ChatGPT’s Sites is seriously underrated. You can connect databases, environment variables, and even your own domain. It’s becoming a place where you can build almost anything. We’re entering an age where anyone can vibe code, and you can just add what you need with plugins. This is a time when optimists have an advantage over pessimists. It’s time to build.

Jeff Dean and Demis Hassabis may be moving in opposite directions from Google’s day-to-day AI business, but the shift could be more strategic than it appears. Google increasingly looks like a company that believes current frontier models are already capable enough to support very large businesses. If that is the case, the priority naturally moves from pushing model intelligence at any cost toward productization, distribution, inference, chips, cloud infrastructure, and applications. That changes the role of frontier researchers inside the company. Researchers like Dean are still focused on how far intelligence itself can be pushed. His new company reflects that belief directly. The premise that a small number of highly committed researchers can invent more than the world’s largest research organizations suggests a model built around concentrated talent, faster iteration, and far less organizational friction. The environment now makes that structure possible. A small group of elite researchers can raise billions of dollars, secure enormous amounts of compute, and receive meaningful pre-IPO equity without remaining inside a large technology company. The gap in resources between a frontier lab inside Google and a newly formed startup has narrowed considerably, while the ownership upside outside Google has become much larger. This creates an unusual structure in which the departure of researchers does not necessarily mean that Google loses economically. Many of these new companies still depend on the same infrastructure. They raise outside capital and then spend heavily on compute, cloud capacity, and AI accelerators. Google can invest in some of them, provide TPU capacity, and collect cloud revenue as they scale. The larger the frontier AI ecosystem becomes, the larger the market for Google’s infrastructure can become as well. In that sense, Google may be evolving toward something closer to a distributed spinout model: frontier research spreads across smaller, highly concentrated companies, while the parent ecosystem captures value through capital, infrastructure, and distribution. The incentives can work for both sides. Researchers gain autonomy, speed, and equity. Google reduces the need to keep every frontier researcher inside one organization while still participating in the growth of the market those researchers create. Hassabis fits into the same transition from another angle. By stepping away from daily operations and spending more time on science, while product and commercialization responsibilities move elsewhere, Google’s internal structure also becomes more clearly divided between scientific exploration and commercial execution. The broader shift is that frontier intelligence and AI commercialization no longer have to advance inside the same organization. Google may be increasingly focused on turning existing intelligence into products and revenue, while smaller research organizations continue pushing toward more capable systems. If those organizations continue to rely on Google’s compute and cloud infrastructure, their expansion can strengthen Google’s economics rather than weaken them. The strategic question for Google may no longer be whether every breakthrough happens inside DeepMind. It may be whether Google can become the infrastructure and application layer that captures value regardless of where the next breakthrough happens.

Sometimes I think the world itself might be one enormous harness. Throughout human history, there have been periods of major invention and technological breakthroughs, followed by periods where the knowledge produced by them is combined, organized, and eventually compacted so that the next generation can continue from where the previous one left off. In a way, this feels similar to how an AI harness manages context over a long-running task. When one context can no longer carry everything that came before it, the important parts are compacted and passed into the next context. The next context does not begin from zero. It inherits a compressed representation of the work that has already been done and continues from there. Civilization seems to work in a strangely similar way. No generation can carry the entire context of human history with it, so knowledge is continuously compressed into language, mathematics, scientific laws, engineering principles, institutions, tools, and technologies, and that compacted context becomes the starting point for the generations that follow. From there, another period of application and deployment begins. Existing knowledge is recombined and applied in new ways, and through those applications we encounter things that were not contained in the previous context: new problems, new phenomena, new laws, and new unknowns. That expanded context is eventually compacted again and passed forward. So perhaps history is not simply a sequence of inventions. It is also a continuous process of context expansion and compaction, with each generation inheriting enough of the previous context to push the frontier a little further. What is interesting about AI is that this process no longer has to happen only across human generations. Harnesses can already expand context through work, compact it, pass it forward, and continue the process again. Perhaps what we are beginning to automate is not just intelligence or labor, but this loop itself. We may already be entering an automated singularity.

