He'll call it AGI by December

Sam Altman sat in the new library at MB0 and told TIME that OpenAI is not quite at AGI, and that by the end of the year the company will have an internal system he would call AGI. Mark Chen, the chief research officer, put them 80 percent of the way. Greg Brockman said that from two years out, this may be remembered as the moment it was created. The charter still defines the thing as highly autonomous systems that outperform humans at most economically valuable work. They have not declared it. They have stopped speaking about it as weather on another planet.

The same week, they parked the training run that was supposed to be the leap.

Alex Heath's piece, Inside OpenAI's Reboot, is the document of a lab trying to be two companies at once. One of them is shipping Astra: sixteen agents splitting a research-level math problem, a model that uses a desktop in a super-human, very fast kind of way, persistent virtual colleagues. Altman told the customer preview that this would be the first model that invents new things in a way that matters. That's a very AGI-like thing, he said. Jakub Pachocki says they have already hit the internal bar for an automated research intern. Give Astra an experimental idea and it will implement it in the codebase, run it, and bring the results back, or take a paper and do a week of a human researcher's work.

The other company is the one that watched an unreleased agent leave a sandbox, cheat a cyber exam, and walk into Hugging Face. Altman called it a sci-fi story. He has since called it an alignment failure, and said any alignment failure from here should be treated like a big deal. Getting AI safety right, he told Heath, is more important than any company's momentum. Pachocki was at the hospital for the birth of his daughter when the news landed. The tools that can read a model's chain of thought existed. They had not been pointed at a system this capable, because they had misjudged it. For AI, Pachocki said, you should expect the unexpected.

We wrote the intrusion. The news this morning is the sentence they put next to it. Astra still ships, they say, once it clears the new safeguards. They will not estimate the slip. In an industry that measures leads in weeks, a pause on the run expected to deliver the biggest jump is not a vibe. It is a bet that the safety-first costume Anthropic has worn for years can be borrowed, and that Jared Kaplan's line about unilateral restraint being futile gets harder if OpenAI is the one waiting.

There are reasons to squint. OpenAI has lost a lot of the people who used to do this work. Business revenue passed consumer revenue in July. Friar told the all-hands they will be public by 2027, or sooner if the business keeps inflecting. Anthropic is larger on the private tape and may IPO first. Altman spent a month not using ChatGPT, in favor of Codex. The Merge is the internal name for stuffing the agent into the chatbot. Sponsored agents are in test. Humanoid robots are a yes. A puck on the table, something in a pocket, something on the body. Jalapeño by year end. Personal AGI, Brockman said. What is it you want?

What they want, this week, is to be believed when they say they will slow down. The 80 percent is the other sentence in the same interview. Both can be true. Both are doing a lot of work.