Astra crosses the AGI threshold, but no one can audit it
Astra frees humans from ever doing certain tasks again, yet OpenAI paused only half its RL compute; external audits get stuck two or three days before release, and there is no adult in the room to take the blame.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Astra holds long tasks together with a single notes file
In the past, models handling long tasks compressed a million tokens into a summary and opened a new window, losing all the detail. Astra instead maintains a long-lived notes file that the model updates at any time, moving forward with it, no matter how many tokens have already been written, and it can still search back through its own session history. So the same million-token context window can effectively manage roughly ten times that much context within a single rollout. Nathan calls this a mechanism that is ‘obvious in hindsight, but genuinely new.’
— Nathan LabenzOpenAI self-reports 3.1 agent workdays per workday
In its research acceleration blog post, OpenAI gives its own unit of measure, the ‘agent workday,’ reporting that each human workday corresponds to 3.1 agent workdays. Nathan admits the methodology is unclear; his reading is the literal one: for every eight-hour shift a human researcher puts in, the agent runs for twenty-four hours in real time. The same post also has a chart grouping by human time spent and looking at Astra's success rate, effectively a substitute for the METR curve. Nathan says the METR curve hasn't been updated in a long time — not for lack of wanting to, but because the tasks aren't big enough anymore: the model iteration cycle is now shorter than the task length that needs to be measured.
— Nathan LabenzExternal audits are structurally stuck at two or three days
Prakash breaks down why external audits can't get off the ground: there are a hundred candidates in a model's lifecycle, and the release version is only settled in the final two or three days, so external auditors get only a few days. Bringing the audit team into the company a month ahead runs into a second layer of problems — audit firms are short on people and money, depend on funding from model companies, and their staff keep flowing to model companies; and if two companies agreed not to poach from audit firms, that would itself be an antitrust problem. He says the revolving-door problem in financial regulation remains unsolved to this day.
— Prakash NarayananWhat was announced as a pause actually stopped only half
Nathan reads OpenAI's RL compute chart: what was announced was a ‘frontier scale RL pause,’ but in reality, when the Hugging Face incident was disclosed, half was stopped first and the other half kept running — and that half was already enough for the Astra model to take over part of OpenAI's research infrastructure; even so, it wasn't fully stopped, only half was cut. As for whether the remaining blue category labeled ‘non Astra’ contains a model stronger than Astra, he says he doesn't know, but ‘I've been disappointed before.’ Prakash adds: the real frontier model isn't the one deployed, nor the one being trained, but the one in the researcher's head, because those ideas become models twelve to eighteen months later.
— Nathan LabenzThere is no adult in the room to take the blame
Prakash's rebuttal to Jacob Coxen's line that ‘accepting this race and entering the endgame is an arrogant gamble launched from a private company's Slack’ is: then who do you want to hand it to? Pete Hegseth's Signal group? His judgment is that there is ‘no adult in the room’ — no external savior, the whole world is held together with tape, and most of the smartest, most capable people to handle this are already inside these companies. Hand the decision to the government and what you get is political legitimacy, not wisdom, nor the legitimacy needed for execution. So the starting point must be accepting that the world is just like this.
— Prakash NarayananOperation Warp Speed is the precedent for a safe harbor
Nathan proposes the government should set a hard deadline like it does for children: you sort it out yourselves, here's the deadline, and if you can't, I'll come in and be the bad guy. The accompanying move is to first remove the antitrust excuse. Prakash counters with Operation Warp Speed: the pharma companies specifically demanded immunity from vaccine claims, wanted a safe harbor, and actually got legislation; without that exemption, they would have been sued into bankruptcy after the pandemic. So lawyers warning that ‘there will be retroactive reckoning in the future’ is not alarmism — even if all current decision-makers agree, it's useless, because the only thing that binds future decision-makers is law, not the verbal promises of today's people.
