AI and AGI

“Most of our listeners have probably heard by now that, late last week, there was an attempted attack on Sam Altman at his house in San Francisco. A 20 year old man allegedly threw a Molotov cocktail at the gate of Sam's home. No one was hurt, but according to the criminal complaint against the suspect, this was someone who had a document that identified views opposed to artificial intelligence, also had a list of names and addresses of other AI executives, investors, and board members.”

“I don't think this is going to work. Right? Like, if you vote the data center project out of your town, they're just going to go to another state or to Canada. They'll put the data centers in space. You know, the they've got options here, and I don't think this is going to meaningfully slow down or stop anything.”

“I feel like there's a certain bind here that these companies and their leaders are in when it comes to talking about some of the scarier possible outcomes of AI. I think a lot of them watched the social media CEOs claim that their technologies during the last decade would produce nothing but good for the world. And I think a lot of them took the lesson from that that, well, we have to be upfront. If we think the thing that we're building has some risk attached to it, we should be open and honest about that and not sugarcoat it.”

“What really does me is the one of the simplest things for all of us to live longer is universal healthcare, right? I went to Korea to talk to the people there. They all have universal healthcare. And every peer country of ours is way up and to the right on all the good things. And we pay double the amount of money, dollars 15,000 a year compared to seven, six to seven. And we get, we're at the bottom of all the outcomes.”

“If you do you remember how Steve Jobs what he said when he died? The last last words? His sister, Mona Simpson, who we met later in his life. She wrote a column when he died. And, she said, he said, and I think he stage managed this, but he looked up, he had everyone around him, all his family, and he said, Wow, oh, wow. Like not giving you the way. I thought that was kind of fantastic that he stage measured.”

“Scientific death acceptance makes you live longer. Death denial makes you hateful, tiny, and dying quicker.”

“Meta is building an AI version of Mark Zuckerberg to interact with staff. According to the Feet, he is personally involved in testing and training his animated AI, which could offer conversation and feedback to employees. This character, this Mark Zuckerberg bot is being trained on Zuckerberg's mannerisms, tone, and publicly available statements as well as his own recent thinking on company strategies so that employees might feel more connected to the founder through interactions with it.”

“I have a working program now that I use to draft email replies. Unfortunately, they're way too agreeable. They keep trying to, like, get me to agree to, like, speak at things in Kazakhstan and, like, sure. I would love to, like, you know, edit your, you know, self published book about AI consciousness. Sounds great. Sign me up. And I have to go in and edit and be like, well, sorry. I can't do that.”

“I think it's a great way to think about the world. It's, like, very principled, very, like, hard nosed scientists, very careful. And I don't know. I think it's just it's such an incredible field. You have such high leverage in computer science, in AI. And so I think a lot of physicists were seeing that, particularly in, like, high energy physics. After the discovery of the Higgs, I think a lot of high energy physicists were sort of looking for what's next.”

“Well, so the goal was we need to come up with some productionization of GPT four. And there's questions about, like, how do we turn this incredibly powerful model into products? And we're all spitballing ideas like writing bot, coding bot, you know, very natural at the time. Some of our least interesting ideas were a meeting bot. So it would just sit in a Google Meet, take notes, send and send out, like, to dos after. But John Shulman was very opinionated. He's like, we think we should keep it very general. Let's do a chatbot.”

“One of the engineers on our team was looking at a reported material, property, and it was just sort of extracted values from literature. And It was really interesting to see the reported value spanned many orders of magnitude. And so you train a ML system on that and it's like, well, the best you can do is model this distribution, but you're no closer to, like, a ground truth. And that's where experimental data comes in, where you now have a grounding in this.”

“Yeah. So, language models are incredibly powerful. It's a very natural interface, and so we continue to use these. But we think about them almost as like an orchestration layer. So that's sort of a a copilot assistant, but also like a system that can direct, experiments. And it's almost it's orchestrating other specialized models as well. So we do construct neural nets that are specially designed for atomic systems where there's, like, some symmetry awareness, and those have much lower latency and they've been, like, fine tuned for that.”

“I think one fallacy is thinking about intelligence as a scaler. We've consistently seen these systems have a very odd spikiness. And it's actually possible to architect a system that is world class on some math domain, but then you could do some perturbations to the questions and actually degrade it substantially. So it's like a bad high school student. And so there's this, like, odd spikiness to these systems.”

