314 | Top AI Labs begs Washington to slow them down, Altman says Intelligence is a Commodity Multiple models ship, and more important AI News, week ending July 31, 2026

27:30To realize AI potential, industry, government, and society at large may need the option to buy time to address emerging risks. Development security measures and strengthen oversight. But each company and country is under intense competitive pressure not to unilaterally slow that acceleration. And today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress. Building on work already underway to monitor frontier model releases, we request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.

28:15Now, as you know, if you've been listening to this podcast, I strongly agree with the motion, right? I've said multiple times for a very long time on this podcast that I'm praying for the day that this will happen. Not that the US government will take control, but that there will be an international collaboration of everybody that needs to be involved in order to figure out where this can go and how can we get there safely so we can enjoy the benefits and avoid or at least reduce the risks and the negative outcomes. So I will start with that. I strongly agree with the general direction. Let's see who signed the letter.

28:51So initially there was just over 1 ,100 signatures. Now there are over 1 ,300 signatures that include some of the top leading people in the top leading labs in the world. So the list includes chief scientists from the top labs of Jakub Pachoki, who is the chief scientist of OpenAI, Mark Chen, who is the chief research officer at OpenAI, Jared Kaplan, the co-founder and chief science officer at Anthropic, Shen Jiazhao, the chief scientist at MetaAI, Shane Legg, co-founder and chief AGI scientist at Google DeepMind, John Shulman, Chief Scientist of Thinking Machines, previously an OpenAI co-founder.

29:26So Dario Amadei, CEO of Anthropic. Ilya Saskover, the CEO of Safe. Again, previously one of the co-founders of OpenAI, founders of Senior Safety Leadership. So Woodsheet Zarimba, I hope I don't butcher his name, who is the head of AI resilience at OpenAI Foundation. Jack Clark, co-founder and head of public benefit at Anthropic. Chris Ola, co-founder and interoperability research lead at Anthropic. Benjamin Mann, the co-founder of Anthropic, Jan Leakey, Anthropic, formerly co-lead of OpenAI Superintelligence Team, and so on and so forth. There's a very, very long list of names of people who are definitely heavy hitters in the industry who knows what they're talking about.

30:06So the most important people in AI research in the world today, in the Western Hemisphere, at least most of them, signed this letter and agree with the motion. Now, a few people that are clearly not there, at least as of right now, are Sam Altman, Greg Brockman, Mira Morati, Demis Assabis, and Jan Lekul. So again, not everybody signed the letter, but people from their organizations that are senior people from their organizations signed this letter. Now, going back to connecting the dots from things we talked about in this episode, Sam Altman said in the interview that was released on the day the letter was released, meaning it was recorded before, that he believes that we need to pace the progress and yet he did not sign the letter.

30:51So that's a little weird to me. I'm also surprised that Demis Asabes, the CEO of DeepMind, did not sign it because he clearly stated exactly these things multiple times before. Now, the letter drew immediate substantial reactions across the board on multiple platforms from Ready to X to different articles and so on. And there are arguments to both sides. I must admit that I felt really negatively when I read the letter, and I'll explain why in a minute. But I will give you some examples of other people who think like me. The main arguments against the letter are the following with the most common objection, which I completely agree with.

31:28that is saying that the people that are building these systems, that leading labs are asking the government to impose constraints on what they are doing. You can just stop it yourself, right? If you think that what you're building puts humanity and this planet at risk, just stop. Don't ask the government to stop you. Don't ask for regulatory approach. Just say, what we are doing is really, really dangerous. Let's stop now and let's convince other people that what they're doing is dangerous and potentially get political pressure to then help do this. But stop yourself. Don't ask other people to stop you.

32:03Another big argument was regulatory capture, basically saying that if the leading labs are going to have the government slow everything down while they're already ahead, it is mostly going to slow down the smaller competitors and new entrants into the market, keeping the leading labs ahead because other people cannot do one, two, three, four, because these are new regulations. The third argument against it is that regardless of what regulation and handshake from an international perspective is going to put in place, it is not going to put any real constraints on China or other actors that would not be willing to share exactly what they're doing behind closed doors.

32:40And hence, it is going to leave the US behind in the race, putting us at risk of having the Chinese or somebody else in the world with superior AI capabilities that could dramatically change the balance in the world. The other interesting thing is that there are more than several different people who signed both letters. One letter is saying, don't put any constraints on open weights models. And the other letter says, let's slow everything down. And that obviously doesn't make any sense. Now on the arguments supporting the letter, there were a few also very good arguments. One is the labs are on framing that saying that they have a serious coordination problem and they cannot have an agreement just between the labs.

