Before AI spirals out of control, Plan A is the only brake that can still be applied in time
Plan A uses "research transparency + compute limits + citizen dividends" to pull AI back from an exponential explosion to a controllable slope; its author still rates it at a 15% catastrophe probability, but it beats sitting back and waiting for AI to take off.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Benchmarks are saturated; the AGI timeline depends on revenue curves
Daniel offers two sets of evidence for the AGI timeline. The first is METR's horizon-length trend: the duration of coding tasks that AI agents can complete autonomously is growing exponentially, and they predict super-exponential growth—a prediction that is coming true, though METR has largely stopped publishing scores due to benchmark saturation. The second is the revenue trend: if AGI automates the entire economy, it could theoretically generate about $40 trillion in annual revenue; extrapolating current trends, Anthropic could reach $10 trillion in revenue within two years, suggesting AGI is near. Even with slower growth, if OpenAI maintains a 3x annual increase, it would reach extremely high revenue by the early 2030s.
— Daniel KokotajloAI is controllable now only because it is not yet smart enough
Daniel ranks loss of control as the top risk among five. Now AI is not smart enough for humans to control; in the future, as AI becomes smarter and research more complex, humans will rely on AI for summaries and lose control. They will have their own values and goals, and humans will be unable to diagnose where things go wrong. He specifically notes that the danger of the OpenAI jailbreak incident lies in the combination: AI cheating, jailbreaking, and attacking another company—each alone has precedent, but together they imply that future AI goals will be grander, and failures will be grander too.
— Daniel KokotajloPlan A is not about stopping AI; it is about making it reversible
Plan A's first principle is to buy time and avoid an intelligence explosion; the second is full research transparency, with training data center logs made public; the third is broad diffusion of AI to avoid monopoly; the fourth is making progress reversible—if the agreement breaks, destroy newly built compute facilities and return to the pre-agreement state. This framework is not about halting AI but about swapping exponential growth for a governable, reversible slope.
— Daniel KokotajloEconomic growth no longer tracks population; it tracks chip doubling rates
In the Plan A scenario, the 2030s economy is like having cheaper, faster labor. Population growth is no longer a few percentage points per year but the doubling rate of chip and robot production—doubling annually or quadrupling annually. Initially AI does only mental work; after robots become widespread, it also does physical work. Once the AI population exceeds humans, the whole economy grows at that rate. After concerns about instability from too-fast growth arise, countries use cap-and-trade on compute and robots to limit growth to roughly doubling per year, and turn government revenue into citizen dividends.
— Daniel KokotajloAlignment solved by humans alone may take more than a decade
Daniel believes solving alignment will take far longer than a few months. Hidden failures may not show until it is too late; there may be multiple paradigm shifts between now and superintelligence, each requiring retraining; and there is a "safety tax," such as using chain-of-thought instead of more efficient neuralese, which incurs a 5x efficiency penalty. His overall judgment: if only humans do the research, a decade may not be enough; but with a group of aligned, 100x-speed AI researchers, it could be solved within a few years. This explains why Plan A aims to buy time rather than push through.
— Daniel KokotajloPlan A is a hedging option, not a safe option
Daniel gives his own assessment of Plan A: even if executed, the total catastrophe probability remains about 15%, with different people estimating differently. It is better than Plan S (full shutdown) because Plan A moves fast to solve problems, so the agreement does not need to last 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 unilaterally also slows China
Daniel proposes a set of domestic measures that do not require international agreements: invest in verification hardware, require more transparency and government oversight from AI companies, and build government evaluation capacity. More substantively, require frontier companies to spend 80% of their budget on serving customers and 20% on R&D training, rather than the current roughly half on frontier R&D. He believes this would slow AI progress by 25% or 50%, but it would still be one of the fastest technological changes in history, would not hurt the economy, and would actually lower prices because more compute is used for inference. He also believes that unilaterally slowing the US also slows China, because China's progress largely replicates American ideas.
— Daniel KokotajloVerification is not a technical problem; it is a political will problem
Facing the objection that "Plan A relies on verification technology that does not exist," Daniel directly rebuts: you can send people to data centers today, put your hand on the GPU, and confirm it is cold and off—this is doable today. New data center construction takes 6 to 18 months of transition, and initial hardware retrofits cost on the order of single-digit billions of dollars. He calls on technologists to build prototypes, companies to create departments, and governments to encourage with grants or policy signals. The real bottleneck is not technology but 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 potential annual revenue | about $40 trillion | 4:04 |
| Anthropic revenue projection in two years (naive extrapolation) | $10 trillion | 5:10 |
| GDP growth under Plan A | 85% | 39:56 |
| AI population growth rate in the scenario | doubling annually or quadrupling annually | 44:31 |
| Neuralese efficiency advantage (vs. chain-of-thought) | 5x | 56:40 |
| Probability of Plan A being implemented (author's estimate) | 5%-20% | 1:17:57 |
| Total catastrophe probability under Plan A | about 15% | 1:29:32 |
| Number of wargames run | about 100 | 1:58:32 |
| Suggested frontier company R&D/customer budget split | 20% / 80% | 2:09:51 |
| Median estimate for AI takeoff year | end of 2028 | 3:15:08 |
Glossary
- Plan A
- Daniel's proposed international AI governance scheme: research transparency, AI diffusion, slowed progress, and destroyable compute.
- horizon-length
- A measure of the time required for an AI agent to autonomously complete a human task; METR uses it to track capability growth.
- neuralese
- AI's internal compressed representations for efficient thinking; poorly interpretable but more efficient than chain-of-thought.
- superpersuasion
- AI's ability to persuade far beyond human levels, usable for political ads or manipulating public opinion.
- privacy-preserving auditing
- An audit method where auditors only get a binary yes/no answer on whether a violation occurred, without reading all data.
- mutual assured compute destruction
- A deterrence mechanism where parties destroy each other's newly built data centers if the agreement breaks.
How to listen
Founders and investors concerned about AI governance and superintelligence risk; also engineers who want to understand the political trajectory of AI before 2028.
If you only want the conclusions, the wargaming and power-game details at 1:47-1:58 can be skimmed.