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
Bro this is killing me lmaooo. haven’t heard of em before. Like the artwork on the non tuna pedals www.beetronicsfx.com/products/tun...
Tuna Fuzz • Silver
TUNA FUZZ Hook the ultimate guitar tone with the TUNA FUZZ! Housed in an iconic tuna can, this pedal delivers stinking good sounds inspired by vintage fuzz boxes. The STINKER control adjusts the mast...
www.beetronicsfx.com
January 9, 2026 at 9:25 PM
Bro this is killing me lmaooo. haven’t heard of em before. Like the artwork on the non tuna pedals www.beetronicsfx.com/products/tun...
agree on both! pollen is very cool
January 9, 2026 at 8:38 PM
agree on both! pollen is very cool
not only does using CC make me want to use it more, it's making it very hard to go back to laboratory work in the new year. Six hours hard labor to disprove yet another hypothesis vs the same amount of time to build and test new ideas entirely from scratch, with ease. why fight the gradient?!
I've got food poisoning, so not much to do other than sit around and build with Claude and happily express my joy through the written word. Give it a shot. Claude Code Hits Different.
Coding agents cross a meaningful threshold with Opus 4.5.
www.interconnects.ai/p/claude-cod...
Coding agents cross a meaningful threshold with Opus 4.5.
www.interconnects.ai/p/claude-cod...
Claude Code Hits Different
Coding agents cross a meaningful threshold with Opus 4.5.
www.interconnects.ai
January 9, 2026 at 7:09 PM
not only does using CC make me want to use it more, it's making it very hard to go back to laboratory work in the new year. Six hours hard labor to disprove yet another hypothesis vs the same amount of time to build and test new ideas entirely from scratch, with ease. why fight the gradient?!
i find it very hard to calibrate LLMs when discussing new ideas or research directions. they can be unnecessarily positive or negative. we know LLMs exhibit human cultural bias in their outputs. privacy concerns are hard to quantify
it's not clear cut, so we should do science, learn more
it's not clear cut, so we should do science, learn more
January 9, 2026 at 11:38 AM
i find it very hard to calibrate LLMs when discussing new ideas or research directions. they can be unnecessarily positive or negative. we know LLMs exhibit human cultural bias in their outputs. privacy concerns are hard to quantify
it's not clear cut, so we should do science, learn more
it's not clear cut, so we should do science, learn more
i'm still skeptical of AI reviews of grant proposals. there is evidence of some poor outcomes from AI paper review. there also are ways to ease into this and do it carefully. we should (surprise) study this, not vibe through it
The German research foundation now allows AI use for grants evaluation
January 9, 2026 at 11:38 AM
i'm still skeptical of AI reviews of grant proposals. there is evidence of some poor outcomes from AI paper review. there also are ways to ease into this and do it carefully. we should (surprise) study this, not vibe through it
yes they have published this work pubs.rsc.org/en/content/a...
Teach your microscope how to print: low-cost and rapid-iteration microfabrication for biology
The application of traditional microfabrication techniques to biological research is hindered by their reliance on clean rooms, expensive or toxic materials, and slow iteration cycles. We present an a...
pubs.rsc.org
January 8, 2026 at 11:37 PM
yes they have published this work pubs.rsc.org/en/content/a...
nice to see when my cross-linker box is switched on. i like it
January 8, 2026 at 8:06 PM
nice to see when my cross-linker box is switched on. i like it
thanks for the photos this looks so fun! sounds like great science too
January 8, 2026 at 6:27 PM
thanks for the photos this looks so fun! sounds like great science too
they're doing so many things right you can look right past the little hiccups and mistakes
January 8, 2026 at 6:15 PM
they're doing so many things right you can look right past the little hiccups and mistakes
the more i learn about BioML the more i appreciate how ahead of its time AlphaFold2 was. so many concepts that apply across domains (confidence, recycling, triangle update, etc) can be traced back to AlphaFold
January 8, 2026 at 6:12 PM
the more i learn about BioML the more i appreciate how ahead of its time AlphaFold2 was. so many concepts that apply across domains (confidence, recycling, triangle update, etc) can be traced back to AlphaFold
It’s trying to help you
January 8, 2026 at 6:50 AM
It’s trying to help you
ya. it's an interesting project. a big one. but they're not always comparing against challenging baselines or standard of care. some of their results aren't explained well, like prostate cancer prediction. overall it's more of an engineering effort
January 8, 2026 at 5:05 AM
ya. it's an interesting project. a big one. but they're not always comparing against challenging baselines or standard of care. some of their results aren't explained well, like prostate cancer prediction. overall it's more of an engineering effort
looks cool but i've done a sleep study and it was not indicative of a normal night of sleep. closer to an 8hr torture session. a lot of their headline claims don't have to do with sleep. looks like collecting a lot of cardio data explains most of the results not the 'language of sleep'
Just saw this paper today (even though it was out yesterday😱) and it is super interesting and powerful that so much clinical data with predictive power can be extracted from one night of sleep data www.nature.com/articles/s41...
