Latest Turing Test results: GPT-4.5 is now capable of simulating a hyperrealistic persona which is judged to be more humanlike than actual humans.

The Scott Alexander & Daniel Kokotajlo episode. Scott and Daniel break down every month from now until the 2027 intelligence explosion. Misaligned hive minds, Xi and Trump waking up, automated Ilyas accelerating AI progress. I went in quite skeptical. But I learned a tremendous

"How, exactly, could AI take over by 2027?" Introducing AI 2027: a deeply-researched scenario forecast I wrote alongside Scott Alexander, Eli Lifland, and Thomas Larsen

Excited to reveal a project I've been working on the past few months - an AI forecast with Daniel Kokotajlo , Eli Lifland, Thomas Larsen, Jonas Vollmer and others.

I have long said that Dwarkesh runs the best podcast because he is willing to go deep. We spent many hours arguing about takeoff speeds (and everything else) here and I had a blast! Scott Alexander volunteered to help us with AI 2027, many thanks to him both for assisting with

This scenario does a great job spelling out what the development of really powerful AI might look like, under somewhat aggressive but IMO plausible assumptions about AI progress. I think that it's like 25% likely that things progress at least this quickly.

As part of the design of AI-2027 I got to make this really cool animated dashboard (on the site it keeps track of the state of the world for you as you progress through the scenario). The whole scenario is worth reading, but let me share some design notes in this thread 🧵

screenshotted with permission

ME, REPEATEDLY, OVER THE PAST SIX MONTHS: Daniel! Publish already! Your predictions keep coming true before anyone even sees them! DANIEL: Oh, that one wasn't even a real prediction, everyone knew that would happen.

okay time to get serious

gf: "and this one's about insect welfare..."

tariff is short for tarif and only tarif

AI 2027 features 100M H100-equivalents, 10^28 FLOP models (1000xGPT-4), and 300K AI copies running at 50x human speed by the end of 2027. Takeaways from the compute forecast behind these predictions. 1/7🧵

grilled cheese sigilism

I've been arguing for years that internal deployments are plausibly most of the risk, and I've been unhappy about focus on the external deployments--they're most of the (relatively unimportant) downside of AI now, but it would be a mistake to fixate on them. Glad to see this!

please use bouba objectives like GFlowNet instead of kiki objectives like PPO or DDPG

Commander: I'm thinking we fortify these positions. Fellas? [ATTACHMENT: plan.png] (😂158379) (👎67943) (🤦‍♂️38264) (🙄28267) (👍1)

You can leave a horse to water but you can’t have a penis on your balls