AI Chatter Unfiltered: Weco's Self-Improving AI Sparks a New Wave of Debate

Weco AI just claimed it has the first evidence of recursive self-improvement in an AI system. That’s the holy grail, the moment an AI starts making itself smarter, and it’s a claim that changes the entire texture of the conversation. Last week, Admin, we were talking about the drama around Anthropic's agents and how they use tools, but this week the chatter jumped from agents using tools to agents improving themselves. What's happening on the ground, in the labs, and on the balance sheets suggests the AI race is not just accelerating—it's fracturing into several different races at once. The single biggest financial shockwave this week came from a simple revenue report. Anthropic just passed OpenAI in quarterly revenue for the very first time. The Wall Street Journal laid out the numbers: OpenAI’s revenue grew eighteen percent to six-point-seven billion dollars.

Respectable. But Anthropic’s revenue more than DOUBLED, hitting eleven-point-six billion. And here’s the kicker: while OpenAI’s operating losses widened to a staggering twelve-point-three billion, Anthropic actually posted a small operating profit. This isn't just a lead change; it's a fundamental challenge to the idea that having the biggest consumer-facing model automatically wins the war. It suggests the real money, right now, is in the enterprise. And as if on cue, OpenAI’s CFO Sarah Friar told employees this week not to worry if Anthropic beats them to the public markets, confirming OpenAI itself "will be a public company in 2027" or even sooner. That’s a massive shift in posture, putting a firm timeline on an IPO and signaling that the pressure to turn their billion users into a profitable business is becoming acute.

Meanwhile, the physical world is catching up to the digital one. A new generalist robot model, GEN-1.5, demonstrated it could learn new, complex physical tasks from just three to twelve seconds of a single human demonstration. Think about that. Not hours of training, not complex coding—just watching once. It achieved fifty-nine percent success on average across ten different tasks right out of the box, and that jumped to eighty-three percent success after just a few minutes of learning updates. This is the kind of breakthrough that closes the gap between a robot in a lab and a robot on a factory floor or in your home. Speaking of your home, Amazon is making two big moves. First, its Prime Air drone delivery service is set to expand to nearly five hundred U.S. cities by the end of 2026.

After a key regulatory approval for longer-range flights, an executive expects they'll hit about one million deliveries this year alone. Packages under five pounds, arriving in as little as thirty minutes. Second, Amazon just made its conversational Alexa-plus AI available at no additional cost on compatible Fire TV devices in the U.S. You don't even need a Prime subscription. It's an automatic upgrade, pushing more capable AI directly into the living room. And the money keeps flowing to the infrastructure that powers all this. Fractile, a British AI-chip startup, is in advanced talks to raise six hundred million dollars at a six-point-five billion dollar valuation. That’s more than six times its valuation from just back in May. The reason for the jump? A roughly two-hundred-fifty million dollar supply deal with—you guessed it—Anthropic, for chips that won't even be delivered until 2027.

The market is betting on the hardware needed for the next generation of models, and it's betting big. Even SpaceX is getting in on the action, with Bloomberg reporting that the company approached the AI-coding startup Cognition about a potential acquisition. Everyone is trying to secure their piece of the AI stack. Now, on the pure research front, two developments are driving the conversation. The first is that OpenAI model that Aravind Saravu highlighted. It discovered an entirely new family of mathematical constructions, solving a problem in a way that performs better than any known human-designed method. This is not just pattern recognition. This is AI generating novel, verifiable, and superior knowledge in a highly abstract field. It’s a clear win for the science.

The second is a quieter but powerful technique called Recirculation. It’s a change to how transformer models work during inference—when they’re actually generating answers. It feeds deeper-level understanding back into the earlier layers of the network, essentially letting the model "think" more deeply about a problem without needing to be retrained. The results are impressive: on Gemma 3 models, it led to a twenty-one percent improvement on math problems and a twenty-three percent overall improvement in understanding, with very little extra latency. It’s a software trick that makes existing hardware and models suddenly more powerful. And finally, there's the human element. Or, the simulated human element. A new paper called "Physics of Agents" studied what happens when you put more than ten thousand language model agents into communities and let them interact.

They found that the way opinions spread—whether they reach consensus, polarize into factions, or just end in indifference—can be predicted using the laws of statistical mechanics. It’s a step toward creating a predictive science for emergent AI behavior, moving us from simply being surprised by what agent swarms do to actually engineering the outcomes. So what does it all add up to? You have a major power shift in the business of AI, you have robots learning from a glance, you have regulators circling, and you have research pushing into both provably correct new knowledge and wildly speculative new territory. The board is changing. Fast. Okay, let's go deeper on the two poles of this week's conversation. The money, and the magic. First, the money. That revenue reversal between Anthropic and OpenAI is the most important business story in AI this year.

