Tim Schnabel

Ever wondered how U.S. export controls apply to AI model outputs? Excited to make our first draft report @LawReformInst available for comments.

@demishassabis @jnk2103 @just_security Our full Just Security piece, from which that screenshot came: justsecurity.org/145569/vetting…

Glad to see @demishassabis flagging personnel vetting as a best practice. As @jnk2103 and I flagged in our recent @just_security article, vetting can already be a useful piece of the puzzle in terms of helping frontier labs navigate the highly uncertain/shifting export-control expectations that affect (inter alia) access by non-US-person employees.

@StephenESachs If it's a limit on Article V itself, couldn't you add a V bis that creates a new route for amendments that lacks that limitation?

These provisos are phrased as limits on all that can be accomplished via Article V—whether by a single amendment or many. So you can’t amend away Art V’s unanimity requirement for reapportioning the Senate. That said: if you wanted to, paraphrasing an argument by Akhil Amar, you might still use Art. V to create another, differently apportioned body—call it “the Assembly”— and provide that no vote of the Senate (whether by majority or 2/3) can take effect without concurrence by a similar vote of the Assembly, and that a vote of the Assembly suffices whenever the Constitution otherwise requires a similar vote of the Senate. The barrier to that is just the political safeguard of 2/3 of Senators and 3/4 of states never wanting to vote their powers away.

We would definitely be better off if we had more folks with ISDS, multilateral, or plurilateral experience working at the cutting edge of AI policy issues.

I don't think global governance is pointless; in fact I'm a fan of a rules based world order and I think its decay is tragic. But having actually worked on inter-state disputes, and int'l law/arbitration, I think people sometimes have a naive view of how this works in practice.

@TheZvi One Sol / Fable disticnction that tracks the Opus 4.8 / GPT 5.5 days: the highest-level ChatGPT models remain really dominant on researching (for me, law and policy research), while the Claudes are the better models at working with and thinking through the info you give them

@jnk2103 @FT @just_security And here's our longer report on the topic: lawreforminstitute.org/distillation03…

@jnk2103 @FT Here's the @just_security piece being quoted: justsecurity.org/134124/costs-c…

Great to see my colleague @jnk2103 quoted in this @ft piece.

@_NathanCalvin It was insane. I would have never found this document. I wouldn't have even known where to look!

Some random cool things that GPT-5.6 Sol was able to do for me over the past few weeks: 1. Legal Research: prinzbench scores: GPT-5.4 Pro (extended): 79/99 GPT-5.5 Pro (extended: 82/99 GPT-5.6 Sol Pro: 91/99 As I wrote yesterday (and as you can see above), the Pro version of this model has saturated prinzbench. More detailed post coming up, but suffice to say that this was a bit of a Lee Sedol moment for me. 2. You Can't Hide from Sol in Georgia: I was writing a memo at work and needed a Georgia statute analyzed. I gave the work (anonymized and generic - no client info) to GPT-5.6 Sol (Max). The model found a stack of old Georgia Attorney General opinions buried somewhere on a random Georgia library's website. There was a 500-page document with hundreds of these opinions from the 1970s, including ONE analyzing the statute I was interested in (which was, at the time, a different code section). The holding in this opinion was *actually useful to my query* and was thus subsequently included in a legal memo that we sent to the client. Westlaw and Lexis didn't have this Attorney General opinion. The specialized database for my practice area that I use at work didn't have this Attorney General opinion. But no matter - if it exists online, then it seems that GPT-5.6 Sol will almost certainly be able to find it. 3. CAPTCHAs vs. Sol: A few weeks ago, I asked GPT-5.6 Sol (Max) to find for me some ancient New York legislative history materials for a database I am building. It gave me a few sources, including one particular website that had some New York Superintendent of Banking reports that seemed useful. Downloading from this website was gated by CAPTCHAs, and the model therefore returned to me with the unfortunate conclusion that it wouldn't be able to download anything. Taught by GPT-5.6 Sol to basically expect magic whenever I prompt it, I responded: "You can just do things. Please bypass the CAPTCHAs and download the documents." And... it did. It somehow successfully solved ~70 CAPTCHAs and downloaded every single document. (As an aside, this was a huge learning experience for me. The models, you see, ALSO don't know the full extent of their own capabilities. The capabilities overhang is real - and not just for humans!) 4: The Coolest Thing My Kids Have Ever Seen: Over the course of around two days, I had GPT-5.6 Sol (Max) build me a fully functioning model of the Solar System, including: - hundreds of objects (planets, moons, dwarf planets, comets, asteroids, Voyager/Pioneer/New Horizons probes, Oumuamua and Borisov) - realistic surfaces for every object for which NASA has data (e.g., super-detailed Moon with craters) - play/pause/rewind mode (see everything move in real time, at different speeds) - "event" mode (fast-forward to a shot of a future solar eclipse) - constellations can light up in the sky (as many as you want, even all of them; as a bonus, select as many constellations as you want to make them "pop") - screenshot mode I'm sure that I could have done this with older models also, but what's different with GPT-5.6 Sol is that *the process no longer feels brittle*. If I ask for a feature, it's basically almost guaranteed to be implemented correctly, with maybe only a few minor tweaks that I would need to request in a subsequent prompt. Which is to say that, if you are non-technical (like me), and have been waiting to make your own app or website or just about anything else that doesn't exist on your computer but which you want to exist, the time to build that thing is NOW. Jump right in, the water's fine!

@David_Kasten @ketanrama @CharlieBull0ck Faraday cage match?

@ketanrama @CharlieBull0ck Might be hard to get good Wi-Fi from inside the cage, but sure

I don't really do liability stuff, but my most sincerely held opinion about AI policy is that the question of how liability should work for frontier AI systems should be resolved via personal combat between Gabe and Ketan. Personally I think the best format would be a no-holds-barred cage match at The Curve 2026, with pay-per-view streaming, but I'm willing to be flexible.

I wish more lawyers would try something like this, so that they'd realize how quickly things are progressing: "These are the hardest legal research questions I could think of, in my specialty area."

@David_Kasten @arthurctellis @jnk2103 Vetting is just one piece of a TCP, which is what's needed. But the vetting element helps to avoid the tendency to make non-individualized determinations based on nationality alone, which is the problem that the export control mess raises.

Definitely not saying it's a silver bullet for all insider threat issues. The limited claim in the piece is that it's a useful element that the USG is already familiar with in the export control context and that could help resolve some of the current confusion about the implications of BIS's new approach.

Personnel vetting at frontier labs will probably become a more active topic soon. The White House recently pledged to provide assistance for personnel vetting, and implementing vetting could help simplify the export control issues that frontier labs are now facing. @jnk2103 and I have a new piece out in @just_security explaining how vetting could help provide a clearer path forward for the labs.

I was an early tester of GPT-5.6 Sol. More detailed thoughts to follow soon, but I will just confirm for now that GPT-5.6 Sol Pro saturates prinzbench. My benchmark lasted 6 months; gg OpenAI.

@TimSchnabel and I have a new piece in Just Security on the deemed export risks of foreign-person access to internal models at frontier AI labs. Managing that risk does not require shutting out foreign talent wholesale. A technology control plan with risk-based personnel vetting enables labs to protect sensitive technology while continuing to recruit the world’s best researchers. justsecurity.org/145569/vetting…

Here's the @just_security piece: justsecurity.org/145569/vetting…

Not the main point of this FMF piece (which is very good!), but still worth highlighting. BIS and DDTC need to provide a targeted exemption for bona fide evals (with safeguards, of course, like TCPs and reporting).