Jase Gehring
@skyjase.bsky.social
scientist at UC Berkeley inventing advanced genomic technologies
lover of molecules, user of computers
https://scholar.google.com/citations?user=63ZRebIAAAAJ&hl=en
lover of molecules, user of computers
https://scholar.google.com/citations?user=63ZRebIAAAAJ&hl=en
looks cool!
January 7, 2026 at 5:05 PM
looks cool!
I think we’ll soon see a “double-check” mode that lets you spend tokens for adversarial verification. this could be combined with “no really, do it all yourself” for more walk-away development
These things really need a better UI to let you use the downtime for learning though.
January 6, 2026 at 11:59 PM
I think we’ll soon see a “double-check” mode that lets you spend tokens for adversarial verification. this could be combined with “no really, do it all yourself” for more walk-away development
January 6, 2026 at 6:58 PM
research-focused LLM agents will likely fille an important niche. there are patterns of behavior in science, easily adopted by an LLM, that are not part of a "chatbot" or "coding agent" interface. comprehensive literature search and literature context management ... i should make a /skill for that
this is tempering my excitement a lot and i'm a bit embarrassed for posting through it. two possibilities:
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
lmao this idea was so good it was published last June 🤡
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
January 6, 2026 at 6:36 PM
research-focused LLM agents will likely fille an important niche. there are patterns of behavior in science, easily adopted by an LLM, that are not part of a "chatbot" or "coding agent" interface. comprehensive literature search and literature context management ... i should make a /skill for that
this is tempering my excitement a lot and i'm a bit embarrassed for posting through it. two possibilities:
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
lmao this idea was so good it was published last June 🤡
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
claude came up with a beautiful idea today. think it's time to tell my boss about this project
January 6, 2026 at 3:59 PM
this is tempering my excitement a lot and i'm a bit embarrassed for posting through it. two possibilities:
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
1. claude is laundering training data as novel ideas. quite bad. bearish
2. it's re-inventing published ideas from scratch. true scientist behavior. happens constantly
my guess: the former 😬
lmao this idea was so good it was published last June 🤡
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
claude came up with a beautiful idea today. think it's time to tell my boss about this project
January 6, 2026 at 1:40 AM
lmao this idea was so good it was published last June 🤡
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
hard to know if a paper from June made it into Opus 4.5's training data or if it was a true original proposal. either way, a lot less exciting.
yesterday was one of the most remarkable in my ~15 years as a scientist. we're on the verge of an explosion in computational science. everything until now has been the lag phase. we're about to enter log phase
January 5, 2026 at 6:54 PM
yesterday was one of the most remarkable in my ~15 years as a scientist. we're on the verge of an explosion in computational science. everything until now has been the lag phase. we're about to enter log phase
Attempting to melt the lab server through the floor
January 5, 2026 at 3:24 PM
Attempting to melt the lab server through the floor
claude came up with a beautiful idea today. think it's time to tell my boss about this project
January 5, 2026 at 4:21 AM
claude came up with a beautiful idea today. think it's time to tell my boss about this project
this is cool. taking pLMs, which implicitly learn MSAs, and using their embeddings to explicitly generate MSAs faster and better than traditional methods
Fast, accurate construction of multiple sequence alignments from protein language embeddings www.biorxiv.org/content/10.6... 🧬🖥️🧪 github.com/Singh-Lab/AR...
January 4, 2026 at 6:48 PM
this is cool. taking pLMs, which implicitly learn MSAs, and using their embeddings to explicitly generate MSAs faster and better than traditional methods
Reposted by Jase Gehring
2025 was a year full of great scientific discoveries and breakthroughs. Here, I want to highlight 10 papers that got me really excited. Of course, this is a highly subjective selection and I am sure I am forgetting some papers that I would love to include, but here it goes: (1/12)
January 4, 2026 at 6:44 AM
2025 was a year full of great scientific discoveries and breakthroughs. Here, I want to highlight 10 papers that got me really excited. Of course, this is a highly subjective selection and I am sure I am forgetting some papers that I would love to include, but here it goes: (1/12)
gonna keep saying this because it's true and important.
the only way out of this mess in the next three years is to make the 2026 midterms an impeachment referendum.
Candidates, i.e. most dems and some reps, need to explicitly run on this message and get elected with a clear mandate.
the only way out of this mess in the next three years is to make the 2026 midterms an impeachment referendum.
Candidates, i.e. most dems and some reps, need to explicitly run on this message and get elected with a clear mandate.
January 4, 2026 at 4:16 PM
gonna keep saying this because it's true and important.
the only way out of this mess in the next three years is to make the 2026 midterms an impeachment referendum.
Candidates, i.e. most dems and some reps, need to explicitly run on this message and get elected with a clear mandate.
the only way out of this mess in the next three years is to make the 2026 midterms an impeachment referendum.
Candidates, i.e. most dems and some reps, need to explicitly run on this message and get elected with a clear mandate.