33:20They need somebody to help them get to such an agreement. They also clearly mentioned the US-China factor and how that puts pressure, as well as they said competitive pressure, which may mean different things. You will hear my opinion in a minute. The other clear arguments for this was the incidents that we just reported about. So you had three, four different incidents in which recent AI models and not previous AI models doing stuff that just didn't happen before. And now we have a very real problem on our hand, at least from a cybersecurity perspective. But a cybersecurity perspective means it can hack into any real infrastructure system.

33:58So think about our power systems or water systems. Everything that you know as modern life can be hacked right now and potentially influenced or completely shut down or destroyed by an AI model, whether if it's used by bad actors or without any malicious intentions like we just saw, but just trying to solve another problem and the electrical power in a specific county or state or the entire United States would be in the way of it achieving the goal and you will find a way to shut it down. So here are my personal thoughts on this after I had a few days to think about this. The first thing is one of my favorite quotes from any movie ever from the good, the bad and the ugly.

34:36If you have to shoot, shoot, don't talk. Or in this case, if you need to stop, stop. Don't send letters, it blows my mind that these people literally think that they are putting the future lives of themselves, their dear ones, and the humanity at risk. And what they're doing in order to potentially try to stop is they write a letter and they send it to the government. You are grownups. You run a company. You run some of the most successful, capable organizations in the world. You can make your own decisions. You can stop. Yes, there is very significant pressure. We're going to get to that in a minute, but you can stop.

35:13And by stopping, I think you're going to have a much bigger impact than by signing a letter. For individual people in these companies, you can resign, right? If you think the company you're working for is doing something that puts humanity at risk or that does not align with your moral values, leave the company. Find another place to work that does align with your moral code and values and where you think you should be and investing your time and your genius brain that now being used for something you believe is really, really dangerous. So this is number one. Number two, I really think that people need to grow up.

35:46And when I say people, I specifically mean Sam Altman and Dario Amadei. Those two that hate each other's guts, that won't shake hands in pictures with other leaders in the industry, and that will avoid each other and will go at each other's throats at any given point. If we got to the point that, again, both of them believe, and they both clearly stated that the future of humanity and the world as we know it is at risk, grow up. Sit in the same room, and by doing so, you will very easily attract other people to participate in the conversation. I would make a serious bet that Demis Asabes, who multiple times in the past said that he thinks the current race is crazy and it's not what AI should be used for, will gladly join them.

36:28By having these three people in the room, that will definitely attract more or less anybody else in the AI industry to join them, most likely including Chinese labs. Now, the fact that these Chinese labs will join the organization means by definition that the Chinese leadership will join the conversation because they are in and a part of these labs, which means that we still don't know what's going to be the outcome. But at least we know who's involved in the conversation and does not require the U.S. government to participate in that. I actually think having the U.S. government participate makes it a lower chance that the Chinese leadership will participate.

37:04it. I will get into more of that in a second. I also think that it is not a government thing. I think the government might be the last organization that needs to be involved. I think everybody else needs to be involved. The research companies, the leading labs, academia, people who are experts on social and political and education and economy and every different aspects of how our world operates today needs to be a part of the conversation. And in the end, including governments, because they are the ones that needs to eventually put the infrastructure and the budgets and the people and the resources in place in order to actually, once we get to a conclusion on how to do this, can actually do this right, rather than having to figure out how to make this into laws and regulations and resources and so on inside multiple countries.

37:49But I think they are potentially the last ones after a more deeper, less political conversation actually happens. Now, the only argument that I can think about why the U.S. government is required to participate is influencing China. I strongly believe that in the current situation, the U.S. government has zero leverage over the Chinese government. And we've heard that before, right? The Chinese government, if anything, will retaliate to anything the U.S. government is trying to impose on them, even if their labs that they're protecting are doing stuff that is currently stealing from the U.S. companies.

38:25So I think that having the US government in this particular case is not valuable to get China to participate, but if anything, the other way around. Now, speaking about China, nobody knows what China would do, right? They may say, of course, we'll do something completely different behind the scenes. They may say, absolutely not. We completely disagree with this motion and we don't think that what you're doing is the right thing and we're not willing to participate and we won't let our companies participate either. I have to assume that the Chinese governments do not want rogue AIs doing whatever rogue AIs can do, potentially more than people out in the West.

