Leading AI companies have agreed to share their models with the US AI Safety Institute for pre-deployment testing, says Gina Raimondo. This and more in my profile of the Secretary of Commerce. https://t.co/XWknuRnRwr

https://t.co/a0qc5pCaKe

I was pretty surprised by their estimates of how much of the cost of developing a frontier AI model is accounted for by researcher compensation. ~30-50% for the four models they made estimates for!

A new report from Epoch AI examines how the cost of the computational power required to train AI systems has been increasing over time. According to their estimates, it's been doubling every nine months

Audrey Tang is stepping back from her ministerial duties to embark upon a world tour to promote the ideas that she helped flourish in Taiwan—ideas captured in Plurality, a book Tang has co-authored with E. Glen Weyl and more than 100 online collaborators https://t.co/QJsIU4nHCh

https://t.co/yy2SL2UFut

Excellent reporting from @KelseyTuoc

(Out to @billyperrigo for the reporting assist)

Microsoft and Amazon are starting to compete with their investees, OpenAI and Anthropic. Can the smaller companies stay ahead of their compute-rich big tech backers? “For the next few years, I don't have concerns about this,” says @jackclarkSF https://t.co/OBspFAefvI

I spoke with Michelle Donelan, Secretary of State for Science, Innovation and Technology, about the AI safety testing agreement she just signed on a flying visit to DC. https://t.co/OJRhC7UvqM

Exclusive: New research provides a way to measure whether an AI model contains potentially hazardous knowledge, along with a technique for removing the knowledge from an AI system while leaving the rest of the model relatively intact. https://t.co/KTEg0fhDkV

On Tuesday, the House launched a bipartisan Task Force on Artificial Intelligence. I spoke with its members to understand their priorities. https://t.co/H5CUSNTH6W

“I can't help but read it as a 2000 year old blog post, arguing with another poster,” says @natfriedman . “It's ancient Substack, and people are beefing with each other, and I think that's just amazing.” My piece on the Vesuvius Challenge winners: https://t.co/J1Bqkmhz9O

Read the piece here: https://t.co/a2mpvLqd2W

I reported on the NAIRR pilot and the problem it seeks to address. The gap between industry and academia on some measures is pretty stark.

Link to the full piece 👇 (6/6) https://t.co/uPN8oOZfXz

Many questions feed into these wildly varying predictions: How impressive are current AI systems? Will simply scaling them up produce AGI? And a lot could hinge on these forecasts, given the risks that many experts worry AI might pose (5/6)

And if you ask people with a track record of making accurate predictions—the superforecasters—they're even more skeptical (4/6)

If you ask a wider set of AI experts, they're less bullish. While around 10% make similar predictions to scaling hypothesis believers, most think AGI is a couple of decades away at least (3/6)

If you ask scaling hypothesis believers—which includes the leadership of many prominent AI companies—then AGI is probably less than ten years away. A model developed by AI forecasting org Epoch arrives at a similar conclusion (2/6)