i love my M3 but when i tried to train some models on it things got pretty spicy! with prices the way they're going i'm gonna be on cloud/servers for the next couple years at least

gemini: "It looks like a tombstone or the top half of a thumb."

i'm watching claude code do real science on real data in real time. a sight to behold

Andrej, it comes with a manual! code.claude.com/docs (i should read it)

found a bug in my code that severely hindered data processing. fixed it and now my models are 1% worse. i am officially doing deep learning

peter thiel get the fuck out of california challenge

Have you ever eaten at Wahaca? Yes, that’s correct: they spelled it phonetically and trademarked it.

a bicycle for the mind

i tried to teach myself "AI" in 2019, starting with the classics. I read Minsky (Society of Mind, '86) and Franklin (Artificial Minds, '95). fascinating books but very challenging, often just plain confusing. they have almost nothing to do with how AI works today 🙃

I’m surprised at how excited I am about future local LLMs. But it looks far off for my uses. I would want models at least as capable as early 2025, and you’re looking at like >$20k and a hefty power bill to serve those at home.

Gemini says:

the models want to learn que voulez-vous the data are crap c'est la biologie

excuse me?!

set up a bunch of training runs and most of them SUCKED. then run R008-G went and made my best model ever??? 🥹 was about to scrap the project entirely or start over or phone a friend or something, but now? the suffering continues still probly good idea to phone a friend 🙃

gonna keep sharing my experiences using LLMs for Science. these tools are being rolled out in real time. in some ways they are already transformative. they are also flawed, and their development raises serious legal, ethical, and societal concerns. seeing both sides is the only way be truly informed

1/4 LLMs learn language competency, factual common knowledge and factual tail knowledge. Factual knowledge has a heavy tail and should be stored in a database. “It doesn’t belong in the model weights”: Cornell's Kilian Weinberger at the Simons Institute. Video: simons.berkeley.edu/talks/kilian...

This is the plot of THE ABYSS with one very specific alteration to the protocol

sorry, having too much fun and doing too much science. acting responsibly and leveraging the tools in powerful ways. the problems i'm working on are very challenging. i'll take all the help i can get

finished wiring up my fluidics control system yesterday while Claude overhauled the UI frontend/backend. I will open source the control software once we've testing it fully and completed real experiments. the UI is.. crazy. more advanced than comparable commercial systems in certain ways.

i am doing computer without needing to know computer, and honestly? that's beautiful as a chemist/biologist i'm very comfortable using the tools of science to work with systems i only mostly understand

Heyyy the Baronas paper was co-accepted at Science 😎 congrats to all the authors! This must be, I hope, very gratifying for Allon and Linus personally as they are co-inventors of the inDrops technology! i share more of my own capsule experiences in the thread below

Maybe we can solve the affordability crisis by telling people they’re wrong, it’s already solved, and showing them lots of charts. Have we tried shouting and holding charts at the same time?

Hey @parafilmm.bsky.social let the good times roll www.nature.com/articles/s41...

bro i'll take some AI cites idc give em here

git status

the "CAGEs" semi-permeable capsule technology from Allon Klein's lab came out in Science yesterday. So many great experiments in this paper. I've been using the CAGE chemistry lately, and it's the best SPC formulation I've tried, out of many! i think their chemical intuition is spot on

talking with some graduating biotech PhDs and the salaries they're quoting are completely flat over the last five years. you bet i'm paying a lot more attention to those numbers than national statistics

people's hatred of AI is logical and justified. they make bad arguments, but they're right on the merits. AI is not being developed in a way that will do the most good for the most people

i'm rich!! how much can i get um.. ah.. 3Gb of DDR3 and a 256 Mb, 133 MHz card? freaking academia stealing shit isn't even worth it

I am literally in this photo and I don’t like it

MONETIZE THE DYSFUNCTION. SOCIALIZE THE PAIN

why does Gemini claim to be bad at patent analysis? and then it actually is really, really bad at it?? this is a top use case for me. a recent experience was a complete waste of time. felt like stepping backwards two or three years

Gotta hand it to Texas for living out their pro business principles. You can’t deny they know how to build

Just had a long, powerful, and pleasant lucid dream. in a place I know well but haven’t been in years. Realized I couldn’t actually be there. I got to explore the dreamscape 😌

Becometh the mind flayer

Berkeley has its "Savio" home cluster, which everyone seems to dislike. BAIR, i believe, has other resources i probably don't have access to. A larger "Doudna" cluster won't come online until late next year. if i do this i'll probably shell out credits to AWS, Voltage Park, or CoreWeave

i am, of course and unfortunately, thinking of training a mid-size autoencoder (for science). so doing some GPU shopping to find my pain threshold

bump. any fixes out there? between my OpenAI perma-ban and no Gemini3 from work... i really need claude to work for me 🙃

Obvious problem with working with agents on complex projects is how do you know if everything was done properly, the way you meant it. Code needs to be built to give you lots of feedback

title says it all. anybody got a fix? trying their VSCode extension and it's pretty different - doesn't run in python environment or folder+claude.md. need to learn how to use it, don't really want to - plenty of visualization problems of its own

In the wet lab I can get about one experiment done per day. that’s relatively fast. It’s a lot of work. most experiments yield limited data and actionable insight. right now I’m doing like three training runs per day with rich metrics and analysis. Completely alien. The dopamine is hitting

at least half the progress in deep learning is surely the result of the addictiveness of building models. line go up is just as strong a reward signal for the models as it is for the humans making them

👀 👀

if you want to make a positive contribution to politics and discourse, hard to think of a less effective means than making posts on social media, especially a platform like Musk‘s. Like stopping climate change by turning off your lights at night

Just shipped runwise 🚀 token-efficient ML training run analysis for AI agents Claude says: "The tool is already quite useful - runwise live, compare, and local saved me a lot of time versus manually parsing logs or navigating W&B." pip install runwise 🧵

just learned that 'J' represents 'I/L' so i can finally spell my name in protein space! the ambiguity is a bit... androgynous? like i could only be represented by splice isoforms or a mass spec that can't tell the difference or something

been using Claude Code for development and Gemini-3 chatbot for discussion. really impressed with Gemini's creativity, ability to diagnose problems, and clear explanations. it's really good. wish i had access to it in the gemini-CLI. having used the CLI with gemini-2.5, i think CC is a better agent

can't unsee this. a true info hazard in the wild

biggest myth about SV and VC culture is that it’s meritocratic. People are constantly rewarded for failure. Knowing the right people, having the right background, and being over the top audacious are way more important than having the best ideas and results