Normally just leave this stuff on youtube but figured I'd give it a try posting it here as well. My talk about the history of large language models and natural language processing

Being a great hypeperson that can move people is extremely underrated. Everyone wants to go in their own direction and work on their own thing, but convincing a group of people that this specific thing is important enough to sacrifice their own thing, that is powerful

Not that I do this very often but every time I accidentally do rm -rf in the wrong place, it is very inconvenient that authorized_keys is one of the first things that gets deleted and then the next time I try to ssh in I cannot

One of my favorite lessons I’ve learnt from working with smart people: Action produces information. If you’re unsure of what to do, just do anything, even if it’s the wrong thing. This will give you information about what you should actually be doing. Sounds simple on the

I don't mean to brag but sonnet says my sophisticated observation deserves a nuanced answer

Incredible result

Really happy to announce I've joined Nvidia on the multimodal retrieval team 🎉 big open source OCR models to come

One of the dirty tricks I see virtually everywhere is acting like models are a technical advancement when really they're just a distillation of some other model that had an actual innovation or better data, or bigger model size. It often hides the signal in the noise. Bonus

>be me >reading state of the art small model paper >i ask: “Is it a good model or just distillation?” >paper:“It’s a good model, sir.” >read data section >it's a distillation

Look at the data. If you have images, look at them. If you have text, read it. If you have time series plot it. No aggregates. No EDA. Just look at it

Casually popping into #NeurIPS throughout the week if anyone wants to hang out and chat document ai, rag, ocr, diffusion lm's, looped transformers, arc agi, etc.

😢 This is one part I'm not enjoying about neurips. Sent an RSVP for this fairly early. Either have to sign up for things crazy early or social engineer your way in.

I enjoyed neurips. Met lots of awesome people, but I can't help but feel like I didn't get full value out of it. Hard to find people interested in the same areas. The crowd I mostly hung with stemmed from old San Diego friends working on mech interp and that's just not my jam.

I've had multiple flights where one person decides to just pick a random seat and then that cascades and forces a whole chain of people to just find an alternative. The flight attendant tried to get us in the right seats and then after she realized how far lost everyone was she

One of the social causes I'm passionate about is that people in the middle seat deserve the right to both armrests. Middle seat would be much less miserable if people respected this. Aisle gets extra leg room. Window gets something to lean on. Middle gets both armrests.

Nemotron 3 Nano is released (and it's a banger), but more importantly: It's just as open as the last one, and it's ONLY THE FIRST ONE. Super and Ultra: OTW > Model Weights - RELEASED > Pre-Training Data - MOSTLY RELEASED > Post-Training Data - MOSTLY RELEASED > RL Environments -

Saw people asking gpt5.2 to try to make ascii are of certain things. Tried to get it to turn my dog ascii... It sure did try

arxiv.org/abs/2512.13586 This looks really promising. Been really excited for diffusion lm's to eventually breakthrough

The Iranian regime arrested my cousin. We are cut off, with no news from my parents and family and no way to help. Please be our voice and don’t let this be ignored. @Amnesty Human Rights Watch UN Human Rights #Iran #DigitalBlackOutIran #Shayan_Abadi