Daniel Saltzberg
3.5K posts
Daniel Saltzberg
@dargason
Data science, compchem, and structural biology. I turn numbers into other numbers to make digital arts and crafts.
Durham, NC Katılım Eylül 2009
1.8K Takip Edilen528 Takipçiler
@anshulkundaje @Ronalfa @gama_search In a system of a million-ish parameters, what's a few more, right?
English
@Ronalfa @gama_search A famous PI building scFMs once told me very confidently that their models can automatically correct technical artifacts without even including any covariates in the model & with no constraints on the expt design 😂😂😂
English
One of the core fundamental research threads we've been pursuing over the last few months at @GoodfireAI is finally out: tightly linking representation geometry and behavior! Hit us up if this spikes your interest!
Goodfire@GoodfireAI
Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵
English
You can now use
`hermes profile create <name> --no-skills`
to create a new agent with no built in skills whatsoever, start with a blank slate, fresh canvas!
hiutiu@hiut1u
@Teknium @chriskim_dev @NousResearch Using them since i started in March, one request: i would like to be able to create a profile without bundled in Hermes skills. My use case: root profile for orchestration, another profiles for specific tasks like contributions. In the end, I have to delete skills frequently.
English
Full write-up: methods that worked, methods that didn't, and what I'd do differently next time. Solo build at Minimal Bio LLC on one GPU.
open.substack.com/pub/mikeminson…
English
9 of 322 designs bound in the GEM × Adaptyv RBX1 Binder Design Competition (ICLR 2026). 8 of those 9 engaged the disordered N-terminus, not the structured RING domain where most groups appeared to be aiming. The IDR was the productive epitope. The RING was where the field was looking.
English
@j0hnparkhill Without a doubt. And for those who will scream "agents expensive!", the solution is the same as it was before: figuring out the standard workflows you can precook and smart engineering. Just more at the agent instead of the code level.
English
@dargason Yeah. I mean. ofc doing all that stuff. But under realistic scenarios where the yaml involves hitting 5 tables and adds value for the users who cant write SQL, unless you have basically precooked all these results, that's 100k's of tokens.
English
@design_proteins @DdelAlamo Also, I find great satisfaction in the idea of designing proteins specifically to annoy squirrels.
English
@design_proteins @DdelAlamo The squirrels in my yard laugh at adding chili flake to birdseed. You're making me wonder now if they have a TRPV mutation - now just how to get a sample to sequence...
English
Do spicy proteins exist?
Can we design and engineer our own?
Here's what we know and what the possibilities are: 🧵
English
@DdelAlamo Perfect, we can engineer a squirrel specific agonist
English
Two dead caps in the same really weird size back to back is bad luck.
English
Daniel Saltzberg retweetledi
@danshipper 3) one "agent engineer" who helps users/teams build their custom agents.
English
@biocheMichael Nice! No need for fancy tools.
Jealous of your lab being in your garage. Have fun setting that up!
English
@dargason Just a spectramax. Nothing crazy but it’ll do
English
Incorporated it
HPLC in my garage rn
Plate reader on the way
Novel Design Build Test Loop
New platform technology incoming
Four stepwise increases in complexity
Great companies start in suburb garages
Stay tuned folks - I’m coming in hot and fast
English
@jrkelly Love and miss Boston. Walking/T-ing everywhere is so freeing.
Cost of living keeps us from moving back, though.
English
Come to boston! Most family friendly urban city in US 😄
Jonathan Berk@berkie1
Active corners in residential neighborhoods are wildly underrated. 📍South End, Boston
English
@Michaelzsguo Nice! Thanks for sharing your workflow. What are you training it to do?
Also, what is the reasoning for dense model being better for predictability than MoE?
English
Made another crucial update. while reviewing the training plan, we dropped the 26B MoE as base model in favor of 31B dense model. this will give us more reliability and predictability.
English
use case #3 for Gemma 4: get ready to fine tuning the model. wish me luck. #AllInGemma4Weekend @googlegemma
English
@jlechh @marcrandolph Yep on cost of living and quality of life. Raising a family in Durham; we left SF because the math on having multiple kids just didn't work.
English
@marcrandolph doing this in durham, nc after living in sf and ny. It has pros and cons that are worth considering (cost, talent density, investor access, etc). It’s an underrated strategy that more folks should be doing, although it's highly dependent on the individual's context and goals
English
If you want to succeed with an early-stage startup, your best move might be far away from Silicon Valley, to a small city with an active startup community of its own.
English