←
Yonathan Arbel4,835 posts
Yonathan Arbel avatar

Yonathan Arbel

@ProfArbel

Let's build safe AI! Law prof @ Alabama Contracts, Defamation, Legal NLP, & AI Safety

◷ Joined July 2014▧ 884 Photos and videos
Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Sep 29

Enjoyed speaking with the wonderful Rosemary Church on CNN last night about AI governance and what it really means to win the AI race. Building safe, reliable AI is part of what winning requires.

Post attachment
Danny Wilf-Townsend avatar
⟳ Yonathan Arbel retweeted
Danny Wilf-Townsend@drmtown·19h

When these other AI public safety suits have been filed, one of my reactions each time has been "this feels like an argument that would be more effective coming from the California AG." And now the California AG has filed. A few thoughts on why this suit is different. /1

Post attachment

Between this, the Florida AG suit, and the LASST suit, there is now a small wave of lawsuits challenging frontier AI companies on the basis of broad public safety concerns. They involve different causes of action, and address real risks, but I think they share similar hurdles. 1/

Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·20h

lol everybody is hating on 6.2. I gave it one small writing test and it failed hard, but I haven't tested coding and design

Andres Sawicki avatar
⟳ Yonathan Arbel retweeted
Andres Sawicki@apublicgood·Nov 17

Just yesterday, we launched @spotlightschol, a new law review submission platform. Within hours, authors had signed up, articles were moving through the platform, & another journal had come on board. The response exceeded our expectations. So I want to share why we built it.

Post attachment
Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Nov 13

(Just to make my own confusion public, the Nvidia story appears more complex than I initially thought and despite my displeasure with Jensen, he may have been wronged)

Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Nov 13

maybe it's the Outlook thing, but so far I received zero (0) requests to sub for Jensen

Rob Wiblin avatar
⟳ Yonathan Arbel retweeted
Rob Wiblin@robertwiblin·Nov 13

Incredibly important role here - 'Pacing the Frontier' lead at Anthropic. Would benefit from someone very sharp and strong willed. Please please apply if you could do a great job at this: job-boards.greenhouse.io/ant…

Post attachment
Yonathan Arbel avatar
⟳ Yonathan Arbel retweeted
Yonathan Arbel@ProfArbel·Nov 13

🛠️Academics working with AI, a new powertool🛠️ ibid is a small site where you paste your text and it keeps your footnotes, formatting, and comments. You edit with AI, and get back formatted text to put back in Word. Private, quick, and easy

0:51
Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Nov 12

startrek level of cool

VoiceClaw Realtime 1.4.85 is out on the App Store — now with improved GPT Live on watchOS on cellular. Additionally, you no longer need to set up Tailscale or a Cloudflare tunnel to use your watch or iPhone to talk to your computer’s Codex app, OpenClaw agents, or Hermes agent.

Post attachmentPost attachment
0:50
1:25
justin curl avatar
⟳ Yonathan Arbel retweeted
justin curl@curl_justin·Nov 12

Yes! More legal scholars should be writing about frontier AI legal issues — so many intersect perfectly with their expertise. And yet not enough do, citing concerns that they’ll say something dumb or don’t know enough about AI. For what it’s worth, I don’t think you need special insight into the technical details around AI models. It’s often enough to imagine what would happen if the cost of intelligence approached zero or if analysis could be completed much more quickly. Some questions that seem pressing are 1. What can Flock teach us about AI-powered surveillance? Could ongoing automated monitoring count as per se unreasonable under the 4A? 2. Government deployments: What should AI models do when faced with unlawful or awful requests? They could refuse, push back but comply, answer normally, or escalate to a human? If they escalate, who do they escalate to? 3. If tort liability is to be a valuable deterrent, how can you pull forward timelines to a judgment? Are there procedural shortcuts that you can take to make sure it’s influencing behavior? 4. Is an order the president issues through an AI agent an “official act”? And are conversations with AI agents subject to executive privilege?

It was great to talk with @ghadfield and @zittrain about regulating frontier AI at HLS. This is an urgent issue and I hope more legal scholars start working on it. In my view, it’s high time for serious government regulation that aims to (1) ensure that all frontier labs are implementing state of the art security and control techniques (2) ensure that no lab is courting misalignment unnecessarily by doing very risky but commercially tempting forms of posttraining (3) pace the frontier and mitigate collective action problems and (4) rapidly generate evidence about misalignment by empowering independent evaluators, selected by the government or external experts rather than the companies, to conduct rigorous risk evaluations by intensively carrying out evidence-generation measures like model organisms, probes, and honeypots/sting operations. In short, although AI alignment and control are (unfortunately) an extremely nascent science, I think there is a lot that regulation can already do, and it can help figure out what else to do. The odds of catastrophe grow substantially higher every month we postpone enacting a serious regulatory regime.

Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Nov 12

We wrote a paper on it and now it comes to life (well, sort of)! "Governing Generative AI Beyond the Grave" asks what rights people have in their post-mortem emulations. We also survey Americans about their attitudes

An AI app that lets you comunicate with dead relatives.

1:46
Yonathan Arbel avatar
Yonathan Arbel@ProfArbel·Nov 12

Unrelated to my thoughts on the substance, the California lawsuit against the labs is a masterclass in Topic Sentences.

Post attachment