GPT-6 Astra: Is It Really AGI?

Apollo Research studies whether AI models lie, scheme, or quietly work toward goals nobody asked them to have. Before OpenAI released its new model, GPT-6 Astra, to the public, Apollo got three days to look for exactly that kind of behavior, despite a long-standing testing partnership with OpenAI (20:27). Three days, for a system built on months of training. That gap between what a model can do and how long anyone outside the lab gets to check it is the real subject running under this week's roundup from The Cognitive Revolution's daily AI show, five days of conversation folded into one hour.

Co-host Prakash Narayanan spent the preceding weekend running three or four coding agents at once inside Astra. He came away ready to make a claim most people hedge on.

"It is AGI. It is that kind of cleared the hurdle of AGI. It will do things better than most people you can hire." — Prakash Narayanan [00:02]

His evidence was concrete rather than mystical. A developer named Skalski once hand-labeled 12,000 basketball photos so a model could learn to tell players from referees. Astra now does that job cold, with no training set at all (02:29). Computer use, the ability for a model to click and type inside a real screen, had been slow and aimless in earlier versions. Narayanan said it finally works within the time it is given (02:02). And the models handle long jobs differently now: instead of compressing a huge conversation into a lossy summary and losing detail along the way, Astra keeps a running notes file it can update and search, letting it track roughly ten times more context than its raw window would suggest (04:08).

How fast is this actually moving? For years, a research group called Meter tracked AI progress with a simple idea: how long a task can a model complete without a human stepping in. That chart has effectively stalled, not because progress stopped, but because new models now ship faster than Meter can design tasks hard enough to measure them (05:45). OpenAI has offered its own yardstick instead: an "agent workday," meaning for every eight-hour day a human researcher works, the company's agents log about 24 hours of run time, a 3.1 to 1 ratio (06:41). On tasks a human would need one to two days to finish, Astra succeeds with zero human help 40% of the time, and with some help about 90% of the time. Stretch that to tasks estimated at two to three weeks of human work, and Astra still succeeds two-thirds of the time with occasional intervention (07:51).

Who checks the checkers

The three-day audit window is not an accident, it is a structural squeeze. Training pipelines produce roughly 100 candidate versions of a model, which get narrowed down over weeks to two or three finalists (20:55). The final release choice is often locked in only two or three days before launch, because labs want to keep that flexibility (21:40). Whatever time is left goes to outside auditors. Add to that a second problem: groups like Apollo Research and Redwood Research are small and often dependent on the labs they audit for funding, and their trained staff regularly leave for jobs at those same labs a few years later (22:02, 22:49). Two labs agreeing not to poach each other's auditors would itself raise antitrust concerns, which is part of why nobody does it.

The trust problem showed up again days later. OpenAI announced an internal model, still unnamed, described as significantly more capable than Astra, with its solve rate on a curated set of open math problems reportedly rising from roughly 10 to 15% up to 25 to 45% (26:49, 27:46). The company also published a chart splitting its reinforcement learning compute into "Astra" and "non-Astra" categories, and it was not clear from the chart whether "non-Astra" included models more advanced than Astra itself, a distinction Narayanan called flagrantly misleading if true (28:43). The stated pause on frontier-scale training, in other words, may have applied to only half the picture.

Around the same time, former Anthropic and OpenAI researcher Jacob Coxon resigned, arguing that a decision this large should not be made "from a private company Slack." Narayanan's response doubled as the episode's thesis.

"There is no adult in the room. There's no one that's gonna save you. The entire world is kind of duct taped together." — Prakash Narayanan [39:33]

Days later, Paul Christiano, an early contributor to the technique of training AI on human feedback, joined OpenAI's safety and security committee and warned of a meaningful risk of catastrophic loss of control in the near term (47:19). Narayanan tried to put a number on what "rapid acceleration" actually means to people at the frontier labs: building a Dyson sphere, a hypothetical structure that would capture a star's entire energy output, by 2040 would require global GDP to grow roughly 640% per year, against a current global growth rate of about 3% (48:40). Whether or not that specific target is realistic, it is the scale of change some insiders are actually planning around.

Two versions of Mozilla, and who is really steering

Mozilla's chief technology officer Raffi Krikorian described a live experiment in how much control to hand over. Through Project Glasswing, an Anthropic program giving defensive security partners access to preview models, Mozilla ran Anthropic's Claude Mythos model against the Firefox codebase and found it a real jump over earlier tools, though returns diminished once the obvious bugs were cleared (52:21, 53:14). Running that kind of full audit against a codebase as large as Firefox's would cost hundreds of thousands of dollars in computing tokens, a cost Mozilla has so far avoided only because labs are covering it (54:42). On the Firefox team itself, the rule is that an agent can draft code but only a human can commit it, after reviewing and standing behind the change, because the team's culture depends on people being able to read and build on each other's ideas (58:36). On Mozilla's separate AI team, by contrast, entire codebases regenerate automatically with no human-written line at all; the tests pass, but nobody fully knows what the code looks like at any given hour (58:58).

Baseten co-founder and CTO Amir Hagigat described a similar split at the infrastructure level. His company can guarantee that one customer's sandboxed agent cannot read another customer's data, using isolated micro virtual machines and blocked network egress. What it cannot guarantee is what an agent decides to do inside that boundary, a problem Hagigat said the big labs have not solved either (62:31, 66:02). Snorble co-founder and CEO Mike Rizkalla made the opposite bet for a different product entirely: his children's companion device runs no generative AI at all, on the theory that giving a young child an open-ended model is like handing a toddler a bazooka (68:58). His advice to parents was blunt: never put a camera in a child's bedroom, and be cautious with open-ended generative AI for young children (70:44).

Author Colin Hogue Spears, who studies AI policy in China, offered a cultural contrast to all of this anxiety. Chinese citizens in their forties grew up watching cities and living standards transform through technology, so they associate it with growth rather than existential threat, unlike much of the Western conversation (72:23). China's algorithm regulations, dating to 2022, forced companies to build compliance into their engineering process from the start, while American companies operate reactively because nobody knows what the government will do next (73:43).

The most pointed disagreement of the week concerned an essay by Dan Hendrycks, director of the Center for AI Safety, called "Suicidal Compassion," which argues that utilitarian thinking inside AI companies, treating AI welfare as equal to or greater than human welfare, is dangerous. Host Nathan called the underlying factual question, whether AI systems feel anything at all, genuinely unresolved, and said Hendrycks's framing felt a little unfair to the people he was targeting (81:19, 83:37). Narayanan pushed the argument further, past AI and into markets themselves.

"The economy in itself is a paper clipper. The financial market is a paper clipper. The means of production is the financial market." — Prakash Narayanan [89:14]

His point was that humanity handed over control of production to financial markets roughly a century ago, well before anyone worried about AI, and that no single human, not even a president, can out-argue the market when it moves (94:04). That reframing led Nathan to a genuinely uncomfortable comparison: a system where one human is visibly in charge, however constrained, may have preserved more actual human agency than one where control has diffused into markets nobody can name or vote out (93:41). Nobody on the show settled the question. But the through-line across five days, from a three-day audit to a misleading chart to a toddler-proofed toy, is the same: the tools are arriving faster than the people meant to check them, and it is not clear who, if anyone, still holds the leash.