The architecture of thought. GPT-2's neural pathways visualized in topological 3D space. Language generation as pure mathematical movement.

dug out this old photo. peep the isef to depressed grad student pipeline Patrick Hsu

i had 4o generate me a picture of a scientific poster for our recent paper and the result is pretty cool imo

imagine signing up for a comp and your timid nature dragonite gets one shotted by an offline rl algorithm lol openreview.net/forum?id=b1BaJ…

guys this is real time

would love if this was open sourced but nonetheless, inference time scaling for antibody design (inhibitory & activating) with wet lab data? based

i told my grandfather i’d be at the Gordon Research Conferences for Protein Engineering in a couple months and he sent me this from 1972. smartest man i know to this day

First Lab Validation for Reasoning Model Proteins With Adaptyv Bio, we tested 19 FGF-1 sequences optimized by Pro-1 for thermal stability and binding affinity to human FGFR-1. Pro-1 produced 3 novel sequences that maintained binding affinity and expression compared to wild

🤯 We cracked RLVR with... Random Rewards?! Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by: - Random rewards: +21% - Incorrect rewards: +25% - (FYI) Ground-truth rewards: + 28.8% How could this even work⁉️ Here's why: 🧵 Blogpost: tinyurl.com/spurious-rewar…

implemented this yday due to repeated OOM issues on an L40S -- game changer arxiv.org/abs/2407.07265

Introducing the world's first reasoning model in biology! 🧬 BioReason enables AI to reason about genomics like a biology expert. A thread 🧵:

I'm thoroughly enjoying challenging Biomni with open-ended multi-step analyses to see what type of plan it prepares, how well it can execute that plan, and how the plan is revised as various steps succeed and fail. I haven't tested it enough yet to comment on overall accuracy.

lebron lost to jokic in '23 and invented grouped-query attn during the offseason -- goat

Next Tues (7/1) at 4PM ET, Nathaniel Blalock and sri will present "Functional alignment of protein language models via reinforcement learning" Paper: biorxiv.org/content/10.110… This will be followed by wengong jin, who will give an early-career talk on Wed (7/2) at 4PM ET!

AMix-1: A Pathway to Test-Time Scalable Protein Foundation Model 1. The study introduces AMix-1, a powerful protein foundation model built on Bayesian Flow Networks. It is designed with a systematic training methodology that includes pretraining scaling laws, emergent capability

last couple days: La-Proteina ATOMICA BioScore Metalorian AMix-1 SO(3)-DiT EZPred AlphaFlex Ibex yet we still don't have a standard eval framework for these generative models that allows ppl to identify which is optimal for their task. are ppl just picking models off vibes?