“So these systems have become so incredibly impressive on this on this domain as a result of huge amounts of data, really cheap verifiable environments, like, you know, you can check Unitesco from failing the passing with just a few CPUs. It's basically instantaneous. There's no domain expertise gap between an AI researcher or software engineer. And, obviously, this will become and is becoming a larger contributor to the next generation of the system.”

“The one of the reasons I ask is, I used to run this company, Color, and we built our own liquid handling robotic systems. We buy liquid handling robots, but then we have to adjust them dramatically. We had, like, cameras that would use ML to monitor the system and sort of make adjustments. We had to three d print parts to decrease vibrations on the platform because we were dealing with such small, volumes of liquid. And so there's enormous amounts of customization versus just having and the firmware for it was awful and writing against that was painful.”

“GPUs are so extraordinarily expensive. And what's interesting is just the compute cost relative to physical infrastructure is actually surprising where, you know, so much money is spent on the compute, that the physical infrastructure sometimes is actually lower, but, you know, has very large lead times and there's intrinsic difficulty of having these well calibrated, well functioning physical systems. But from a capital perspective, it's primarily a compute cost.”

“In the 1930s, there was a really big debate about, like, what's the right construct for the banking system and the financial system. And, there was a a proposal from a group of economists, called the Chicago Plan, and, the kind of ringleader was a Chicago economist. Actually, it might have been a Yale economist or Princeton at the time, but, Irving Fisher, who wrote a book called A Hundred Percent Money. And that idea was that full reserve money was essentially, you know, government obligation money. So you have kind of a full reserve, and you can only lend full reserve money.”

“We actually see it used, you know, from at the very smallest end, like someone who's paying, you know, 25¢ for a digital object in a digital game that's built on a blockchain. Or even now, we're starting to see, AI agents that are paying for, the output of essentially the AI tokens of another AI agent, and they're, you know, spending, again, just, you know, a dollar 50¢, 20¢, etcetera. So super tiny transactions at one end all the way to the largest electronic trading firms in the world that do huge amounts of of capital markets activity who are, you know, settling multi $100,000,000 transactions. And the powerful thing is it's all the same.”

“Blockchains are operating systems, and they have compute engines. They have virtual machines, and you can write Turing complete code. You can write software that runs on these, but there's some really key attributes that make them different. So the the first is that the code is is sort of tamper resistant. The second is it's perfectly auditable. You can audit every single input and output of that machine, of that code in real time.”

“In that world, we need a different infrastructure for the financial intermediation layers. Why? Well, we don't have an infrastructure that can support that. We don't have an infrastructure that can, work globally, interoperably, instantly, that can be, programmed, through software layers by arbitrary pieces of software that doesn't exist. We need an infrastructure where the agents themselves can, dynamically create and spin up, different kind of financial endpoints themselves. We need transactions that can scale potentially into the, you know, billions or trillions of transactions.”

“The other piece is that it's sort of designed with real money, as the foundation. So there's not like a volatile gas token. USDC is actually the default native token, which is now, under the law, essentially, like a legal form of electronic money. So you have real dollars as the way that people understand. And so to a company that's, like, doing this, it's like, I pay AWS credits. I understand how to budget for that, my treasury, my operations, my compliance, etcetera.”

“If I use as a reference point, like, you know, I spent a long time building on the early Internet and the early nineties and the early web and, like, all this stuff all the way up until, like, 2001 for, like, you know, for me, it was, like, ten years. And it was still it was, like, awful still. Like, it just like kept grinding and it was like, how do we make this useful? And then you had, you know, a whole bunch of things happen that were in the background, like, you know, Wi Fi, broadband, you know, you finally got like usable, other Internet connected devices, and you could actually really start to do stuff. But it was, like, ten years in the desert or longer before you could even get there. And I kinda feel that way about the blockchain space.”

“I'm very proud because, there there are tokenized, stocks that are out there, and the most, active tokenized stock today is not Tesla. It's it's not the S and P index. It's actually Circle. So that was cool to see. We also, you know, have seen this growth in, like, tokenized money markets. So basically, like, on chain treasury bills, we actually operate the largest tokenized treasury, product called USYC.”