Gemini 3 Pro is really smart. i've had some really rewarding conversations on chemistry, math, and deep learning. probably my favorite model for learning. i also use it frequently to review Claude's Code, to propose new ideas, and to critique proposals
January 4, 2026 at 5:29 AM
Gemini 3 Pro is really smart. i've had some really rewarding conversations on chemistry, math, and deep learning. probably my favorite model for learning. i also use it frequently to review Claude's Code, to propose new ideas, and to critique proposals
this is me, and i know how to code! my background skews toward bio and chem. i never coded full time, never got really good. with agentic tools, i'm building software projects >30x faster than i could on my own. not an exaggeration
If this technology reduces barriers to coding significantly enough, it will also change the equation for lots of fields that don’t *think* of themselves right now as bottlenecked by coding
The opportunity cost of most of the world *not* learning about technology is huge. Now I have no silver bullet, magical solution fantasy about this, I worked for too long in learning science for that. But tools that propel you into active learning are really possible here, not just offloading
January 3, 2026 at 10:12 PM
this is me, and i know how to code! my background skews toward bio and chem. i never coded full time, never got really good. with agentic tools, i'm building software projects >30x faster than i could on my own. not an exaggeration
Reposted by Jase Gehring
No. We need to tax the bejeezus out of billionaires and fund our scientific agencies to support scientific innovation the way they should be.
We need a Kickstarter for grants.
January 3, 2026 at 1:31 PM
No. We need to tax the bejeezus out of billionaires and fund our scientific agencies to support scientific innovation the way they should be.
If you think AI is gonna bring in four day work weeks and universal high salaries, you don’t understand America or economics itself. Four day work weeks are the result of high demand for labor, not an abundance of it
While a growing number of U.S. employers are mandating workers return to the office five days a week, some companies say AI is saving them enough time to launch or sustain a four-day workweek.
These companies say AI is key to their four-day workweeks
Some companies are giving workers back more time as artificial intelligence takes over more tasks.
wapo.st
January 3, 2026 at 2:04 PM
If you think AI is gonna bring in four day work weeks and universal high salaries, you don’t understand America or economics itself. Four day work weeks are the result of high demand for labor, not an abundance of it
Yup this is when I knew he was gonna get away with it. We’re gonna paper over international crimes and run cover for … trump
President Trump said on Saturday that the U.S. had captured the Venezuelan president, Nicolás Maduro. Follow live updates. www.nytimes.com/live/2026/01...
January 3, 2026 at 1:37 PM
Yup this is when I knew he was gonna get away with it. We’re gonna paper over international crimes and run cover for … trump
impeachment is not plausible right now, but it is realistically on the ballot this year. the 2026 midterms should be about one issue: the impeachment of the president. this is effectively the only way to impeach a president in our system, by popular mandate in a midterm election
January 3, 2026 at 8:30 AM
impeachment is not plausible right now, but it is realistically on the ballot this year. the 2026 midterms should be about one issue: the impeachment of the president. this is effectively the only way to impeach a president in our system, by popular mandate in a midterm election
what part of the world is not willing to put with imperial war right now?
I think this administration has vastly overestimated the rest of the world's willingness to put up with imperial oil war bullshit right now. I don't think this ends well. No war.
January 3, 2026 at 7:11 AM
what part of the world is not willing to put with imperial war right now?
GPU: ⚡
Line: 📈
Claude: 🤖
Jase: 😴
Line: 📈
Claude: 🤖
Jase: 😴
January 3, 2026 at 12:31 AM
GPU: ⚡
Line: 📈
Claude: 🤖
Jase: 😴
Line: 📈
Claude: 🤖
Jase: 😴
we're converging on an equilibrium where the US generates massive inequality internally and globally via "free market" deregulation and falls back on the EU to regulate the most grotesque excesses of its industries
Best shot is probably EU regulators?
January 2, 2026 at 7:37 PM
we're converging on an equilibrium where the US generates massive inequality internally and globally via "free market" deregulation and falls back on the EU to regulate the most grotesque excesses of its industries
every paper deserves two separate social media posts: normie mode and sickos mode.
January 2, 2026 at 6:03 PM
every paper deserves two separate social media posts: normie mode and sickos mode.
for the less worldly 🥖
January 2, 2026 at 4:43 PM
for the less worldly 🥖
made a Claue /experiment skill that's been great for understanding and exploring complex datasets
i prompt with a question or specific study in mind, and it will:
- propose experiment or analysis
- define methodology, sample size, statistical analysis, success criteria
(1 of 2, cont'd)
i prompt with a question or specific study in mind, and it will:
- propose experiment or analysis
- define methodology, sample size, statistical analysis, success criteria
(1 of 2, cont'd)
January 2, 2026 at 4:12 PM
made a Claue /experiment skill that's been great for understanding and exploring complex datasets
i prompt with a question or specific study in mind, and it will:
- propose experiment or analysis
- define methodology, sample size, statistical analysis, success criteria
(1 of 2, cont'd)
i prompt with a question or specific study in mind, and it will:
- propose experiment or analysis
- define methodology, sample size, statistical analysis, success criteria
(1 of 2, cont'd)
i've re-implemented PoPE from scratch as part of a model overhaul (w/ Claude Code). i don't have clean ablations yet but it's working. took about two hours to re-write my entire model
i think the PoPE idea is really cool and a great fit for scientific data
i think the PoPE idea is really cool and a great fit for scientific data
They find that RoPE (the positional encoding used in most modern LLMs) has a fundamental flaw. It entangles "what" (content) and "where" (position) information.
They propose PoPE (Polar Coordinate Position Embeddings), which eliminates the what-where.
They propose PoPE (Polar Coordinate Position Embeddings), which eliminates the what-where.
January 2, 2026 at 5:19 AM
i've re-implemented PoPE from scratch as part of a model overhaul (w/ Claude Code). i don't have clean ablations yet but it's working. took about two hours to re-write my entire model
i think the PoPE idea is really cool and a great fit for scientific data
i think the PoPE idea is really cool and a great fit for scientific data