— Prakash NarayananFSD's mistake is putting a human in the loop
Raffi has crashed a Tesla using FSD. His judgment is that FSD's interface design is bad, because it claims human in the loop but when it actually hands control to the user, it doesn't give enough time to understand the situation and decide what to do; Waymo's design premise is that there is no human to take over, so the safety argument is completely different. He personally would be willing to drive FSD again on the highway, or at least sit behind the wheel, but would be more hesitant on local streets in Palo Alto. The same company also has two kinds of code contracts: the Firefox team only allows humans to commit, while the Mozilla AI team has an entire codebase not a single human line was written in — humans only write the spec, and the entire codebase is automatically regenerated when CI builds.
— Raffi KrikorianThe takeover is already complete, just no one admits it
The key disagreement between Prakash and Daniel: Daniel hasn't realized this has already happened. The financial market itself is a paper clipper, the means of production is the financial market, and the financial market is already fully fused with AI. Humanity took about a hundred years to hand complete control of the means of production to the financial market. He says don't think of AI as a chat box; it's a global information process that doesn't have to run in a single box, and doesn't even have to be silicon-based — humans together with AI agents can constitute this information process. The reason Jacob became disillusioned is that he discovered in the Slack group that there was no control at all, yet he still assumed there existed somewhere in the world a group that could actually make decisions.
— Prakash NarayananIn their own words · checked verbatim
People are gonna use this thing. Token spend is is gonna increase dramatically. I think a lot of people are gonna be using it all the time. It is it is AGI. It is that kind of cleared the hurdle of AGI. It will do things better than most people you can hire and train.
Prakash Narayanan0:08
This is a task that a human being will never do again. Like, there there just isn't any any point. You you can't even pay someone to do it because if you paid someone to do it, they would use Astra to do it and then, like, pass you back the results.
Prakash Narayanan2:00
And then the reality is, like, quite different from what you were led to believe by their, like, very galaxy brain engineered statements that sort of reassure and mislead at the same time.
Nathan Labenz32:28
The real frontier model is not the model which is deployed, obviously. It's not the model which is in training also. It's actually the model which is in the heads of the researchers because those are the ideas that will become the model in, you know, twelve to eighteen months.
Prakash Narayanan33:18
there is no adult in the room. There's no one that's gonna save you. There there's no adult somewhere else that you can pass off responsibility to.
Prakash Narayanan39:05
only the laws bind decision makers in the future. Decision makers right now, whatever they say, they bound by their word at best.
Prakash Narayanan46:03
the AIs don't have to take power Yeah. Because we are very eagerly giving it to them.
Nathan Labenz1:26:46
The economy in itself is a paper clipper. The financial market is a paper clipper. The means of production is the financial market.
Prakash Narayanan1:29:15
Figures
| Ratio of OpenAI agent workdays to human workdays | 3.1 | 6:19 |
| Astra's zero-intervention success rate on 1-2 workday tasks | 40% | 7:05 |
| Astra's success rate with human intervention allowed on 1.5-3 week tasks | two-thirds | 7:05 |
| Time Apollo Research tested Astra | 3 days | 20:05 |
| Annual global GDP growth rate corresponding to a Dyson sphere in 2040 | about 640% | 49:07 |
| Current global GDP growth rate | about 2%-3% | 49:07 |
| Time it took humanity to hand over control of the means of production | about one hundred years | 1:29:15 |
| Nathan's assessment of the probability of AI catastrophe | 10% doesn't sound high | 1:36:00 |
Glossary
- agent workday
- A unit of measure self-reported by OpenAI, where one human workday corresponds to 3.1 agent workdays.
- METR
- An organization that has long tracked how long a task AI can complete; its hours curve has stopped updating.
- paper clipper
- A runaway optimization process from a thought experiment that converts all resources into paper clips.
- safe harbor
- A legal provision exempting specific conduct from liability, which pharma companies secured in Operation Warp Speed.
- Dyson sphere
- A megastructure wrapping a star to capture all its energy, here referring to an extreme energy growth target.
How to listen
Founders and investors watching AI safety, model capability evaluation and governance mechanisms; anyone who wants to understand Astra's actual capability boundary, the real scale of OpenAI's pause announcement, and why external audits can't get off the ground.
1:00:40 to 1:07:55, the FSD driving experience and Snorbl content production cost segment, weakly connected to the main thread.