This is a genuinely insane world. What’s even more remarkable is how quickly we’ve adapted to it. OpenAI president Greg Brockman pointed out that GPT-4 finished training four years ago today. It wasn’t released until March 14, 2023. OpenAI spent seven months evaluating it, red-teaming it, and tuning it for safety. So when GPT-4 blew everyone away, the model itself was already more than six months old. By the time the public got access to GPT-4, the leading AI labs were already at least half a year beyond what the rest of us could see. The model felt like the future to us. Inside the lab, it was already yesterday’s work. Chinese models are now closing the gap fast on public benchmarks. Kimi, Qwen, and DeepSeek are competing near the top, and the gap between them and the leading US models looks much smaller than it did a few years ago. But public benchmarks only compare models that have already shipped. They don’t tell us how far ahead the labs are behind closed doors. If there is still a six-to-seven-month gap between training and release, the race we’re watching may look very different from the race unfolding inside the labs. And six months means far more now than it did in 2022. There is more compute. The algorithms are better. Synthetic data, reinforcement learning, inference-time compute, and AI-assisted research are all speeding up the development cycle. Sam Altman has also said he isn’t particularly concerned about distillation, which suggests that OpenAI is confident in what it has coming next. Of course, no one outside these labs knows how large the gap really is. But GPT-4 showed that the gap can be real. That also changes how we should think about unreleased research systems like Astra. The ten hard problems Astra revealed on August 1 were solved by a research system, not a polished consumer product. The system hadn’t even reached the public frontier yet. If public releases still trail internal capabilities by several months, the results we’re seeing in papers and demos may already be well behind the real state of the art. GPT-4 finished training only four years ago. At the time, having a conversation with it felt unreal. Today, it feels painfully slow and limited. It’s now something we use as a baseline when comparing the cost and performance of newer models. In those four years, AI has learned to write code, use tools, browse the web, and carry out tasks that take hours. Now it is starting to make headway on math and science problems that have resisted human efforts for decades. AI is no longer something only researchers talk about. It comes up at work, at school, in boardrooms, in politics, and around the dinner table. All of that happened in four years. The speed of human adaptation may be almost as remarkable as the speed of AI progress. A capability feels mind-blowing when it first appears. A few months later, it feels normal. A few months after that, it feels slow and outdated. The danger may be that we’ve become numb to the pace. In a linear world, you can look at the last four years and make a reasonable guess about the next four. That doesn’t work when progress itself is accelerating. If we use the last four years as our yardstick, we will keep underestimating what comes next. It took only four years for GPT-4 to go from science fiction to legacy software. We shouldn’t picture the future moving at today’s speed. We should expect the pace itself to keep picking up. And exponential change is hard to grasp from the sidelines. Sometimes you have to jump in before you understand how quickly the ground is shifting. How long will it take before today’s frontier models feel just as outdated? We’re already living through changes that would have sounded too far-fetched for science fiction not long ago. That may be why even people at the heart of Big Tech and frontier AI labs are leaving secure positions to build what they believe comes next. Even Jeff Dean, after 27 years at Google, has decided to take on something new. People at that level are walking away from comfortable, established careers to bet on the next chapter. Maybe this is a moment to spend less time bracing for change and more time thinking ahead, taking risks, and building. Every major technological shift has given people plenty of reasons to be pessimistic. But the people who moved the world forward were usually the ones who saw possibility beyond the immediate risks and got to work. Pessimism has always sounded smart. Over the long run, optimism has had the better track record. If we’re living through a future stranger than fiction, I’d rather help shape it than watch from the sidelines.