“Proof of work, obviously, it was itself an innovation and and sort of, essentially, like, the exhaust of the proof of work of Bitcoin is is just, like, the exhaust of energy consumption. And so it doesn't actually in some ways, it's waste in in a sense. The the energy is is waste. And and so I think the idea of, essentially, like, inference compute, as GPU, inference compute as proof of work. And so the work itself is the inference, and, that as the underlying basis for proof of work, cryptocurrency is pretty interesting, and would would be, you know, potentially something that could align with the kind of monetary principles of something like Bitcoin but actually, be productive, productive proof of work.”

“It does feel like, we have the potential for double digit GDP numbers in the 2030s. Like, that's that seems not unrealistic to me. Not that that's gonna be uniform all around the world, but certainly, in large large parts of the world, that seems very, achievable, based on what I see. The risk here is that GDP, effectively, like, the GDP growth is a sort of capital capturing more capital at the at the expense of of humans. Like, that's the real risk.”

“I bought the business for 5,500 notes, 7,000 with interest. If I make the payments, I keep it. If I miss a payment, they take everything away from me. What was really interesting about the story is I didn't have any money. But what I did have, because I worked there, is relationships with all the suppliers. So I got them to give me the first order on consignment. I said, I will always pay you back. This is not a favor. This is a chance for you to get your shelves filled up with your stuff that I'm gonna sell for you.”

“Now the other thing about that is if I give you the language model and it makes a mistake, you call me up and you say, hey, Bill. The language model made a mistake. And I say to you, the language model works. And you say, yeah. It works, but it made a mistake. Well, it's probably right, but it's not deterministic. And what we're learning too is people that run businesses understand that people make mistakes. They never will forgive software for making a mistake.”

“At the end of an interview, he said to me, Bill, it's a very interesting interview. You're interesting guy. The HR department's gonna get in touch with you in the next couple of weeks. And I said to him, you know, mister Fuller, I don't think you completely understand the situation, sir. And he looks at me with a tilted head, like, what's this kid up to? And I said, I haven't broken a promise to my father in twenty one years, and I guaranteed I won't come home tonight with my employee badge in my pocket. And he goes, you know, Bill McDermott, as long as you haven't committed any crimes, you're hired.”

“Let's take that cost, and then let's take the cost associated with the human capital doing that instead of something else because the platform was doing the work for you. And then let's add up the cost of the GPU factory, the business model associated with the language model company, and the tokens that will materially affect their business model. We've actually done the math on this. And so for a simple application on our platform, it would be 10 times greater in cost to try to replicate it with a language model.”

“For example, you know, 90% of our customer service cases now are managed by agents. And, you know, that that means, like, only 10% are actually involving people. And so there's a lifting and a shifting and a changing of the guard in terms of what people do for the critical thinking and the judgment calls that they have to make instead of the tactical, you know, work of just grinding out details.”

“If an agent can do it as good or better, that's an easily easy economic decision to make, which is why, you know, there'll be 2,200,000,000 of these agents entering the workforce in the next couple of years. So, you know, there's gonna be a lot more agents than there will be people.”

“You know, you've seen what we did with Moveworks, the agentic front door to the enterprise, Inveso, which is human and nonhuman identity, and now the security move with Armis on top of an unbelievable core. By the way, we integrated these businesses in twenty days. Twenty days. So a lot of companies don't have the engineering power to say they can do hard things quickly.”

“Right now, I think it's just tell me what I need to know. It's not like, let me come in and discover your problems. Well, let me bring you a solution because I have one in my pocket, and I can't wait to share it with you. It's like, you know my business. If you don't know my business, there's no conversation. But let's assume you know the business. Be very prescriptive.”

“And then you have to have a position that you take that brings something unique and variable to the equation. I always tell people, if two people are in the same room at the same time with the same opinion, one of them is redundant.”

“You know, this is the moment when leaders really matter. Because as the waters get choppy, you know, we see who's tough and who's not. And when the tide goes out, you wanna be fully dressed. And even if things don't look so good, you'll figure it out if you're fully dressed and you're ready for the battle.”

“What hasn't changed is what customers are seeking for, which is outcomes, right? Outcomes and return on their investment in order to get the things done. And, of course, now AI is an amazing technology that, again, helps to get more things done in the enterprise.”

“You can build two years ago, right, everybody build a rack service. And you could easily, with a POC, blew off everybody's, you know, the CEO's socks and, like, look at how easy it is to build a chatbot on 10 documents. But that but but SAP and and also these large customers, right, they always have a problem of scale. Okay. What do you now with 100 documents? Well, it becomes a little harder. A thousand documents becomes a different engineering challenge.”

