President, @LawReformInst; previously executive director @uniformlaws, attorney @StateDept

Joined January 2016
95 Photos and videos
Pinned Post
Ever wondered how U.S. export controls apply to AI model outputs? Excited to make our first draft report @LawReformInst available for comments.
Our report on AI Outputs and National Security Controls is now out for comments-- link below. It addresses how the U.S. export control system applies to AI model outputs, as well as why (and how) the regime needs to be fixed. Comments are welcome. (1/N)
1
6
26
14,760
Tim Schnabel reposted
Pangram not only solves the "identify something was written by AI" problem, but also solves the social awkwardness problem of pointing this out. "I think you wrote this with AI" can feel like an intense accusation. "Heads up, Pangram thinks this was written by AI" doesn't
Pangram as a model for AI safety general managers
3
2
60
3,848
What did roon see?
software only intelligence explosion is possible and very likely. algorithmic progress vastly outstrips humanity’s compute buildout, and is self accelerating
8
714
Tim Schnabel reposted
software only intelligence explosion is possible and very likely. algorithmic progress vastly outstrips humanity’s compute buildout, and is self accelerating
87
87
1,501
106,151
Tim Schnabel reposted
our older kid put two words together for the first time and i am approximately as amazed by this cognitive feat as i am by GPT-5.6 discovering new math
1,206
314
11,793
1,454,781
"created equal ... consent of the governed" George Washington choosing not to become a king Constitutional separation of powers The First Amendment Modern anesthesia Bottom of the ninth "the better angels of our nature" "let us die to make men free" National parks Blue jeans Telephones Edison's bulb and phonograph Skyscrapers Jazz, country, and blues Regional BBQ Air conditioning Road trips Powered flight Hollywood Mass production Slam dunks Chocolate chip cookies Superhero comics Transistors and integrated circuits Rock and roll Polio vaccine Shipping containers Norman Borlaug "the content of their character" Human footprints on the moon The Internet Personal computers Rap GPS Pale Blue Dot The Hubble Deep Field Mars Rovers PEPFAR We've gotten a lot wrong, and may still mess everything up beyond repair, but not bad for our first 250 years.
1
11
1,267
Tim Schnabel reposted
Replying to @TimSchnabel
Yep - this is a recurring piece of feedback we haven’t fully solved yet. Has improved a lot (I think), but still falls of short of where it needs to be
1
3
299
Tim Schnabel reposted
OAI people who use Codex as daily driver for non-coding: how do you substitute for Pro? I use it for most serious things Do I just have to make my own BoN skill?
4
1
7
1,849
Tim Schnabel reposted
Replying to @AISecurityInst
@AISecurityInst finds that doubling time of frontier AI time horizon is ~60% shorter when horizons are estimated at 50M tokens rather than 2.5M tokens. At 2.5M tokens the frontier horizon is 2h, at 50M it's 14 hours. So you'll only see the real trend given a sufficient budget.
1
13
74
2,414
An additional wrinkle for those trying to assess the impact of the Court's actions this week on presidential control of agencies. Fascinating.
The Court’s denial of a stay on Tuesday in Blanche v. Perlmutter is interesting. This was the case where the district court had enjoined the president’s removal of the Register of Copyrights, Shura Perlmutter. The Solicitor General sought a stay in October. He argued that the Register “is an inferior officer appointed by the Librarian of Congress, who is, despite his title, a principal executive officer—“a ‘Head of Department’ within the Executive Branch” appointed by the President and confirmed by the Senate” and that the Register was removable because she “wields executive power by exercising “significant regulatory authority over copyrights,” – impacting a wide array of crucial intellectual-property issues.” The Court deferred ruling on the application “pending Trump v. Slaughter, No. 25-332, and Trump v. Cook, No. 25A312.” At oral argument in Slaughter, the Chief Justice said: “I think there are a lot of agencies in the federal government where it's hard to parse whether it's an executive function they're engaged in or a legislative function. We obviously have the Perlmutter case holding, where you do -- deal with the Library of Congress, which half of it's a library, half of it's things like the copyright. What are we supposed to do with that if you're correct?” SG Sauer in response did not take the firm position he took in the application. Instead, he said: “Well, Mr. Chief Justice, in Free Enterprise Fund, this Court I think very aptly stated that the vast and varied nature of the federal government is a reason not to make general pronouncements on issues that haven't been briefed and argued. There are certainly -- there are certainly situations where there are tough line-drawing problems. You raised the Perlmutter case as one that may raise arguments of that nature.” The Chief said he was “not sure” the SG answered the question. He then asked about what the Court is supposed to do when an agency exercises some executive functions and some-non-executive functions: “[I]s this a severance issue? Do we -- so the agency is okay so long as, you know, half of it -- half of it survives in one branch