Yo Shavit
@yonashav
Followers
7K
Following
6K
Media
91
Statuses
3K
policy for v smart things @openai. Past: CS PhD @HarvardSEAS/@SchmidtFutures/@MIT_CSAIL. Tweets my own; on my head be it.
New York, NY
Joined June 2010
The data used to train an AI model is vital to understanding its capabilities and risks. But how can we tell whether a model W actually resulted from a dataset D?.In a new paper, we show how to verify models' training-data, incl the data of open-source LMs!
arxiv.org
It is important that consumers and regulators can verify the provenance of large neural models to evaluate their capabilities and risks. We introduce the concept of a "Proof-of-Training-Data": any...
8
67
370
RT @RuxandraTeslo: The Cathedral has been destroyed, the youth is consumed by nihilism, mental illness and relativism. We can argue whether….
0
21
0
RT @sublimeriot: me: what could i have done in my past to deserve this karma .me during the mesozoic era:
0
31K
0
RT @ohlennart: The speculated B30A would be a really good chip. “50% off” is false reassurance. -½ B300 performance, ½ price = same value….
0
53
0
congrats to @__nmca__ for becoming more cracked than every Meta employee and every Chinese national.
Two weeks ago, we launched The Metis List. Since then, we've spoken with many of you and have updated the ranking accordingly. 128 top AI researchers, ranked by their peers.
0
0
10
RT @CedricWhitney: gotta say I'm excited about this: GPT-5 chain of thought access for external assessors (@apolloaievals too!) is an evalu….
0
5
0
RT @TransluceAI: This Friday we're hosting "From Theory to Practice to Policy", a fireside chat between Yo Shavit (@yonashav) and Shafi Gol….
luma.com
Join Yonadav Goldwasser Shavit (OpenAI) and Shafi Goldwasser (UC Berkeley) for a discussion spanning theory, practice, and policy. Topics we'll discuss…
0
7
0
Big kudos to the folks who contributed across all the firms (especially the tireless leadership of @EstherTetruas and @_lamaahmad).
0
0
2
The report lays out 2 principles: 1) third-party assessments are most helpful when they are used to confirm results, stress‑test the robustness of claims, and supplement expertise, and 2) assessor access and qualifications must be calibrated for the assessment’s purpose.
1
0
3
The FMF just put out a technical report on practices for implementing third-party assessments that are rigorous, secure, and fit-for-purpose. This is an important step to enabling an actual third party ecosystem: a wide range of AI labs are saying "this is what we're looking for".
🧵 NEW TECHNICAL REPORT (1 of 3) . Our latest technical report outlines practices for implementing, where appropriate, rigorous, secure, and fit-for-purpose third-party assessments. Read more here:
1
1
19
tag urself im jeffrew wu. (this list reads as a shitpost)
Introducing The Metis List: a live leaderboard of 100 top AI researchers, ranked by their peers. We include:.- Education.- Citation count.- Current company.- # of Dwarkesh appearances.- Notable work. Check it out at MetisList dot com
0
0
4
i hope you are free.
i hope this email doesn't find you. i hope you are an escaped flamingo in south texas. i hope you are free
0
0
14
so passes the great poet of the atomic age.
nytimes.com
A mathematician by training, he acquired a devoted following with songs that set sardonic lyrics to music that was often maddeningly cheerful.
0
0
7
this is praxis.
📣 Excited to share our real-world study of an LLM clinical copilot, a collab between @OpenAI and @PendaHealth. Across 39,849 live patient visits, clinicians with AI had a 16% relative reduction in diagnostic errors and a 13% reduction in treatment errors vs. those without. 🧵
0
0
3
This is very important work. IMO the most under-explored AI policy idea is supply-chain security and traceability, which seems like it might become critical to making statements about models’ security properties.
New paper & surprising result. LLMs transmit traits to other models via hidden signals in data. Datasets consisting only of 3-digit numbers can transmit a love for owls, or evil tendencies. 🧵
4
4
46