Period. For two years, the narrative has been simple: OpenAI is the sun, and everyone else is a planet in its orbit. They had the consumer hit with ChatGPT, the brand recognition, the developer momentum. But that story just ran into a wall of financial reality. Let’s break down the numbers again. OpenAI grew, but its losses grew FASTER. Twelve-point-three billion dollars in operating losses in a single quarter. That is an astonishing burn rate. It’s the cost of serving a billion free or low-cost users and training the next generation of frontier models. Now look at Anthropic. They more than doubled revenue to eleven-point-six billion and eked out a profit. How? By almost completely ignoring the consumer market. Anthropic went all-in on enterprise. They sell to large corporations with big budgets and specific, high-value problems to solve.

Their strategy is built on long-term contracts, deep integration into existing workflows, and a focus on safety and reliability that big companies demand. It’s less glamorous than a viral chatbot, but it turns out to be a waaay better business. At least for now. Each enterprise customer is worth thousands, or millions, of consumer subscribers. The margins are better. The customers are stickier. This is the context for Sarah Friar’s announcement about a 2027 IPO. It’s not just a casual statement. It is a signal to the market, to employees, and to competitors that OpenAI knows it needs to pivot from a research lab with a hit product into a sustainable public company. The clock is ticking. They have to figure out how to monetize that massive user base more effectively, or the sheer cost of running their models will crush them.

You can't lose twelve billion dollars a quarter indefinitely, no matter how good your tech is. So what you're seeing, Admin, is a fundamental divergence in strategy. OpenAI took the "blitzscaling" approach—grow at all costs, capture the users, and figure out the business model later. Anthropic took the classic enterprise software approach—find the customers with money and solve their problems first. And for this quarter, at least, the boring strategy won. It also connects directly to the hardware story. That Fractile chip deal isn't just about Anthropic buying silicon. It's about them securing a dedicated supply chain, tailored to their needs, to serve their high-paying enterprise clients. They are building a moat, not just with their models, but with their entire business and hardware stack.

Now, let's talk about the magic. That claim from Weco AI about "AIDE-squared"—Autonomous, Iterative, and Directional Evolution. They are claiming to have demonstrated recursive self-improvement. Let's be very clear about what that means. This is not just an AI learning from data. This is an AI analyzing its own architecture, its own code, and then generating improvements to make itself smarter. Then the new, smarter version repeats the process. It's the feedback loop that many researchers believe is the critical step toward Artificial General Intelligence. It’s the stuff of science fiction. And, frankly, it's the stuff of nightmares for AI safety researchers. The evidence so far is just a website and a tweet. We don't have a peer-reviewed paper. We don't have a public demo.

But the claim itself is what matters to the conversation right now, because it forces everyone to confront the possibility. For years, this has been a theoretical milestone. Now, a company is saying, "We're here." The immediate reaction in the community is, of course, extreme skepticism. But it's not zero. Because we are seeing adjacent breakthroughs that make it feel more plausible than it did even a year ago. Look at the OpenAI math discovery. That's an AI creating knowledge that was unknown to humanity. It’s not recursive self-improvement, but it IS a demonstration of superhuman cognitive ability in a specific, formal domain. It proves that these models can do more than just remix their training data. They can reason their way to novel insights. Or look at the "Physics of Agents" paper.

We are starting to build the scientific tools to understand and predict the behavior of large-scale AI systems. This is the groundwork you’d need to even begin to safely manage a recursively self-improving AI. You can't control what you can't predict. So you have this spectrum. On one end, you have concrete, verifiable, but narrow breakthroughs like the math discovery. In the middle, you have practical engineering advances like Recirculation that make current systems better. And then, way out on the bleeding edge, you have this explosive claim from Weco AI. Here’s the thread that connects it all. The field is professionalizing and specializing at an incredible rate. The business of AI, as shown by Anthropic, is about solving real-world enterprise problems for profit.

The application of AI, shown by the GEN-1.5 robot, is about moving from screens into the physical world. But the frontier of AI research is still chasing the ultimate prize—an intelligence that can think and create and improve on its own. The tension between the pragmatic, profitable present and the speculative, world-changing future has never been sharper. This week’s chatter shows that you can't just follow one of these threads anymore. They are all interconnected. The profits from Anthropic's enterprise deals will fund the next generation of models. The breakthroughs in robotics will create new markets for those models. And the chase for recursive self-improvement, whether it's real yet or not, sets the ultimate destination for the entire industry. What this week sets up is a new kind of competition.

It’s not just about model quality anymore. It's about business models. It's about supply chains. It's about public perception and regulatory pressure. And it's about who can navigate the tension between building a profitable business today and building the technology that will define tomorrow. The monolithic era of AI, dominated by one or two names, is ending. We're entering an era of specialization, where the winners in enterprise, in robotics, in consumer, and in pure research might all be different companies. The race just got a lot more complicated. And a lot more consequential.