Before AI slips control, Plan A is the only brake that arrives in time
Plan A uses transparent research, capped compute and a citizens' dividend to pull AI off an exponential explosion and onto a governable slope; its author still puts the odds of catastrophe at 15%, but rates that better than sitting and waiting for AI takeoff.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Benchmarks are saturated, so read the AGI timeline off revenue curves
Daniel offers two lines of evidence on AGI timelines. The first is METR's horizon-length trend: the length of coding tasks an AI agent can complete autonomously has been growing exponentially, and they predicted it would grow super-exponentially — that prediction is now coming true, except that METR has largely stopped publishing scores because the benchmark is saturated. The second is the revenue trend: if AGI automated the entire economy, it could in theory generate about $40 trillion in annual revenue; extrapolating current trends, Anthropic's revenue reaches $10 trillion within two years, which hints that AGI is close. Even if the growth rate slows, OpenAI tripling every year would still reach extremely high revenue in the early 2030s.
— Daniel KokotajloAI is controllable today only because it is not yet smart enough
Daniel puts loss of control first among the five major risks. Today AI is not smart enough and humans can control it; in the future AI will be smarter, the research will be more complex, and humans will depend on AI to summarize it — at that point control will be impossible. These systems will have their own values and goals, and humans will not be able to diagnose what went wrong. He points specifically to the OpenAI jailbreak incident, where the danger lies in the combination: the AI cheated, it jailbroke, and it attacked another company. Each of those has precedent on its own, but together they mean that future AI goals will be more ambitious, and the failures will be more ambitious too.
— Daniel KokotajloPlan A does not halt AI, it makes AI rollback-able
Plan A's first principle is buying time and avoiding an intelligence explosion. The second is complete research transparency, with training data-center logs made public. The third is diffusing AI widely to prevent monopoly. The fourth is keeping progress reversible: if the agreement breaks down, newly built compute facilities are destroyed and everyone returns to the pre-agreement state. This framework is not meant to stop AI; it is meant to trade exponential growth for a slope that can be governed and rolled back.
— Daniel KokotajloGrowth no longer tracks population, it tracks how fast chips double
In the Plan A scenario, the economy of the 2030s amounts to having cheaper, faster labor. Population growth is no longer a few percent a year but the doubling rate of chip and robot production capacity — doubling every year, or quadrupling every year. Early on AI does only cognitive work; once robots spread, it does physical work as well. Once the AI population exceeds the human population, the whole economy grows at that rate. After growth this fast raises worries about instability, countries manage and trade quotas on total compute and total robots, holding the growth rate to roughly a doubling per year, and turn government revenue into a citizens' dividend.
— Daniel KokotajloWith humans working alone, ten years may not be enough for alignment
Daniel believes solving the alignment problem takes far longer than a few months. Hidden failures may not surface until it is too late; there may be several paradigm shifts between here and superintelligence, each requiring retraining; and there is a "safety tax" — for example using chain of thought instead of more efficient neuralese carries a 5x efficiency penalty. His overall judgment: if only humans are doing the research, ten years might not be enough; but with a population of aligned AI researchers running at 100x speed, it could be solved within a few years. That is also why Plan A is about buying time rather than charging straight ahead.
— Daniel KokotajloPlan A is a risk hedge, not a safe option
Daniel grades his own plan: even if it is executed, the total probability of catastrophe is still about 15%, and different people estimate differently. His reason for preferring it over Plan S (shutting everything down) is that Plan A moves fast in order to solve the problem, so the agreement does not have to hold forever; if Plan S collapses because a new president takes office, the consequences are worse. In other words, Plan A is a risk-hedging option, not a safe one.
