AI 2040 Plan A
Slow frontier AI, expose the science, distribute capability, and preserve reversibility
- Difficulty
- Expert
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 99%
AI 2040 Plan A is the AI Futures Project's recommended route for changing the default race to superintelligence. Its first goal is to slow frontier capability growth before companies automate the full AI-research loop. Its second is radical transparency, allowing scientists and governments to inspect training recipes, architectures, risks, and safety evidence rather than relying on company assurances. Third, capability should remain distributed across multiple companies and countries so no single project controls the technology. Fourth, new infrastructure should be reversible: if an international agreement collapses and racing resumes, newly built data centers can be disabled rather than intensifying the race. The proposed transition pauses new training while permitting existing-model inference, establishes reciprocal inspections, creates transparent training facilities, and then allows bounded progress while dangerous activities remain restricted.
Origin
The AI Futures Project created Plan A after readers of AI 2027 asked for a positive recommendation rather than only a forecast of the dangerous default path.
Core principles
- 01Slow capability growth before recursive self-improvement begins
- 02Make frontier research transparent to scientists and governments
- 03Distribute advanced capability across countries and companies
- 04Build infrastructure so a failed agreement can be reversed
- 05Allow lower-risk use while restricting dangerous development
How to run it
- 1
Pause frontier training
Temporarily stop development of new frontier models while allowing existing systems to continue serving permitted uses through inference. Use the pause to negotiate a safer development structure.
Pro tip Distinguish training from inference so the intervention does not require immediately switching off every existing service.
Watch out Waiting until most jobs are automated means the most powerful systems may already exist.
- 2
Verify reciprocal compliance
Send inspectors into major data centers across participating countries to verify that facilities are serving existing models rather than training new ones. Make compliance observable to every side.
Pro tip Use reciprocal inspection to reduce each country's fear that another is secretly continuing the race.
Watch out A pause based only on promises preserves the incentive to defect.
- 3
Create transparent facilities
Build designated data centers where future training occurs under open scientific scrutiny. Publish model architectures, training methods, and relevant findings.
Pro tip Give independent scientists direct access to the evidence needed to challenge safety claims.
Watch out An auditor-only system can become adversarial when companies know more than regulators and can withhold novel problems.
- 4
Distribute frontier capability
Use slower progress and shared knowledge to let multiple countries and companies operate at similar capability levels. Avoid concentrating the best systems inside one private or national project.
Pro tip Let transparency reduce technical gaps while governments retain their own regulatory choices.
Watch out Replacing a corporate monopoly with one global control center preserves concentration risk.
- 5
Bound dangerous progress
Maintain an ongoing process for deciding which research and deployments are too dangerous and which can proceed. Prioritize interpretability, alignment, and controllability while preventing an intelligence explosion.
Pro tip Require evidence that grows stronger as the capability and potential downside increase.
Watch out Do not let ordinary economic deployment quietly recreate recursive self-improvement.
- 6
Preserve reversibility
Design newly built compute infrastructure so it can be disabled if coordination breaks down and a race resumes. Return the parties toward the pre-expansion baseline instead of leaving a larger uncontrolled stock of compute.
Pro tip Specify the shutdown mechanism before approving the build-out.
Watch out Reversibility added after infrastructure exists may be politically or technically impossible.
In the wild
In the scenario, governments pause frontier training, retrofit existing data centers for inference, and exchange inspectors who verify the distinction. During the pause, they build transparent training facilities where research later resumes under open scientific scrutiny and restrictions on dangerous activity.
→ AI capability continues advancing at a slower pace while scrutiny, distribution, and safety research gain time to catch up.
Illustrative example: three countries register frontier clusters, suspend training runs above an agreed threshold, exchange inspectors, and permit customer inference. New shared research facilities publish their methods and include pre-agreed shutdown controls if any participant returns to secret racing.
→ Each participant can verify restraint while preserving useful services and a controlled route back to research.
Common mistakes
Regulating after mass displacement
The plan depends on intervention before companies complete AI-research automation and create systems powerful enough to resist meaningful control.
Trusting private safety assurances
Without broad research transparency, companies retain stronger information and incentives than their auditors.
Building without an exit mechanism
Additional compute worsens a renewed race unless the agreement defines how that infrastructure can be disabled.
Is it for you?
Best for
It is best for policymakers and technical institutions designing coordinated frontier-AI risk controls.
Not ideal for
It is not ideal for unilateral company policy because its central mechanisms require domestic regulation and international coordination.
From the transcript
“goal one, slow things down. Goal two, make it more transparent.”
“we actually think it's actively good for there to be multiple AI companies across multiple different countries that have similar levels of very advanced AI…”
“the fourth principle is basically build the new data centers in such a way that if everything breaks down and everyone starts racing again, the…”
From the episode
OpenAI Whistleblower FINALLY Speaks: “AI Has A 70% Chance Of Going Horribly Wrong!“