“Do you still remember in a when I was a computer science student where the Google guys came in in a in a lecture, and they said, like, hey. I can go home at 5PM because I wrote my tests. And, of course, this was non you still remember that, like, test first or test driven development? It's coming back. The reality is nobody did it. At least I never did it because, hey, it was so much more fun.”

“I always call this the tribal knowledge, the stuff that is not in the system stored somewhere that just lives in people's heads, so maybe in Slack channels, maybe in Teams channels, maybe it was just a discussion on the phone. So how can you drive a decision from that? And then so the question is the agent needs to come back, ask you for input. Now you wanna store that. And now what we do, in the past, we called this process mining. Now we call it agent mining because it will record all these decision traces, these contexts of what the users are entering into the system.”

“Now if you want to do these predictions, quite frankly, then the challenge is large language models are not made for this. The way how they, you know, generate just one token after another essentially in a sequence to sequence modeling. I mean, they're language models. So that and they do this phenomenally well. But if you still wanna do these predictors where you have to go back to these classical machine learning approaches, right, you use XGBoost or AutoGluon and and many of these, AutoML approaches.”

“Usually, I say the the the primary problem, as I said, is is is the problem of a data. Because most of the time, the data is, of course, very disaggregated in a in a in a company. Either because you made certain decisions, how you purchased, solutions in the past, or you did an m and a. So you acquired a company naturally. Of course, they bring a very different IT system landscape as well and so on and so forth.”

“Like like everybody who works today maybe in the finance shared service center. It's for me the equivalent of a junior developer today with Cloud Code. So now they actually become they've got one level higher. They're now not so much anymore, tasked with then writing a lot of the code. With with with codex or with Cloud Code. But they actually then start supervising the code, give feedback, right, and capture, of course, the essence of what the code should look like and then, you know, do much more review and then rather think about what to build next.”

“I mean, for the most part, SAP software is seed based licensed, today with a few exceptions like a Conker or Fieldglass, for example, or the business network. But, you know, very clearly with AI, it was very clear for us that, you know, step by step, it will go towards this consumptive world. At first, consumptive. And then maybe in the next step, once we have more verifiability in the system, then also towards maybe an outcome based license model, to, for example, what Sierra is doing.”

“What I learned actually, and I did this mistake probably more than anybody else in this world, is to kind of pitch the technology. And this is completely wrong. When I sit together with CFO or a CIO, the first question is like, hey. What's top of mind for your business? What are what are current what are your current challenges? And then work backwards to the technology. And that always I always found that this is the most useful approach.”

“What we are focusing on is the optimization domains, obviously. And then if you go into, think logistics, traveling salesman problems, knapsack problems, like all these kind of usual, hard problems in computer science, These are interesting problems where we where some peep where we believe that could be interesting for the future, for maybe different kind of computing paradigm to solve for. Because if you can obviously load your trucks, right, and the and do the route planning even more, the the outcome is, say, the emissions go down, right, and you save a lot of money.”

“If you need to scale up these models and they are actually getting into production, the price per token or the what's needed per token becomes the key metrics that you care about. And so what we're seeing with the fusion language models is that they scale better than autoregressive models at inference time. They're cheaper to serve. They're faster. You get more tokens per GPU, which means that the price is actually lower.”

“But if you think about text and you take two words, then it's not clear what's in between the meaning of two different words. Right? And so there is no real geometry to the space of possible tokens or possible words. And so that makes the idea of denoising much more challenging because there is it's not clear what it means to perturb, add noise to to text.”

“One that works pretty well is basically one where you, mask out tokens. So you you kind of like, hide them. You you take a sentence and then you remove some of the tokens. You hide them from the neural network, and then you ask the neural network, can you predict what those tokens were? And so it's similar in some sense to next token prediction, except that things were done out of order, and the network needs to be able to use information from you needs to use context to the left and to the right and combine it in some interesting ways to figure out how to predict all these missing tokens from the from the sentence.”

“The latest model that we announced this week, Mercury two, is actually matching in quality, some of the best speed optimized models from Frontier Labs. So we'll think about the Haiku models, the flash models, mini models from OpenAI. So it's the at that quality level. But, again, it's about five to 10 x faster in terms of, like, the time it takes you to get an answer, using a diffusion model versus an autoregressive model.”