39:01And the reason for that is being the type of country they are, they want to control everything. And having a rogue AI do things that are not within their control is not something they want. Again, my personal interpretation. And so I do think that they may participate and definitely allow their companies like the Chinese leading AI labs to participate to make sure or at least try to control these AI capabilities from going rogue and doing the things that they want versus the things that the Chinese leadership want. Again, what they will do behind the scenes, nobody knows, but we also don't know what the US governments will do behind the scenes when it comes to military development or stuff like that.

39:43And then there's obviously the money side. Hundreds of billions of dollars was already invested in AI development and in this AI race. And there are hundreds of billions and potentially trillions in potential revenue that is supposed to come in as an outcome of these investments. This is very, very serious pressure for these companies not to stop. This is the only real argument that I can think of. You have a board of directors that represents your investors, and they are the ones that can decide whether you stop or not. And they are representing the investors who invested hundreds of billions of dollars in order to make this thing come to life.

40:22So they're not just going to say, let's stop and forget about all the hundreds of billions of dollars we invested. This is the only issue that I see here, right? At the end of the day, there are huge sons of money, nothing like anything we've ever seen before that are involved in this process. However, if we do not figure this all out, it may all come collapsing anyway. It may actually may be a much bigger collapse, colossal collapse. It will come to a very bad event that plug on everything or will lead to whatever catastrophe that nobody wants to see or even imagine. And so slowing down collectively might be the right way to actually recoup and recover these investments instead of hoping and rolling the dice or flipping a coin to an outcome that is the right outcome, not just from a moral future of a planet perspective, but also from a financial perspective.

41:20So bottom line, I think the direction is the right direction. I think signing a letter and sending it to the government is pathetic. I don't have a better word to describe it. I think that the leaders of these labs should grow up. I think they should take responsibility. And I think they should just say, we're stopping. Let's get in a room and talk about this and then see what happens. I think they need to get the buy-in of their investors. I think they will get the buy-in of their investors if they tell them the risk of what happens if they don't. So let's say something really bad happens out of OpenAI and then the government or governments around the world ban OpenAI.

41:53They say, oh, they will never do that. Well, they just did it to Anthropic just a few months ago on something much, much, much smaller. And so this is a real risk. And so I do think that there is a way to get the investors to be on board with this as well, which then may hopefully eventually lead to what I've been hoping now, which is wide international collaboration to evaluate what this may do to the way of life as we know it and figure out the way to progress in a pace that we can control versus that will control us. And now to some rapid fire items. I'm trying to be quick. There's a lot of really interesting things to cover.

42:31One of them is that OpenAI Chief Finance Officer Sarah Fryer announced to employees that the company's annualized recurring revenue in July of 2026 exceeded its total for the entire second quarter. So one month July made more revenue to OpenAI than the three preceding months before that combined together. What this means is that the very dramatic changes that are in response to Anthropic opening a growing gap on the business and enterprise revenue were the right measures to take. and closing stuff things that were more of shiny objects and focusing on developing codings and GPT work, basically copying what Anthropic was doing, is yielding the right financial fruits.

43:17This is really critical because they're planning on an IPO either later this year or early 2027, and having that growth and that strength in their financial capabilities is critical for that being successful. In parallel to that, OpenAI is slashing the rates of their two smaller models. So GPT 5.6 Terra and Luna prices are cut up by up to 80, and that is amid surging competition and the whole enterprise cost scrutiny that is currently happening. So effective July 30th, which is yesterday, the time this is recorded and the time you were most likely listening to this or some of you, just weeks after its public release, they are going to dramatically cut the prices.

44:01So the fastest, smallest model, GPT 5.6 Luna, is going to be cut in prices by 80 % to$0.20 per million input tokens and$1.20 per million output tokens. And the mid-tier GPT-5.6 Terra model saw a 20 % price reduction, now costing$2 per million input tokens and$12 per million output tokens. This is still significantly higher than the Chinese competition, but is significantly lower than it was just a few weeks ago. And to be fair, two days ago. Now, the interesting thing here is that Luna, now again, the smallest model, now delivers performance that is comparable to Frontier class models from just a year ago at roughly six cents per task, nearly nine times faster and cheaper than previous versions of a year ago.

