Courses/Compute Verification

NEXT COHORT SOONFREE · ADVANCED

Compute Verification

Could the world actually verify what’s running in AI data centres? Any enforceable agreement on advanced AI depends on it. A guided, discussion-based read of the AI 2040 reading list, from the team behind AI 2027: one intensive Mon–Fri week, or part-time over six.

Enroll for free

No application · ~2-minute signup

About this course

If the world ever agrees to limit dangerous AI development, someone has to be able to check that the agreement is kept, by verifying claims about what computations are actually happening on the servers in AI data centres. That verification problem sits at the intersection of hardware, cryptography, and international policy, and it’s one of the most neglected, concrete places to contribute to AI safety.

What you’ll walk away with

A map of the verification problem

What's solved, what's stuck, and why it matters for AI agreements.

Technical-policy fluency

Speak both languages, mechanisms and agreements, credibly.

Ground floor of a young field

Few people work on this, so serious early engagement counts for a lot.

A peer network in AI safety

The people you spent the course thinking alongside.

A certificate you can verify

Tiered (Associate, Professional, Expert) with a public link for your LinkedIn.

A clear next step

Whether that's a deeper course, a project, or a career move.

This course is for you, if:

ENGINEER

You build systems

Hardware/security skills, real problem.

POLICY

You work on governance

You want to know what's enforceable.

RESEARCHER

You want a neglected niche

A concrete, under-explored corner.

SKEPTIC

You doubt treaties can work

Could verification make deals real?

What you’ll explore

6 UNITS · ~5H

All readings are provided online; nothing to buy. Draft outline: final unit titles may shift as the course team polishes the curriculum.

01

Why compute is the lever

Frontier AI needs extraordinary amounts of specialized hardware. Why that makes compute the most governable input, and the foundation any serious agreement would build on.

02

What would need verifying

From chip counts to training-run properties: the specific claims about data-centre computations that an agreement would have to check.

03

Hardware mechanisms

On-chip and on-site approaches such as secure elements, attestation, and monitoring, and what they can and can't prove about a running workload.

04

Verification without trust

How adversarial parties could check each other's claims: the cryptographic and institutional designs that don't require anyone to take anyone's word.

05

The international picture

Precedents from arms control and export regimes, what's different about compute, and how verification could anchor real agreements between rivals.

06

Open problems and your role

Where the field is stuck, what would unstick it, and how your skills map onto the work that needs doing now.

INCLUDED · FOR A LIMITED TIME

A 1-on-1 call with someone working on AI safety full-time

This is a coaching call, not a sales call. Bring your questions, your doubts, and your situation. You'll leave with a personalized AI safety action plan you're genuinely excited to pursue.

Enroll for free

You could read the list alone.

Here's what taking the course changes.

ALONETHIS COURSE
Finish the reading listMaybeUnit 6
Test your takes against othersEvery session
Someone to ask the dumb questionsLens Coach
A plan for what to do next1-on-1 call
Proof you did itCertificate

Common questions

Is it really free?+

Yes. We're a nonprofit built to help people understand the risks of AI and decide how to respond. There's no paid tier and no upsell. The course, the coaching call, and the certificate are all free.

How much time does it take?+

About 5 hours per unit: around three hours of self-study spread out however you like (readings, Lens Coach, a short exercise), plus one 90-minute group session.

Intensive or part-time: which should I pick?+

Intensive packs all six units into one full-time week, Monday to Friday, so expect a full working day of focus each day. It suits full immersion: between jobs, on a break, or when you don't want to wait. Part-time is a unit a week for six weeks and fits around work and life. Same readings, same discussions, same coaching either way.

Do I need a hardware or policy background?+

No. The course builds up both sides. Engineers pick up the policy context, policy people pick up the technical mechanisms, and the discussions are richer for having both in the room.

What materials do I need?+

Nothing to buy. The course is built on the AI 2040 reading list, and every reading is provided online.

What if I can't make one of my sessions?+

Life happens. Message Lens Coach and you can swap into another group's session for that unit, or reschedule. We'll help you find a time that works.

What happens after the course?+

You keep your peer network and the community Discord, Lens Coach stays available, and if you want to keep going, our Forecasting, Modeling, and Shaping AI Futures course or a project of your own are natural next steps.

Not sure it’s for you? Come to the first session. If it’s not the right fit, we’ll help you find something that is.

Compute Verification

Enroll free