Evan R. Murphy

📘Preparing for the AI Crisis: A Plan for Canada📘 Frontier AI labs expect smarter-than-human AI in as little as 1-3 yrs. Governments are unprepared for the coming safety, job, and geopolitical crises. Our new white paper outlines a way forward: aigs.ca/news/251021-ai…

Recording of my talk, "AI Risk: Can We Thread the Needle?" at EA Summit Vancouver '25: youtu.be/x53V2VCpz8Q?si… 6/

#EASummitVancouver #ArtificialIntelligence #AI #EffectiveAltruism #DoTheMostGood 5/

I'm excited to get to speak at EA Summit Vancouver this September 19-20! We may only have 5-10 years before AI fundamentally reshapes human civilization. In my talk, "AI Risk: Can We Thread the Needle?", we'll walk through three plausible futures: 1/

JUST IN: @CAGovernor Gavin Newsom signed #SB53, legislation further establishing California as a world leader in safe, secure, and trustworthy artificial intelligence — creating a new law that helps the state both boost innovation and protect public safety.

METR is a non-profit research organization, and we are actively fundraising! We prioritise independence and trustworthiness, which shapes both our research process and our funding options. To date, we have not accepted funding from frontier AI labs.

A new letter from 47 academic researchers in support of @Scott_Wiener's SB 53 was covered this morning in Politico. It's great to see such strong support from across California for this important measure!

📢 New Poll: Canadians Overwhelmingly Want Action on AI Impacts. But No Party Has a Plan. Read more in the thread! #Elxn45 #cdnpoli #AIgovernance #AGI #OurFutureWithAI

@gregpr07 Thanks

@evanrmurphy Hey, cloud.browser-use.com/billing -> manage

@gregpr07 I bought a month of cloud browser-use but I can't find an account page or any way to cancel the monthly subscription.

...The key is policies that automatically trigger when AI hits certain capability milestones. If AI advances rapidly, strict requirements automatically take effect, but if progress slows, the requirements relax accordingly." Nice insight from arxiv.org/abs/2310.17688

"We need governance measures that prepare us for sudden AI breakthroughs, while being politically feasible despite disagreement and uncertainty about AI timelines...

@GaryMarcus > Models that are significantly better than GPT-4 don’t exist. They could exist but just haven't been released.

13.5 months post GPT-4: • Models that are not as good as GPT-4 have become widespread. • Models that are roughly as good GPT-4 have become common. • Models that are significantly better than GPT-4 don’t exist. If that doesn’t hint at diminishing returns and potential convergence at an asymptote absent a breakthrough, what would?

And here's a link to Anthony's LinkedIn post which I shamelessly copied for most of this tweet thread: linkedin.com/feed/update/ur…

It was a privilege to get to work on this with Anthony M. Barrett, @davidevanharris, Krystal Jackson, @Lemkchen, @JessicaH_Newman, @BNonnecke, @areddie89, Stuart Russell, @bbiinnyyuu and other UC Berkeley colleagues who provided thoughtful review and comments!

UC Berkeley colleagues and I submitted a response to the NTIA RfC on Dual Use Foundation Artificial Intelligence Models with Widely Available Model Weights on March 27. A thread 🧵summarizing key points our submission (copying colleague Anthony M. Barrett's great LinkedIn post):