44:49Now, the top of the line model, the GPT 5.6 Sol model price remains unchanged, but it now also offers a new fast mode that delivers up to two and a half x faster speed than the previous processing speed and it gives it to you at at the same level of intelligence as before what does that tell you it tells you that the competitive pressure is on and that if you want the frontier model you will pay for that a lot of money but if you don't you can find cheap faster alternatives and i mentioned that multiple times on this show i don't think they have a choice and the reason i think they don't have a choice because Chinese models are there too, and they're providing significantly cheaper capabilities that are not at the frontier, but that are good enough for most tasks.

45:35And I think that's what we're going to continue to see moving forward, meaning more and more cheaper open source models, fulfilling gaps, and then the frontier labs having no choice to provide similar competitive models to that. Another company that is pushing very hard in that direction is Microsoft. They are also focusing a lot on releasing cheaper models that are specifically customized to specific tasks. They just released a bunch of new AI models that are specializing in specific things, such as image generation, voice generation, transcription, coding, data security, et cetera, and all of those while reducing GPU costs by 50 to 90%.

46:14So I'm not going to dive into all the different models they released, but all of them are optimized for very specific tasks. Some of it is a task that can be used by anybody. Some of it to specific industries or specific clients that can tailor the models to their own specific needs in order to reduce errors and increase speed and reduce cost all at the same time. And these new models that are designed for specific things by OpenAI handle up to 90 % of tasks efficiently while reserving expensive frontier models like GPT 5.4 for 10 % of the overall tasks, which again, proves my point that I've been talking about for a very long time.

46:54You do not need the frontier models for most of the tasks that you're doing today. If you know how to use the existing models in the Microsoft case, it's specifically trained, but you don't even have to go that far. You can use the existing models without training them and maybe not do 90 % of the tasks. You can probably do set out a task with models that are significantly cheaper than the top-of-the-line models out there. Staying on the same topic, DeepSeek just released a new Flash model of their latest model that matches OpenAI's performance at 60 % lower cost. So DeepSeek has released version for Flash 0731, that tells it when it was released, which is aimed to its already very competitive AI model that scores nearly identical to OpenAI's GPT 5.6 Luna, again the smallest model, while dramatically undercutting it on price.

47:43So the fact that Chachupiti cut that price makes no difference. I mean, it makes a difference. It now makes it a lot closer in price. But still, DeepSeek is winning by a big spread. Another company that is playing the same game is Thinking Machine Labs. So Thinking Machines has released its Inkling model just a couple of weeks ago, and we shared that on this podcast. but now they're releasing Inkling Small, which is a quarter of the size of the previous model, but it matches it or almost matches the full version performance. And so similar approach, build a leading model and then work your way backwards from there to develop a faster, cheaper model that doesn't give you a big decline in capabilities and outcome, but that runs faster and significantly cheaper and can run on more devices.

48:34And this has been the trend in the last couple of years and it's just accelerating dramatically right now with everybody's push to save money on tokens. Now speaking on model releases, there have been a bunch of other model releases that do not focus on price optimization. The first one is Black Forest Labs has launched Flux 3. Those of you who don't know Flux, Flux is one of the most advanced image generation capability tools out there. I love it. I use it a lot across multiple things that I'm doing and now they've released a new model that is a new multi-modal foundation that is now already available for early access it also now has top tier video generation performance and its early evaluations is showing it generating videos that outperforms runway gen 4.5 and luma ray 3.2 and it's capable of generating 20 second videos and native audio from text from images from existing footage.

49:29They're also integrating their models into robotics and physical AI with what they're calling Flux Mimic, which is a tool that allows robots to predict the actions they need to do and enable them to run in a more effective way. So a highly capable AI model in the visual space that started as a model that only generated images, now generates images as the best of them, generates video better than most of them, and is now fully focused and integrated into physical AI in order to do things beyond just generating images. We talked about a similar path from MidJourney with their body scanners that they're planning to build.

50:11So again, these are companies who started with very highly capable visual capabilities that are now merging them into other applications in the real world in order to make more money and provide more revenue sources beyond just image generation, which obviously is very hard to compete in. But Flux is open source. So again, you can do a lot of cool things with it with just the image and video generation models. Staying on the visual side of things, Minimax just launched Minimax H3, which is a unified multimodal video generation model that consolidates five separate AI tasks into one compared to the previous models that they had.