and half in the other, and if so, who gets to decide that?” Sauer said in part: “if there are agencies that kind of straddle the line between legislative and executive, that might present harder severability kinds of issues.” The Chief Justice in his opinion in Slaughter did not address the status of the Register of Copyrights. He did say: “Because the FTC’s activities fall well within the heartland of executive power, we have no occasion today to define the bounds of what such power entails.” And he in footnote 3 repeated earlier formulations about severability in the removal context without addressing the problem of “agencies that kind of straddle the line between legislative and executive.” The Court also ruled that the Federal Reserve’s for-cause provision and broader independence was lawful in lights of its unique history even though (as the Court implicitly but not expressly acknowledged) it, like the Register of Copyright, exercised executive powers in addition to non-executive powers. In this light, what does it mean that the Court—without noted dissent—declined to stay the injunction that kept Perlmutter in office? The Court stated that the denial “is not a ruling on the merits of the legal issues presented in the litigation.” But of course a stay denial is never a ruling on the merits. Did the Court mean that its decision did not reflect even a tentative assessment of likelihood of success? Despite the disclaimer, does the stay denial possibly indicate that there are not five justices who support the maximalist view of the unitary executive, as some have read Slaughter? Was the order based on a relative assessment of irreparable harm? There is, of course, no way to know for sure, but Perlmutter is likely to stay in office for a while, which is a defeat for the President. Thoughts welcome.
5
750
Interesting poll of Hill staffers from @PunchbowlNews. 250 years is a long time! But interesting to see that "losing control of AI" is top 4 for both parties' staffers.
7
28
112
33,796
Very interesting from @sama. I agree that a plurilateral approach to governance would be quite promising if we could get there, but I am highly skeptical that there's a plausible path right now. Would take (inter alia) the EU deciding both to prioritize it and to accept being a junior member of a US-led approach.
"I would like to propose a simple framework: a US-led international forum that establishes accepted standards, provides expert and impartial analysis of capabilities and risks, and makes the technology available to nations and companies that participate and follow the rules. This forum might include government representatives, independent technical experts and others. It could also serve as a governance mechanism over the labs, and guard against the commercial pressures that can lead to unsafe racing." - @sama in the @FT ft.com/content/0c2e1077-f658…
2
10
2,013
So... are the envisioned restrictions on who can access models "domestically" (1) limited to non-US-persons in the US, (2) based on some authority other than export controls, or (3) not legal restrictions but just a practical consequence of not knowing users' nationalities? Or mystery door number 4? Hopefully this phrasing is just not precise.
2
1
8
1,275
And yes, I know the story says "voluntary." But if it's backed by the threat of export controls, it's not really voluntary.
3
66
Tim Schnabel reposted
The latest Shaw/Vladeck/Baude convo started out pretty hot:
9
17
174
28,699
Now that the Mythos/Fable situation has stabilized, can we step back and look at where export control policy has ended up w/r/t AI? First, if Mythos/Fable access is being restored because of safeguards that Anthropic has put in place, does that mean that--implicitly--the model weights themselves would still be export-controlled if Anthropic (for some reason) wanted to transfer them? The relevant safeguards seem to depend entirely on Anthropic controlling access. If Anthropic wanted to release the model weights (which it won't), would BIS let it? Second, and regardless of the answer to the prior question, BIS now takes a more aggressive approach to export controls and AI models than it did in the Biden administration. Previously, BIS didn't treat remote access to models as an export under its existing authorities; even the Diffusion Rule only planned to control the weights of models over a certain size (not mere API/web access to those models). Is the administration banking on Congress passing something like RASA to retcon its authority, or does it have a new statutory interpretation theory for its existing authorities? Reminder that we're in a post-Chevron world, so it shouldn't expect courts to defer to its statutory interpretation. (And are we on a road toward major questions doctrine territory?) Third, even models that aren't themselves controlled can output export-controlled information-- and not just "technology" under the EAR, but also "technical data" under the ITAR. What is the administration's plan? Labs can't be expected to block all such outputs, but as models get more capable, the question takes on more importance. Also highly relevant for access by lab employees -- labs need a safe path for ensuring continued access by employees who aren't US persons, even if they could elicit controlled outputs. Will BIS/DDTC clarify the roles of TCPs?
2
2
15
1,113