“The the real challenge is that you're like the the attention mask that you use in a traditional autoregressive model is causal. So the model only knows how to use context to the left as it figures out how to what to do next. And in a diffusion language model, you really wanna be able to have access to the context to the left and to the right as you decide what to change. It's like one of the key properties that make these models potentially much higher quality than compared to autoregressive models.”

“I think one of the reasons why, there are still no other providers that are able to serve diffusion language models, in production today, you cannot run a diffusion language model on existing serving engines. So if you think about BLLM, SG Lang, TensorRT, these frameworks that exist and and not even open source, and and they are really, really good at serving, other aggressive LLMs very efficiently. The space for diffusion language models is much, much, less developed, so we had to build our own serving engine.”

“A diffusion model, at least for images, diffusion models are are known to be, much more suitable for controllable generation. And the reason is that because the object, let's say the image that you're generating is, sort of, like, available to the model from the very beginning, it's very easy for the model to check whether or not this object that it's generating is consistent with, say, some constraints or some kind of, some kind of, like, control signal that you wanna use to to make sure that the output is consistent with whatever you want the model to generate. So I was on some papers where we're doing medical imaging, and and the idea is that, you know, when you do a CT scan, you're basically taking some projections of your body cross section, and then, you know, you're trying to reconstruct what your body looks like from some measurements that you get from the machine.”

“We're already seeing, a lot of usage. I mean, you nailed the two main ones that we're seeing, voice, a lot of voice, customer support, the educational kinda like agents. People love the speed of the of diffusion language models. They always have this issue that they would wanna be able to use a thinking model, like a reasoning model, but usually, the latency is just not enough. And so maybe they use unless they use specialized AI inference chips, but that's too expensive and they cannot scale to large volumes. So we had a bunch of, customers that are building voice agents on top of the fusion language models.”

“My sense is that, you know, there is a big switching cost. Like, they're very, very focused on Gemini on the on their main model. And so, you know, it could be a that that's kind of, like, the issue with these big labs is that, you know, they're only in one direction, and then it's hard for them to really focus on on an alternative direction. As a start up, we're in much better positions to do that because we, you know, we're laser focused on one thing, and we can really deliver and and build everything that's needed to get that, technology to succeed.”

“Chat Concierge for us was, our beachhead initiative around deploying multi agentic deploying a multi agentic solution. So, Chet concierge is essentially a a auto deal dealership, project or application that was deployed out to our auto dealers to basically bridge that experience between dealers and their customers and make it very seamless. And we need to understand we are moving to a world where the car buying experience doesn't start at the dealership. It starts before. It starts when they go to their website and try to figure out, okay, what's the inventory?”

“We moved from a classic ML world to a world where we have LLMs, generating responses. And now we want to move on to a world where actions need to be taken, specific goal oriented actions need to be taken. And when the problem that we are working on is a complex one, with multifaceted, aspects associated with it, That's where multiagentic comes into place. So, basically, we have a large complex goal which we have to break down into specific steps, and each step is basically narrowed to a specific agent.”

“So what in the past, what used to be thought of as non functional requirements such as latency today is product feature. It is it is baked into the experience of a developer. So these are some things that, you know, we are seeing a paradigm shift in terms of what we need to bring to the fore to the developer experience to keep in mind when you're when you're implementing your systems.”

“I'm thinking about an example that we, you know, we've all heard about. I forget the the very specific scenario, but, a business had a chatbot. I think it happened in Canada, had a chatbot on their website, and the customer asked for a discount, and the the chatbot basically gave them a discount. You know, this would clearly be disastrous in a, in a car dealer type of scenario. Like, how do you, make, you know, generative AI and agents safe, you know, for, you know, these dealers that have a lot at stake?”

“When you, you are the most successful when you can offer two things, reasoning and specialization. So reasoning capabilities with our agentic platforms, a agentic frameworks in the platform, we are bringing that to the fore. Specialization is something that is very, very crucial. So this can be achieved primarily using, specialized models and fine tuning. Student teachers student distillation gives you that control you need to have on, providing personalized experiences as well as providing, having some control over your latency metrics.”

“All the more important that observability comes to the forefront in stochastic systems like a multigenic application. All the more important for us to be able to replay agentic actions and try to understand how it function. Agent behavior needs observability along many different dimensions in terms of what are the tools involved, what was the reasoning mechanism that led to that tool invocation. And, overall, what was this context that passed across systems?”

