Director of AI Innovation & Law Program @UTexasLaw | Senior Editor @Lawfare | Adjunct Research Fellow @CatoInstitute |
- Tweets 6,647
- Following 1,840
- Followers 2,697
- Likes 5,748
This is a solvable problem.
Embedded evaluators in AI labs are far better than the status quo of nothing at all.
Yet auditors will lack teeth and trust until backed by government authority: orgs like METR will be stuck with no shield against the accusation (or reality) of capture.
I wrote for @TheAtlantic about the traps of "independent" audits, and what it takes to get this right: theatlantic.com/technology/2…
2
1
15
1,130
In the long run (if you'll indulge me), we need to reorient the incentives and norms of higher education institutions so that we don't see this kind of brain wane form again.
1
1
65
For related reading, check out this piece by me and @BharatKChandar via @lawfare: lawfaremedia.org/article/why…
1
62
On Wednesday, September 30, the next Lawfare & Libations will be held in San Francisco, CA with @KevinTFrazier.
The event is available to current, and future, material supporters!
1
2
3
2,502
Kevin Frazier retweeted
The Scaling Laws Roadshow continues! Yesterday, @KevinTFrazier and @mtokson had a great discussion on AI, surveillance, and privacy at @UNLCollegeofLaw.
1
1
1
126
#IKYK or at least I hope so...the difficulty caliberating under v. overrefusals in AI is one of the most important issues.
While the dangers from underrefusal are obvious, we ought not dismiss the dangers of foreclosing access to info.
Thomas Emerson expertly makes that case:
2
253
"kind of like minority report"
one underrated part of our reward hacking paper - we can detect when the model is contemplating a reward hack from its activations, before it even happens!
kind of like minority report
rollouts are stochastic though, so a flag means a higher chance of cheating, even if the model ultimately doesn't follow through
1
436
"Swarm"y times call for speedy measures.
Help me shape this course. I'm looking for the best minds on these topics to assist with selecting readings, guest lecturers, and designing exercises.
Yes, we will open source all content.
We're not missing a beat here at @UTexasLaw.
2
2
10
515
Are you at @redwood_ai, @AVERIorg, @farairesearch, @Irregular, or another evaluator?
Are you a CS prof looking to help law students learn about classifers, evals, or alignment?
Are you someone with experience at the labs and want to teach it forward?
Send me a note.
102
More here: lawfaremedia.org/article/boo…
63
A great example of (re)inventing the wheel.
AI governance is a wicked problem that can adopt / modify institutional frameworks from other areas.
Read this @lawfare by @RuneKvist, @CristianTrout, & Rajiv Dattani.
2
4
13
753
The future is here (in fact, it's been here), you just haven't read about it.
@hendrycks - July 2023:
1
4
289
full paper here is worth another read (in the event you missed it three years ago): arxiv.org/pdf/2303.16200
67
Kevin Frazier retweeted
🎙️ We chat with @KevinTFrazier, director of the AI Innovation and Law program at @UTexasLaw, about the legal and social impacts of data centers, the realities of workforce disruption, and regulating AI for child safety using existing consumer protection laws.
stackoverflow.blog/2026/09/1…
2
2
3,273
Reason 491 for something like an Office of Technology Assessment: a neutral body to provide demos of frontier AI tools to staffers & members of Congress.
If they can't use it (per internal rules), they should minimally learn how others are putting it to use.
via @NPR
2
6
401
If you don't already subscribe to @lawfare & @scaling_laws, here's your reminder to do so.
This was one heck of an AI conversation w/ @scaling_laws roadshow co-host @AlexBores, @NatPurser, @MackenZ_arnold.
Coming to a pod near you next week.
1
11
62
7,999
Kevin Frazier retweeted
I'm hiring an economics predoc at @ColumbiaSIPA, generally on applied micro projects, but with a large focus on the impacts of AI on economic outcomes using large-scale field experiments with many companies.
Please share widely: apply.interfolio.com/194121
4
51
183
12,927