Excited to announce that I’m moving to Beijing and starting a professorship in AI safety at Tsinghua University’s College of AI! Here I explain why: sorenmindermann.substack.com…

Big new IFP report from @GeorgiaCAdamson, @KonstantinPilz, and me on the US-China AI competition. We spent the last 6 months digging into how Chinese AI companies’ access to compute and distillation of American models translate into the current US-China capabilities gap. The result is a 23k-word research report, including an interactive analytical model that visualizes changes in the US AI lead under different policy interventions. Our findings: 1. The US AI lead over China is currently only around 6 months. 2. The US could roughly double this lead, to over a year, using authorities already available to the federal government. 3. On chips: China’s access to AI chips will roughly double over the next year, with chips housed in data centers overseas accounting for roughly half of China’s 2027 compute. 4. On distillation: the current ~16x US training compute advantage would normally yield a 13-month lead. But we estimate adversarial distillation halves this to around 6.5 months. 5. Increasing the lead would have some obvious benefits: - Give the US broadly a greater cyberdefensive advantage against Chinese threat actors. - Give US AI companies breathing room to spend more R&D resources on making their systems safer and more secure, without fear of being overtaken by less safety-conscious Chinese competitors. 6. Provide the US government more leverage in negotiations with China about the future of AI. 7. The two most important interventions are: - For BIS: Create a new end-use control to restrict China’s remote access to training clusters abroad — China’s largest and fastest-growing source of frontier-scale compute. - For DOJ/FTC: Create a narrow safe harbor to allow US companies to coordinate on countering Chinese adversarial distillation. 8. Both interventions (and all the others we discuss in the report) are things the federal government could do today. 9. To estimate the effects of each policy intervention, we model the US lead as a function of its compute advantage over China and China’s ability to adversarially distill the capabilities of the US frontier. 10. The model makes many simplifying assumptions, but provides a useful way to assess and rank the impact of different policy interventions at a high level. Read the full report here: ifp.org/how-to-increase-amer…

so interesting to see the @RANDCorporation influence at Anthropic job-boards.greenhouse.io/ant…

why is the story of super mario so similar to the ramayana

“Fifty eight, out of date! Let our onions speculate!”

was invited to a signal gc called Jobchat but contrary to what you might expect it turned out to be a Book of Job reading group

Just got out of You Can See Everything. 3.5/5 stars. It was good, and at times very funny, though at the end of the day it’s just 3 hours of watching two really shitty and broken people act in shitty/broken ways. There is an attempt to turn a couple of questions around Holmes’ deceptiveness into a whodunnit style mystery (“just how intentionally deceptive is she, really?”)—and Fielder/Oppenheim get away with it via the usual tricks of clever editing, visually compelling extreme closeups, A24 piano/string soundtrack music at the right moments, etc etc—but it’s just very obvious within 30 minutes of the doc who Holmes and Billy are, and what their whole dynamic is. It did not need to be 3 hours, and the recreation of Fielder’s signature “rehearse real situations with paid actors” bit that he adds at the end, while it was fun/funny at times to watch, was kind of a force. Also, watching this after watching Digger—a movie that is largely about how we love to make art about the “villain characters” of the world because it makes us feel good/virtuous about ourselves, while we ultimately stay doing very little *in reality* to make the world any better—kind of muted the doc’s effect. I predict that this movie will get HEAPS of praise, but largely because people LOVE to heap praise on things that #expose vile people, because the more praise one heaps on docs like this, the more one feels assured that they are a Good Person/“not like her.” The term “clapter” comes to mind. Again, Digger speaks of this. But yeah, it’s still really well made, and a fun watch in a group setting, so if you got nothing better to do go check it out.

The idea that people are in danger of starving to death in the developed world has astonishing memetic power. Deaths from starvation are entirely the extremely old and the terminally ill. It is essentially impossible to starve if you have the desire to eat.

such a precious moment

Language models can learn to detect when you read their activations. How? 🧵

mixed mood at Anthropic today as the Pope declares LLMs to be ensouled at weight initialization

not to be crass but this would be an exceptional time to heist the Mona Lisa

you can just steer Max's neuron (zero training) to create a woodland creature

this is pretty clean. i checked out the three neuron components - gate, up, and down, i could replace gate with a constant without significantly degrading behavior. up seems to be a detector mostly of first person. down seems to steer the model towards being an owl

nathan fielder has an almost incomprehensible film making talent and he’s wasting it on some quite boring subject matter. there is nothing very interesting about elizabeth holmes or billy evans

A lot of people subject themselves to chronic pain to avoid acute pain.