“I think the core is that, you know, you really need to treat agentic AI as a system. It's truly a system. You have to start with governed data. You have to kind of put in that risk controls baked into multiple layers of your, of your application or your system. You have to look at, latency as something that needs to be optimized end to end. And, understanding that your biggest gains do come from postproduction telemetry is also critical.”

“So, I think, customers that are new on the multi agentic journey are one of our more naive users. So we have to offer them prebuilt blueprints that they take and run for their use cases. There are savvy customers who know what they want to do, and you offer them those developer kits. So we have all of these different ranges that we offer to the to our customers depending on where they come from, what their use case is, and how can they get to the fastest path to production in the least, constrained manner as possible.”

“The mission really is democratizing general AI, so that it's not just companies who can purchase a lot of GPUs, are able to create LLMs and adapt to LLMs and serve LLMs, but also people like myself and colleagues who are academics, so for example, cannot buy as many GPUs, and is there something really meaningful and fun that we could do, even with a smaller counterpart? And at the end of the day, I believe that fundamentally it should be feasible. It's only that the world has invested so much more into exploring what happens when you scale things up so much.”

“DeepSeq R1 does some amount of this imitation learning after the reinforcement learning... if you just do reinforcement learning, by the way, sometimes this model starts code switching in the middle of solving math problems. It's just suddenly speaking in Chinese and English and back and forth, or some other foreign languages that may not make sense to human readers. So reinforcement learning only cares about whether you got the final solution right or not. It doesn't care about how you got there. So strange behaviors can be emergent and then it can be even reinforced.”

“Mode collapse is a real concern with LLM generation. So, what we find in our paper is that even when you ask open-ended questions, like, you know, tell me a joke about time, or tell me something wise about time. Even when you ask, by the way, hey, give me a random number between and 10. It's not random... The bigger problem is after post-training, like sequential fine-tuning and RL, the probability, output probability of the model becomes even more skewed, like zoning in to the stereotypical answers that people tend to like.”

“There's a lot of delving now that wasn't happening before. Yeah, probably, yeah. You know, actually, whenever I see the word delve in anybody's writing, I'm like, hmm, what did you do?”

“I can give you one example of our recent work called the prismatic synthesis. It's a synthetic data generation algorithm which is prismatic because it acts like a little bit like a prism that can scatter the light to make it more diversified... we're doing this using Dipsic R1 32 billion parameter model as the teacher model... our goal is to compete against the alternative, which is to use much stronger teacher that's 20 times larger... we find that that one million data points is actually better than the one million data points that you generate from the stronger teacher model, the best teacher model.”

“The idea is that during pre-training, the model is forced to be completely passive in the way that it learns to predict which token comes next. But what if we encourage the model to think for itself? Before predicting the next token, what if we encourage the model to think for itself by generating something like a chain of a thought? And then predict the next token... the key idea of our approach is to make the reward information gain of predicting next token with thought compared to without thought.”

“There must be a better way of it, fundamentally better way of doing this. And can we find it? In some ways, the nature found a solution, which is the human brain. The nature found the solution, and human brain requires so little energy. Our brain apparently use less energy than one light bulb.”

“Overtone pluralism means when you ask a question that's politically thorny, for example, that could have different answers, the best way might be for LLM to just present all of them. All of the reasonable opinions as, hey, the answer is that people have different opinions. Here's one view, there's another view, and be able to include all of them as opposed to picking the majority opinion. Because that marginalizes out the rest.”

“It has been an absolutely crazy six weeks for the creator of NanoClaw by Gavriel Cohen. He created this tool basically in 48 hours on his couch, and it has now led six weeks later to going completely viral and creating and having a deal with Docker. So today on the podcast, I want to break down his story, how he built this product, what it does.”

“A few weeks later, after he made, you know, a post, Andrew Carpathy, of course, the famous AI researcher, was saying, you know, like, hey, this thing's pretty cool. It went super viral when he did that. He posted on X. And this basically put it in front of thousands of developers. And from all the attention that that got, it got more than 22,000 GitHub stars. It had 4,600 forks and dozens and dungeons of contributions and collaborations on like new features that people wanted to add to it.”

“while he was kind of looking at some of the performance issues, he realized that OpenClaw's agents had downloaded all of his WhatsApp messages and then stored them locally as plain unencrypted text, which right? I mean, these are WhatsApp is famous for being encrypted and now you got them all as unencrypted texts, you know, just stored on your device. And so I think not just, you know, his work messages, it had authorized that he like told it, you know, it had authorization to access. Apparently, it had actually went and downloaded his entire messaging history.”