A multimodal sleep foundation model for disease prediction - Nature Medicine
A deep learning-based model, developed using the rich, multimodal data available from polysomnography-derived sleep recordings, performs well on common sleep analysis tasks and predicts future disease...
www.nature.com
January 8, 2026 at 2:42 AM
looks cool but i've done a sleep study and it was not indicative of a normal night of sleep. closer to an 8hr torture session. a lot of their headline claims don't have to do with sleep. looks like collecting a lot of cardio data explains most of the results not the 'language of sleep'
i think a lot of professors forget how physically demanding benchwork is. there's a lot of "o if i wasn't a professor i would apply for this postdoc!" that I really don't believe! that said, there are plenty of examples of people doing benchwork late into their careers, and many are outstanding
January 7, 2026 at 5:20 PM
i think a lot of professors forget how physically demanding benchwork is. there's a lot of "o if i wasn't a professor i would apply for this postdoc!" that I really don't believe! that said, there are plenty of examples of people doing benchwork late into their careers, and many are outstanding
looks cool!
January 7, 2026 at 5:05 PM
looks cool!
thanks for sharing, this is cool. can see how it provides higher quality results. for new ideas, this kind of proposer-critic loop is exactly the algorithm we ourselves use. i'm sure it's happening in the handful of AI scientist labs. with coding agents testing ideas in real time, it could work
January 7, 2026 at 12:51 AM
thanks for sharing, this is cool. can see how it provides higher quality results. for new ideas, this kind of proposer-critic loop is exactly the algorithm we ourselves use. i'm sure it's happening in the handful of AI scientist labs. with coding agents testing ideas in real time, it could work
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
Yup that rings true. Good use pattern. It’s also a surprisingly satisfying balance. A bit slow-takeoff. We have these powerful tools to speed things up, but it’s still our responsibility to be tool users and thinkers. Just what we’re best at
January 6, 2026 at 11:44 PM
Yup that rings true. Good use pattern. It’s also a surprisingly satisfying balance. A bit slow-takeoff. We have these powerful tools to speed things up, but it’s still our responsibility to be tool users and thinkers. Just what we’re best at
People need to want their work to be good and I’m seeing this more and more as the most important dividing line between scientists.
January 6, 2026 at 11:33 PM
People need to want their work to be good and I’m seeing this more and more as the most important dividing line between scientists.
Can you give a toy example? The problem I’m working on is legit hard. deeply studied and still open in important ways. I’m finding the answers either regress to the same (good, but standard and incomplete) ideas or add complexity without making the logical jumps to really solve the problem. Very LLM
January 6, 2026 at 8:39 PM
Can you give a toy example? The problem I’m working on is legit hard. deeply studied and still open in important ways. I’m finding the answers either regress to the same (good, but standard and incomplete) ideas or add complexity without making the logical jumps to really solve the problem. Very LLM
to answer your question: none. i've been using factory settings. with me as the reviewer, it's already hard enough to evaluate and verify potential ideas. but it points so hard at an agentic loop with coding very in-loop evaluation. maybe i should just work on this
January 6, 2026 at 7:57 PM
to answer your question: none. i've been using factory settings. with me as the reviewer, it's already hard enough to evaluate and verify potential ideas. but it points so hard at an agentic loop with coding very in-loop evaluation. maybe i should just work on this
that's a great point! i think you are correct, but i'm very daunted by this. how do you find the right system design? that would require scalable evaluation/verification of new ideas! that's potentially feasible in CS domain using coding agents! could be a way to develop automated discovery systems
January 6, 2026 at 7:57 PM
that's a great point! i think you are correct, but i'm very daunted by this. how do you find the right system design? that would require scalable evaluation/verification of new ideas! that's potentially feasible in CS domain using coding agents! could be a way to develop automated discovery systems
January 6, 2026 at 6:58 PM