiliconAngle) Chinese AI developer MiniMax debuts M3, a new coding model that it says rivals Claude Opus 4.7, costing $0.12 per 1M input tokens, compared with $5 for Opus 4.7 (The Information) A look at Anthropic's hiring process, which prohibits AI use in interviews and features a culture interview that candidates describe as highly intense (Bloomberg) Learn more about your ad choices. Visit megaphone.fm/adchoices

ai pc windows intel arm nvidia gb anthropic opus use ai gpus tsmc m3 1t mediatek minimax hbm

NIO's Swap Network Delivers 16% Of All EV Energy In China | The Bull Case Nobody Is Talking About

Play Episode Listen Later Jun 1, 2026 9:51

Four stories to close out May — starting with the mostunderrated bull case in the NIO thesis right now.NIO's battery swap network delivered 16% of all EV chargingenergy in China over a five-day period. Not 16% of NIOvehicle energy. 16% of ALL electric vehicle charging energyacross the entire Chinese market. From one company'sinfrastructure. In five days. The market is pricing NIOas a car company. The swap network is becoming an energyinfrastructure business — a recurring revenue platformwith a flywheel that gets stronger with every vehicle sold.Gen 5 swap stations arrive mid to late June, unifiedacross all three NIO brands for the first time. The numberthat is already 16% goes higher.Software stocks closed May as their best month since 2001.The SaaSpocalypse — the fear that AI would destroy SaaS —didn't happen. Dell surged 33%. Snowflake 36.5%.AI is making software more valuable not less — selectingfor the companies that use AI as a feature rather thanrunning from it as a threat.SpaceX trimmed its self-assessed valuation to $1.8 trillionand is still on track for the largest IPO in world history.MiniMax — the Chinese AI company recently compared toDeepSeek — is pursuing a China listing. SoftBank announcedplans to invest up to €75 billion in French AI data centers.The AI infrastructure arms race is now a global story.The US and Iran are "mostly agreed" on a 60-day memorandumof understanding — ceasefire extended, Hormuz reopens,nuclear talks begin. The deal still needs Trump's signature.Oil closed May at $92.56 — down nearly 19% for the month —worst monthly performance since COVID. Watch Monday morning.If the deal gets signed the market opens at a new recordand everything that's been held back by the macro overhangsince February starts to move.

官方预告MiniMax M3系列AI模型即将登场

Play Episode Listen Later May 27, 2026 0:32