What do students even want from Lens Academy's Compute Verification Intensive?

LessWrong，LessWrong 是一个专注于人类理性思考的社区博客，讨论认知偏差、决策理论、AI 安全与有效利他主义。关注

This is an independent review, and all views represented are my own. An early version of this draft was approved by Lens Academy, but this post was not commissioned by them.

Executive Summary

I took Lens Academy's Compute Verification intensive during the week of September 7 2026. I had a positive experience and registered to retake it during week of October 26. If you are interested in Compute Verification, please apply by 11:59 PM October 19 AoE.

Review

Motivation

There exist many open problems in ensuring the verifiability of frontier pacing commitments which are not currently tractable under a purely automated approach, where progress requires both large scale global coordination and technological innovation. Talent pipelines are required to get students and early-career professionals up to speed, but rarely (if ever) have evaluators with nonzero research experience.

Quick Takes

  • I enjoyed the small study group discussions and believe the cohort format works well. The course was split into 5 units, and we did one of them every day. I'd like to do a more in-depth review at some point, but it would likely be after my second take.
  • Student feedback was elicited ~2x daily and usually acted upon within ~minutes.
  • I consider Lens' content to be less mature than comparable frontier LLM safety research upskilling programs such as ARENA, but Lens' update rate is faster.
  • The course exposed me to fresh material I wouldn't have come across normally since it isn't in the places where I normally do my reading (e.g arxiv/wikipedia), such as:
    • Practice Guide to International Treaties.
    • Reading Statutes and Bills.
    • Implementation of NPT Safeguards Agreement and relevant provisions of the United Nations Security Council resolutions in the Islamic Republic of Iran.
    • ORDER RELATING TO SEAGATE TECHNOLOGY LLC AND SEAGATE SINGAPORE INTERNATIONAL HEADQUARTERS PTE. LTD.
    • Report of the International Law Commission.

Goals

What were the goals of the people who self-selected into taking this free course?

Participants from Hawking and Socrates group recorded goals in Unit 1, analysis below.

All names anonymized.

Student

Plans

Progress

Problems

Nine Inch Иails

  1. Prepare slides for an intro talk about compute verification.
  2. Test grasp of content with Anki flashcards.
  1. Deck created.
  2. Each session has a document tab for its glossary– cards could be created from these.
  1. Talk delivery still pending, further iteration needed.
  2. Exercises for concept handles (e.g Dresher, Fearon, Putnam, Gintis, Binmore, Skyrms, Farell & Newman) can be confusing.

Weezer

  1. Have a broader view of field's demands.
  2. Start project on Cryptographic AI verification.
  1. Discussed costs, timelines, and implementation of verification mechanisms.
  2. They are exploring projects related to cryptographic verification of AI computation, but started two days before the intensive!
  1. There are ~60 engineers working on verification, "the field" is pre-paradigmatic.
  2. This wasn't a focus in the intensive, but Lens will be running a Part 2 which will go further into technical components.

Portishead

  1. Make tool or product to help policymakers.
  1. Learned about geographical supply-chain bottlenecks.
  1. Underspecified.

DJ Shadow

  1. Finish intensive.
  2. Connect finance or cybersecurity work to material.
  1. Done
  2. Limited evidence of progress.
  1. DJ Shadow may have done all the readings and DJ Shadow may be able to correctly answer questions about them, but completing the week isn't strong evidence of that.
  2. The example DJ Shadow cites as being applicable to DJ Shadow's work does not seem to appear in the material.

Tricky

  1. Help with PauseAI US Advocacy.
  2. Contribute to a technical verification project.
  1. Developed questions about practical feasibility and governance.
  2. Did some learning, no contribution yet.
  1. PauseAI US is controversial.
  2. Again, technical projects weren't a thing in Part 1. But there's always Part 2!

Afrika Bambaataa

  1. Develop project proposal.
  2. Prepare for hackathon research sprint.
  1. Discussed cyber risks, HW/SW combinations, government funding.
  2. Did FFT on GPU
  1. No proposal recorded.
  2. On a smaller dataset where I suspect CPU performance would be comparable.

Massive Attack

  1. Finish course
  2. Conduct additional research.
  3. Articulate views on Substack and in job interviews.
  4. Land a role.
  5. Support progress towards a multilateral instrument.
  6. Mitigate catastrophic risks from AI at the global level.
  1. Done.
  2. Done.
  3. Done.
  4. Nearly done.
  5. *single tear rolls down one eye* ... so cracked.
  6. cracked, fissured, and groovy.
  1. Will not accept an endorsement on LinkedIn 😔.

Björk

No original goal documented.

Got a good baseline, alongside a better understanding of hardware verification's weaknesses.

Related Work

See Marta Kosmyna's review.

Acknowledgements

Thanks to Mark Kagach for helpful feedback and comments.

Session documents written by members of Hawking and Socrates group.

This work almost used compute resources provided by Side Channel Cloud (https://sidechannel.cloud), operated by Amodo Design.

  1. I don't remember who first said this, but I like the quote "There's a very high degree of heterogeneity in results that makes simple statements such as X is better than Y generally wrong except in cases where there's a very large difference in capability between models."
  2. Students cloistered into cohorts based on meeting availability. Hawking group met daily from 14:30 to 16:00. Sometimes we were visited by students from Socrates, which met earlier in the day.
  3. These band names spilled out of my Kaito KA350 Emergency Radio, set to Canada's longest running radio documentary, "Ongoing History of New Music", E1086, "The story of trip-hop".
  4. ...alright, not all band names are trip-hop. Okay so these guys just performed at Scotiabank Arena and I had made plans to go see them (under the Golden Full Harvest Moon! during the Mid-Autumn festival! PayPal on SeatGeek right before motherfucking Petrov day!!!) but in fact reality oh so rarely works out the way you perfectly arranged for it to go; as brother Sanā'ī would spew spli tha mf Moon in two Sun, bruh my truth-seeking balls finna enter tenseigan chakra mode fr.
  5. electric boogaloo
  6. Which hackathon? If it's an Apart Sprint, then the last one was on AI Incident Response, the next one's Collusion, but most relevant for verification is "Ho TAIG hast thou heard the decree?"
  7. remember kids, always account for Peripheral Component Interconnect Express Data Transfer Encoding Throughput Overhead, which varies by generation:a one-two 🎵 each byte of data is converted into a 10-bit data code 🎵 (20 to 25 percent!)a three-four 🎵 ampere's a hundred bee and twenty-eight over a hundred bee plus thirty, so basically that's an anti ochus vii over a phraates tee 🎵 (1.5 percent!)a five-six 🎵 old hopper's the same encoding at twice as many transfers-per-second 🎵 (64 GT!)a six-seven 🎵 for blackwell i won't get into fixed size F.L.I.T 🎵 (94.5 efficiency!)a seven-eight 🎵 and that's all cause vera rubin does away with traditional bus interfaces entirely* 🎵 (NVLink bitches!)*termsandconditionsapply verarubindoesnotuse PCIeforCPUGPUorGPUGPU communication PCIeGen6remainsforIO toNICandDPUandStorage talktoyourfounderaboutC2traffictoday.
添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论