50:49It unifies text, image, video, and audio into a single unified cohesive system that generates 2k video outputs in 4 to 15 seconds and again it can do it all beginning to end including the sound effects including the speech including everything including references of images and so on and from a price point based on third-party evaluations it costs about 13 cents per second generation which is actually really really cheap it's about two dollars for a 15 second full video clip in a native 2k resolution this is very competitive now we mentioned a model for robotics google also is playing in that field and now they released gemini robotics which is aimed to be integrated into advanced embodied reasoning model designed to act as a high level brain for robots enhancing their ability to understand, plan, and collaborate in the physical world.

51:46So again, Google, while not releasing Gemini 3.5 Pro that everybody has been anticipating for a very, very long time, is continuing to release models in other aspects, either smaller models, open source models. And now we see a robotics specific model that is supposed to play in the robotics world. Going back, connecting it to what we said in the beginning, saying we're going to get the Chachapiti moment in robotics in the next two years. You're seeing more and more advancements in the models, huge investments in this field across the board all around the world. So I tend to agree with Sam that we're going to start seeing very capable robots roaming in different places that we're not expecting to see within just a few years.

52:23Now, I should probably tie this to the article where we're talking about the letter to stop the frontier. But another thing that Anthropic has released this week to discover mathematical flaws in cryptographic algorithms, which are being used to protect a lot of really important stuff. So one of the tools that are being used in the cryptographic world is called HAWC. I don't know what the acronym stands for, but Claude Mythos Preview reduced the effective key strength of HAWC by half in just 60 hours. It's cutting the expected cost of a full key recovery attack by two in the power of 60 to just two in the power of 38.

52:59That is still a really, really, really large number, but it is a significantly smaller number. And it did this by finding a novel mathematical property called automorphism. I don't have a clue what that stands for, but what it means, it means that the next step might be able to crack really, really sophisticated cryptographic codes and get into basically any secure data on the planet without having a supercomputer work on it for the next seven years. And I'm going to end with two items that relate to home usage of AI with two. interesting and i must admit surprising or maybe not so surprising uh developments of things that we will be able to start buying in the near future the first one is apple is set to start selling a major smart haul overall of seven new products that are built on the new advanced series so this is a complete overhaul of the apple home series with all its different components and beyond It will include things like a refreshed HomePod, Apple TV, models, a smart display, and other capabilities, all powered by the new Siri AI, which, as you know, is running on Gemini in the background.

54:14they're talking about imminent product releases that is planned for the very very near future including a new apple 4k tv the home pad a refreshed home pod mini and all of these is expected in the very near future running on ios 20 and then the next home component is even more surprising and it is going to come from no other than martha stewart so according to its own website and recent reports, Hint, which is the name of the new system, or Hint Home Intelligence in its full name, is an AI-powered home management platform that has been co-developed by Martha Stewart and has officially launched nationwide on the Apple App Store.

54:56The platform is aiming to transform home ownership by integrating human expertise with artificial intelligence to offer personalized proactive guidance for maintenance, cost savings, and project management. So the idea is to track everything that's happening in your house, everything that you're doing in the house, everything the house needs in order to help you maintain it in a more effective way. All of that under the Martha Stewart brand. I find this A, interesting, B, exciting to see how this is going to work out. I know of several different startups who are in this field of reducing maintenance costs and then connecting it to insurance companies in order to reduce your insurance costs and stuff like that.

55:34So a big name like Martha Stewart getting into this field will be very interesting to track. That's it for this week. There's a lot of other news into the episode that you can still read by signing up for the newsletter. There's a link for that in the show notes and you can then get the newsletter every single week. It has everything that we're talking about during the episodes, but a lot more than that in additional links that you can follow and explore on just the topics that you're interested in. I want to remind you for our September cohort of the multi-agent orchestration course. We started teaching this course back in May and we had several extremely successful cohort studies, truly, truly transforming people's lives and their businesses.

56:13The things that people are doing immediately during the course while the course is still running after every session just blows my mind because they are completely changing people's lives and putting businesses on new trajectories. by applying what they're learning every single session. And as I mentioned, you can now sign up for the September cohort if you haven't done so yet. And as listeners of this podcast, you can get$100 off by using the promo code 100. That's it for today. We'll be back on Tuesday with another fascinating how-to episode that will teach you how to implement something with AI, the exact step-by-step process.

56:50And until then, have an amazing rest of your weekend.

56:57Thank you.