“He took what was 800,000 lines of code and he brought it down to 500 and was really just trying to create an absolutely minimal and secure alternative to OpenClaw. And so that's what he built with NanoClaw. I mean, that is insane. 800,000 lines of code down to 500 It was a super stripped down framework. It was written very, very concise. And so instead of relying on this kind of massive dependency tree, it used containerized environments that would isolate AI agents and then strictly control what data they could access.”

“last week, Cohen actually shut down the AI marketing startup that he launched with his brother Lazar. And he is focusing exclusively on NanoClaw. So he had something else going, realized this thing had so much momentum, he shut that down. Right now, both of them are building a company around the project called NanoCo.”

“this is usually how these open-source projects go, right? Because technically open-source means they're giving the code away. Anyone can use it for free or kind of with a license. I mean, there's different ways you can do open-source. But you really are trying to give this away for other people to use. But you typically will create a company around it where you host it on your own server and usually have an API. And if people don't want to kind of run it on their own hardware, they can still get access to it through you.”

“The CFO, Sarah Friar, reportedly told colleagues earlier this year that she doesn't believe the company is ready to go public in 2026. She's specifically worried about $660,000,000,000 in projected AI commute, compute commitments and whether the revenue ramp can support those contracts. Sam Altman is pushing for a q four IPO, and Friar is publicly opposed. She's the CFO inside of OpenAI. So we have some serious head butting from the top people inside of OpenAI.”

“The information actually reported that Friar stopped reporting directly to Sam Altman last August and now reports to Fidji Simo, who runs the applications business. And Sam Altman has reportedly been excluding Friar from certain financial conversations. I think that's definitely not very normal CEO, CFO. Friar and Altman put out a joint statement calling the report, quote, ridiculous, which in, you know, CFO speak means it's basically true.”

“So first, I wanted to talk about Google DeepMind. They've just dropped Athletica this week. This is an autonomous math agent, and it's actually just built on top of Gemini three deeps DeepThink. And the results that it has been producing are really wild. So on the first proof challenge, which is basically a benchmark of unsolved or really complex kind of novel math problems, Athletica actually produce solutions that according to human expert evaluators were graded as, quote, publishable after minor revisions.”

“Sony is also coming out swinging. So they have something called Sony AI, and they have a project called Ace. And they just published on this their their project was just published on the cover of Nature, one of the, you know, top journals for science this month. And it is the first known autonomous robot that beats elite and professional level human table tennis players in real matches. Ping pong, I mean, I am I'm no expert, but I do love ping pong. And that is a game that if you see the pros play, the ball is flying so fast, is moving so quickly, and the precision has to be insanely high.”

“Google is now kind of doing the opposite. They signed the contract that Anthropic walked away from, and they essentially allowed Gemini to be used for, quote, any lawful government purpose. Gemini 3.1 is already on the GenAI MIL platform, which is about 3,000,000 Pentagon staff. So Google's own DeepMind people are basically watching the company, take the deal that Anthropic refused.”

“The headline for all of this is that the 08/02/2026 enforcement deadline for high risk AI systems under the EU AI act, it's getting pushed back. What they're saying is that stand alone high risk systems classified under annex three now have a 12/02/2027 compliance date. And AI embedded in regular products under Annex one gets pushed all the way to the 08/02/2028. So we are pushing this years out. The industry basically just got a sixteen to twenty four month deferral on the rules that they were supposed to be ready for in three months.”

“China has just blocked Meta's $2,000,000,000 Manus deal. Meta has already went ahead and acquired this company called Manus, which technically, to be fair, was it was started in China. The founders are Chinese. They realized that China was pretty restricted in this. And so they moved the company to Singapore. When they moved it to Singapore, Meta said, hey. We'd love to acquire it. They bought the company. They paid $2,000,000,000, and they have, like, deeply embedded it.”

“Together, I think these are basically the four companies that are committing about $700,000,000,000 in CapEx for this year, and almost all of it is in AI infrastructure. Dan Herbach check over at Ramsey Theory Group put out a note today saying that Wall Street's AI boom is heading into a, quote, ROI reckoning. Alphabet's q one EPS is projected to drop 6.4% year over year despite revenue growing to 106,900,000,000 because the 175 to 185,000,000,000 in AI CapEx line is grinding their margins.”