I spent Wednesday at an event space in East London packed full of developers. The occasion was Code w/ Claude: Anthropic’s first developer-focused event in Europe. It felt like a victory lap for a company that is reportedly now raising funds at a $900 billion valuation—a number that, if accurate, would surpass that of key rival OpenAI. Behind Anthropic’s meteoric rise are the developers who thronged the London hall this week, eating free lunch and dinner, and accepting complimentary mini computers. These power users spend far more money on Claude than the average person. In other words, they are Anthropic’s lifeblood, and the company has an incentive to keep them coming back for more.

In a day of keynotes across two stages, Anthropic executives let rip a firehose of technical details about how coders should best use AI to increase their productivity. The mood in the room was one of nearly unbridled enthusiasm. Boris Cherny, the creator of Claude Code, proclaimed that AI had helped him reconnect with the feeling of magic that first got him into programming on a school calculator. Developers I met spoke of a renaissance in computer programming, with more opportunities available than people to exploit them.

The day was a picture of Anthropic as a product company. There was little talk about the risks that Anthropic has been so vocal about elsewhere. On stage, Anthropic staffers alluded to the continued rate of AI progress—featuring many graphs going up and to the right—but mostly in the context of encouraging developers to harness Claude Code’s many capabilities, rather than to mitigate their risks. When an Anthropic researcher asked a crowd how many of them had shipped code written by Claude without even reading it, a startling number raised their hands. “That’s a dangerous game,” he said, to nervous laughter. Then he moved on quickly to talking about how developers could squeeze the most out of Claude.

Thursday was a very different experience. I hopped on a train to Oxford University, where Anthropic’s co-founder Jack Clark was scheduled to give a lecture. Clark is a fascinating figure. Much of his work over the years has involved measuring, and then warning about, the rate of AI progress and its many attendant risks. This lecture was no different. He said AI posed a “non-zero” chance of killing everybody on the planet, and warned that even if this total crisis were averted, the next few years would contain more disruption than any in living memory.

Clark predicted that by 2028, or perhaps sooner, AI would reach “recursive self-improvement,” or achieve the capability to improve itself—leading to an unpredictable future. “We have not ever encountered a technology that has this property, where it can make a better version of itself without people,” he said. Most of the world is “in denial” about the capabilities of current AI models, Clark said, let alone those coming down the track in six months.

It would be easy—and comforting—to dismiss this as a billionaire executive talking up his own technology for self-interested reasons. But another way to interpret this lecture, and indeed much of Anthropic’s work as a company, is as a warning from the people who are closest to this strange new technology, and who can see where it is headed. The arrival in April of Claude Mythos—Anthropic’s latest model, which has nation-state-level cyber-offensive capabilities—shows that many of Anthropic’s past predictions were correct, and that the future they predicted is in some sense already here. The model is so powerful that Anthropic decided not to release it publicly, and has instead given a small group of companies and governments access in order to patch vulnerabilities in the world’s most commonly-used software.

Clark said even Anthropic itself had underestimated the scale and speed of AI advancement. When Mythos finished training, he said, “We were like, ‘Oh, it's here! It’s here faster than we thought! And we've done insufficient preparation.’”

It is not nefarious, nor even surprising, that Anthropic is speaking differently to policymakers than developers. Companies tailor messages to different groups all the time, and Anthropic has been open about AI’s risks in ways other companies have chosen to stay silent about. Still, experiencing these two different narratives so close together gave me a profound sense of whiplash. In one sense, Anthropic has a credible claim to hold the keys to perhaps the most powerful technology of all time, while taking it upon itself to warn governments and societies of that inconvenient fact. In another very real sense, it is still a normal software company, dependent on the whims of fickle developers in hoodies and jeans.

If Clark is right, nothing about this moment in time will last for long. The developers Anthropic is currently so dependent on will find themselves replaced by machines faster than they can say “recursive self-improvement.” Perhaps that's why his predictions didn't find their way into Wednesday's keynotes.