Gabe 🧬

damn curriculum learning works

you’re building LLM wrappers, i’m skipping rocks in the loss curve. we are not the same

Science Twitter 🚨 I’m working on a talk about applying to PhD Programs during the next cycle. I was wondering, do you know of any potential changes due to cuts in federal funding? Is there anything that applicants should be doing differently for this cycle? Thank you! 🤍

anyone deep enough in a field will experience this. I’ve gotten ghost @redwood_ai github repos and internal research results/ideas from other groups just to see it be given in a talk. I wonder if this would, for securities, be counted as insider info?

Happy to share that I'm continuing my research on bridge recombinases as a scientist in @pdhsu lab at the @arcinstitute! I am looking for a motivated post-grad to work with me as a research associate. If you are interested or know of someone, please use the link below:

Excited to announce that I’m starting a new position as a senior fellow @FutureHouseSF. I’ll be working on applying AI agents to metagenomic workflows. Very grateful for the opportunity to learn and explore my scientific interests. Please reach out if you’d like to collaborate!

(1/7) New research: how can we understand how an AI model actually works? Our method, SPD, decomposes the *parameters* of neural networks, rather than their activations - akin to understanding a program by reverse-engineering the source code vs. inspecting runtime behavior.

>train a new batch >play some chess >notice loss plateaus in middle of game >cancel job, start with new params >keep playing chess >blunder >loss doesn’t plateau now

grandma’s genes + daily 30 SPF + BHA + Vit C

New Anthropic research: We elicit capabilities from pretrained models using no external supervision, often competitive or better than using human supervision. Using this approach, we are able to train a Claude 3.5-based assistant that beats its human-supervised counterpart.

RL with KL penalties is better seen as Bayesian inference

openai/chatgpt team: plz add a non-temporary no universal memory chat. even when you prompt it to use no universal memory there is still leakage.

hot ai research boy summer

I have a 3-year-old relative with a genetic disorder that I really think could be treated with intrathecal ASOs. How much money would it cost to develop the treatment? Speed probably matters.

hi if you're doing academic research on open source models that requires a significant amount of batch inference (billions to trillions of tokens) we at sfcompute would like to support you, please DM

A personal update: I have officially left my role as co-founder and head of bio @spermracing. It's been a wild ride. Leaving was a tough decision, but I have chosen to move on to something new. open to various roles in the ai x bio space! DMs are open

last time the losing side was the one that expelled minority scientists 🤔

The Case for Bridge Editors Many people seem confused about the gene-editing tool released by @arcinstitute. How does it work and why does it matter? Fair questions. The preprint is not easy to understand, and the reasons for inventing a new gene-editing tool are not obvious...…

me when i stabbed myself with GFP AAV9 🤭