2026 was the year AI started making real scientific discoveries. in a single year: > Claude Fable 5 found a counterexample to the 87-year-old Jacobian Conjecture > OpenAI Astra solved 10 open math problems, including proving the existence of a Bisopic group > GPT-5.2 Pro produced the first clear solution to an open Erdős problem > GPT-5.6 Sol Ultra proved the 50-year-old Cycle Double Cover Conjecture > Grok 4.5 disproved the 5-year-old Hypercontractive Boundary Conjecture > AI identified the genetic cause of a disease that had remained undiagnosed for 20 years > OpenAI increased enzyme reaction efficiency by 79× and discovered a new enzymatic mechanism > Isomorphic Labs unveiled IsoDDE, surpassing AlphaFold 3 by more than 2× for drug design > OpenAI solved all five problems in the AtCoder Algorithm Contest and topped the Heuristic Contest > an internal OpenAI model disproved the 80-year-old Erdős Unit Distance Conjecture this is what AI accomplished in 2026.

OpenAI just killed the price advantage of Chinese open-weight models. This is the beginning of the end. Frontier labs aren’t just winning the race anymore. They’re about to swallow the entire market.

Today we’ve raised $52M Seed and we are announcing the public launch of S2.1 Pro. >It can clone a voice from 5 seconds of audio >2x faster than Cartesia & 1/6th the cost of Eleven Labs >most expressive model with word level control over emotion, intonation, pacing etc We support frontier AI companies including HeyGen, LiveKit, Retell, Sanas, and OpenArt all run our model in production. If you're a business and we can't cut your voice AI costs by 50%, we'll give you 1 year of Fish Audio for free. Book a demo: To celebrate our first birthday, we'll give you 1 month of S2.1 Pro for free. Like, retweet, and comment “Fish” to get it.

🚨 HE’S BACK. Last November, Ilya Sutskever said: “The age of scaling is over. Now it’s the age of research.” Now he’s saying: “Our research has reached the point where it is worth scaling.” That’s a remarkable shift. After focusing on research, he’s now signaling that they’ve found something worth pushing with massive compute. AI is about to accelerate like crazy.

Introducing Vals-Smith: turn your code base into a customized benchmark. Public benchmarks tell you which model is strongest overall, not which model is the best on your code. Vals-Smith turns your merged pull requests into real coding tasks and measures the percentage a model can actually resolve. New models ship every week. Vals-Smith tells you which one to trust with your code.

A lot of people are saying Google is falling behind after Gemini 3.6 Flash. I think they're reading it the wrong way. To me, Google has changed its strategy. Yes, Gemini is behind GPT-5.6 Luna, Grok 4.5, and Claude Sonnet 5 in coding. But it leads in computer use, visual understanding, and long context. At 1 million tokens, it scores more than twice as high as Gemini 3.5 Flash. That doesn't look like a company that is losing. It looks like a company building for real work. Frontier models are already smart enough for most thinking tasks. Now the question is not who gets another benchmark record. The question is who helps people do their jobs every day. People on X talk about agents and hard benchmarks. Most companies are still trying to figure out where AI fits. Most workers are not running agent systems. They need a model that can read documents, understand charts, keep track of long conversations, and work inside the tools they already use. That is exactly where Gemini is strong. If AGI is about doing every kind of knowledge work, then vision, long context, and real world understanding matter just as much as coding. As Demis Hassabis has said, intelligence has to bring all of these things together. Many developers think Google is losing the coding race. I think Google has stopped chasing benchmark wins and started focusing on where the money is. A fast, low-cost model that fits into everyday work may end up being the better strategy.

hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)

This is a crazy update. I’ve been using ChatGPT since the GPT-3.5 days, but GPT-5.6 and ChatGPT Work have given me a genuine wow moment. The difference isn’t just that I’m finding new ways to use it. It’s seeing completely different people experience their own “ChatGPT moment” all over again. Now that everyone is trying Codex, it feels like we’re reliving November 2022. You can almost smell that same excitement in the air—the feeling that something fundamental has changed. This isn’t just another model upgrade. It feels like the beginning of another ChatGPT moment.