Avi Krishna

New blog post! I used 976 million tokens to figure out the personalities of different AI models We find that character training recipes are converging, that labs are trying to reduce sycophancy, and discover some models are more creative than others! avikrishna.substack.com/p/eliciting-fr…

@_heyglassy @theo Halide caméra 100%

Theo actually doing t3 codes support at the underscores show smh

the whole controversy here is so overblown cursor spent ~75% of compute on non-pretraining and probably spent a ton on CPT too they scored so highly on benchmarks for way cheaper (this is really hard!!) and this is the best marketing Moonshot (Kimi) is going to get for a while

@0xSero Is this really fair to say? Even Qwen has closed their mosr performant model. They are still an open source company I would say

You think I was joking? Say goodbye to open weight models, we frankly fucked it up for ourselves. x.com/0xSero/status/…

@irvinxyz ur wrong

I don’t think robotics is a data problem .

sf gets 1.5x the sunlight of New York btw

this is an eval that tries to make the directional claim that 'language models can't generalize' by measuring their performance on brainfuck horrendous... you can't claim to measure generalization via OOD task performance if the OOD tasks are genuinely way harder

@ellev3n11 decentralized computing gotta be the future man

fun fact about the Composer 2 RL run: we ran training distributed across 3 (sometimes 4) different clusters around the world using some secret sauce we built together.

@paulg bro loves his wife

One of Jessica's most mystifying qualities is how often she matches the places we go. Almost to the point of camouflage sometimes. I don't think she consciously plans it. She just has this uncannily powerful fashion sense.

every day I miss Claude the Alligator man bro was such a cutie patootie

the truth is that model big model developments happen at most companies at the same time

@heywildrich We're about five years out from "Williamsburg Simulator," where you manage a hole-in-the-wall hipster bar/music venue. The win condition will be booking MGMT after ORACULAR SPECTACULAR drops.

recently lost my photoshop subscription and went through 8 editors before I could found one where 1. I could rotate text and maintain the vectororization 2. Use Helvetica ends up that Abobe has a free Photoshop for web that does this; even GIMP and Photopea couldn't do it??

@wregss oh I saw your ICCV paper, was good stuff! commenting for reach.

Hi ML Twitter! My Summer 2026 internship unfortunately fell through last minute 😵‍💫 If your team is looking for interns, I’d love to connect - RTs appreciated 🙏 My website: aniketrege.github.io

guy who hates SF because of the weather is certainly a new one

@mayukh_panja unsurprisingly, remove berlin from germany and per capita GDP goes up

Berlin is an odd place. When I was working for Deutsche Bank, I found out soon that this is not cool and people don't like it. You are supposed to act a little guilty about working for a big bad investment bank and enabling the evil capitalists. However, if you have a story about how you snorted cocaine in the UBahn that is supposed to ballsy or whatever. You might think I am joking but that is the vibe of Berlin. I have been asked point blank if I felt uncomfortable about working at DB. In case you are curious about how I really feel, obviously no. Investment bankers add way more value to society than drug addicts. It is funny that this is even a matter of debate.

@NoCommas training on test is all you need

man evals suck so bad and i am making a blog post on contamination soon. the problem is so widespread across the ecosystem even if you don't get access to test, the labs are just paying human data companies / creators of the evals themselves to make identical questions that they can hillclimb on

@rennyzucker Come on... This is certainly beat by Hausmannian Paris or even Edinburgh For a modernist, maybe beat by Chicago or even by FiDi near Trinity Church or even near City Hall?

@EverettRandle pre Industrial Revolution shrimp were probably better off but unborn shrimp? now you're talking

There's never been a better time in history to be a shrimp

@bubbleboi ball's in your court, bubble boi

I made a killing this week and I’m trying to spend money on some mutuals