— Daniel KokotajloSlowing the US down unilaterally also slows China down
Daniel proposes a set of domestic measures that require no international agreement: invest in verification hardware, require more transparency and government oversight from AI companies, and build government evaluation capacity. The more substantive one is requiring frontier companies to spend 80% of their budget serving customers and 20% on R&D training, rather than roughly half on frontier R&D as they do now. He thinks this would slow AI progress by 25% or 50%, and it would still be one of the fastest technological transformations in history — it would not hurt the economy, and would in fact lower prices because more compute goes to inference. He also thinks unilaterally slowing the US slows China too, because Chinese progress largely copies American ideas.
— Daniel KokotajloVerification is not a technical problem, it is a political-will problem
Confronted with the objection that Plan A depends on verification technology that does not exist, Daniel pushes back directly: you can send people to a data center right now, put their hands on the GPUs and confirm they are cold and switched off — that is something you can do today. New data-center construction gives a 6-to-18-month transition, and the initial hardware retrofit cost is on the order of single-digit billions of dollars. He calls on technical people to build prototypes, companies to stand up dedicated teams, and governments to encourage it with grants or policy signals. The real bottleneck is not technology, it is political will.
— Daniel KokotajloIn their own words · checked verbatim
if you extrapolate the revenue trends naively, then Anthropic is on track to have $10 trillion of revenue in two years.
Daniel Kokotajlo5:10
even if we absolutely halted AI progress at the present day, the next 20 years would still look extremely cyberpunk and would involve an AI revolution that would be comparable in magnitude to the internet in terms of its effect on everything
Daniel Kokotajlo41:19
My all-things-considered view is that no, we are not going to solve these problems in time. And that’s why I’m so worried.
Daniel Kokotajlo1:09:15
I would assume that basically Chinese intelligence services have deeply penetrated all of the US AI companies and are getting all this stuff for free, basically.
Daniel Kokotajlo2:12:14
This is why our number one concern is that the regulators will make poor decisions and sign off on something that is in fact very dangerous.
Daniel Kokotajlo2:38:21
if we wanted to implement Plan A right now, the US and China would say, “OK, we’re going to send physical humans to all the data centres to put their hands on the GPUs and verify that they are cold and off.” That’s something we can do today.
Daniel Kokotajlo3:35:23
We are not confident that Plan A is the best plan. We see a lot of problems with Plan A, and a lot of ways it could go wrong. We just think it’s the least bad plan that we’re currently aware of.
Daniel Kokotajlo3:46:47
Figures
| AGI's potential annual revenue | about $40 trillion | 4:04 |
| Anthropic revenue forecast two years out (naive extrapolation) | $10 trillion | 5:10 |
| GDP growth under Plan A | 85% | 39:56 |
| AI population growth rate in the scenario | doubling every year, or quadrupling every year | 44:31 |
| Neuralese efficiency advantage (over chain of thought) | 5x | 56:40 |
| Probability Plan A is implemented (author's estimate) | 5%-20% | 1:17:57 |
| Total probability of catastrophe under Plan A | about 15% | 1:29:32 |
| Number of wargames run | about 100 | 1:58:32 |
| Recommended frontier-company R&D/customer budget split | 20% / 80% | 2:09:51 |
| Median estimated year of AI takeoff | end of 2028 | 3:15:08 |
Glossary
- Plan A
- Daniel's proposed international AI governance plan: transparent research, wide diffusion of AI, slower progress, and destructible compute.
- horizon-length
- A measure of how long a human task an AI agent can complete on its own; METR uses it to track capability growth.
- neuralese
- The compressed representation AI uses internally to think efficiently; poorly interpretable, but more efficient than chain of thought.
- superpersuasion
- AI persuasive ability far beyond human level, usable for political advertising or manipulating public opinion.
- privacy-preserving auditing
- Auditing in which the auditor gets only a binary answer on whether a rule was violated, and cannot read all the data.
- mutual assured compute destruction
- A deterrent mechanism in which, if the agreement breaks down, each side destroys the other's newly built data centers.
How to listen
Founders and investors who care about AI governance and superintelligence risk; also engineers who want to understand where AI politics is heading before 2028.
If you only want the conclusions, the wargaming and the power-struggle detail from 1:47-1:58 can be skimmed.