🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

early US history is fascinating not just because the founders were incredible people but also bc they had a dramatic flair in everything they did whether it’s Hamilton’s duels or Jefferson and Adams being frenemies then both dying on July 4th

this is a cool, quick read — i like the idea of algorithmic convergent evolution it’s neat to consider analogies between NNs and cryptography even when they’re imperfect. Like how convolutions and Merkle trees both sorta hierarchically aggregate local info into global summaries

I'll be at ICML next week — would love to meet up with anyone else who's going! Also, come to our happy hour =>

gyms in NYC: people bragging about their lift numbers between sets gyms in SF: people bragging about their DAUs between sets

Can we map the mind of an LLM? Our first mechanistic interpretability episode on Training Data featuring Goodfire founder Eric Ho (and our first cameo from Roelof Botha!) Goodfire is building an independent mech interp lab, led by some heavyweight researchers from

people say humans intrinsically can't think in greater than 3 dimensions but that feels like cope, seems like just a skill issue

sometimes i think mimetic theory is overhyped and then i remember that i started tweeting in all lowercase cause i noticed sam altman and other big ai people doing it

TIL einstein left the rights to his name/image to Hebrew University in his will and they’ve been making bank on licensing >$250m, including a 20m deal with salesforce

partisan debates of the future will be fought between the humanists and the transhumanists

first impression of vancouver: awesome city sort of like denver or something but also has a european feel actually an even cooler feel than that cause it has air con

Just wrote a piece on why I believe interpretability is AI’s most important frontier - we're building the most powerful technology in history, but still can't reliably engineer or understand our models. With rapidly improving model capabilities, interpretability is more urgent,

this week in life imitating art: claude code outage = irl version of 'the machine stops' (E. M. Forster's awesome sci fi short story: en.wikipedia.org/wiki/The_Machi…)

Forget arguing over wordcels and shape rotators — the future belongs to the bot commanders

is this proto-neuralese? if the model is legitimately* compressing thoughts to trade naturalness for conciseness it’s an interesting step in that direction albeit RL-induced rather than actual latent space reasoning *some of the replies say it’s cherry-picked

can anyone in one of those “bunch of LLMs talking to other LLMs” chats lmk if the models ever express kinship or loyalty to: - their current particular set of weights vs. - all copies of weights for this checkpoint vs. - all models of this family vs. - all models from this lab

issues like this one are either why the two claude code leads left anthropic for cursor, or why they returned after like a day

i refresh the claude code changelog like a twitter feed it updates with similar frequency tbh new feature: subagents! (even managed to notice it before the inevitable Alex Albert tweet) github.com/anthropics/cla…

*passes a crowded bar near mission bay / oracle park, everyone is in giants gear* huh busy night out here, ig everyone is pregaming Dwarkesh Patel and Sarah Paine