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 understand.

Nuno

Anyway, best of wishes to both of them. I’m not saying this facetiously. Getting married is a wonderful sacrament as well. I wish that they’ll have a little bit of time off.

Bertrand

It’s a tough time to get married.

Nuno

It’s a tough time. It’s a tough time. We’ll wish them all the best in a very thoughtful way.

Bertrand

I’m not sure I’ve seen the pace of launch of new models, closed and public open source, moving that fast on the frontier side. It’s pretty insane what has happened the past few weeks. At the same time, I’m not sure I have seen regulatory actions happening so fast and in some ways, randomly to the space. We have seen security breaches. We have seen multiple open letters now from all the biggest tech companies in the world in a matter of days. It’s pretty insane. I think it’s a pretty clear reflection on the pace of AI, on the opportunity on one side, on the risk on the other side at every level.

I will say personally, I’m a very big optimist into all the positive that AI can bring from intelligence brought to the masses, from acceleration of knowledge of science. Personally, I’m pretty excited, but I’m also worried that overregulation can hurt from a security perspective, but also research and development perspective, and slow down what could be done. Very interesting times.

Nuno

Conclusion

To conclude, this has been an incredible few weeks, incredible 2 months. The pace of everything that happened from regulation to launch of models to interventions of CEOs to interventions of presidents around the world, et cetera, has positioned us with something that looks like a geopolitical war, China versus the US, and it’s getting framed under close versus open. Shockingly enough, close is the US. We’ll see what happens in the next few weeks.

The underwriting of these markets is probably the most complex underwriting I’ve seen in my career in venture capital. Really understanding where all this is going to lead us is anyone’s guess. We think there is a bubble. We’ve discussed it before. If the bubble doesn’t burst now, who knows if there’s really a bubble? Maybe we’ve gotten this wrong all along. But the next few weeks will be exciting. Honestly, a really exciting time to be alive, if not just to read the news on AI and everything happens within latest petitions by Dario or by our friend Sam. Thank you, Bertrand.

Bertrand

Thank you, Nuno.