Gal Sapir

15 May

I'm on @science.org ! (eLetter, but still)

the "lack of speaker" being the strongest signal imo. the same way my eyes just skip over 75% of codex/cc output is this feeling, the same feeling you get when you interact with a customer service bot. the sense you are reading a person's thoughts is important in reading (signal of actual 'effort'l

Ted Underwood@tedunderwood.com

Modelspeak may be a little worse than the typical cliché, in fact, because it makes the reader wonder whether there is any speaker — and that profoundly undermines communication.

terrytao.wordpress.comWhy are we even doing all this?[This is a guest post by Rikab Gambhir. This blog post was initially written in a different file format and converted using AI. — T.] First, they came for the artists, and I did not speak out…

An interesting read! I love reading stuff on Terence's site even if it's a guest post. terrytao.wordpress.com/2026/10/15/w...

so natural yet crazy to think that transplanted organs match their "age" (basically methylation patterns++) to recepient's

www.biorxiv.org/content/10.6...

biorxiv.orgbiorxiv.org

ic.unicamp.bric.unicamp.br

engineering is science for impatient people is great! i think also these days CS is science for impatient scientists as well

Tom Silver@tomssilver.bsky.social

This week's #PaperILike is "Artificial Intelligence: An Empirical Science" (Herbert Simon, AIJ 1995). Timely as new results make us revisit foundational questions. And hot takes: engineering is "science for people who are impatient." PDF: ic.unicamp.br/~wainer/cursos/2s2006/epistemico/simon-ia.pdf

yeah I dislike this phrasing too. In the "all models are wrong, some are useful" universe, I think it leans less towards "useful" when applied like that

Vladimir Salnikov@v4ldelund.bsky.social

Sorry guys, but "System 1 Decision model" is something on claudish Can we just call it classifier ?

domain-specific harnesses are all the rage (or so it sometimes seems) wrote a bit about my own experience working on one tldr - i think its a pretty effective way to try and 'encode' expert knowledge and save time+cost when working on projects in specific areas models aren't trained on

sparsethought.comhow do you design the right domain-specific harness?what belongs in a research harness when the models keep improving, and how do we get the expertise into it?

sparsethought.com/2026/09/23/h... might be of interest to @eugenevinitsky.bsky.social

youtube.comJev's kinda neat. It might change how you do some stuff.YouTube video by ed3d dot net

19 Sept

this is a *great* explanation! thanks @ed3d.net

Ed@ed3d.net

Everyone on this blue sky wants to know: what the h*ck is a Jev? Why are nerds excited about this? Reader, I wish I knew. Watcher, I’ll tell you, don’t tell the reader guys. youtube.com/watch?v=HojO... (As usual, obeisance to the algorithm is appreciated. 🫡)

16 Sept

this is the best exchange I've read this whole week!

Vladimir Salnikov@v4ldelund.bsky.social

"just look at the data" final boss

16 Sept

can i just say that Google calling for clinical trials wasn't on my bingo card in 2026? an interesting and imo smart take (still need to write a bit more about it)

"Trust is not a fixed property of a model and cannot be measured as a benchmark. It is earned incrementally through direct human experience with the AI agent and rigorous scientific analysis of the study results. Only prospective studies can observe this dynamic."

16 Sept

can i just say that Google calling for clinical trials wasn't on my bingo card in 2026? an interesting and imo smart take (still need to write a bit more about it)

"PCPs reported that AMIE shifted consultations “from data gathering to data verification”; patients arrived organized with coherent narratives, enabling more collaborative conversations and shared decision-making." i think this is really a first step

14 Sept

a really nice paper they articulate an important (emerging) gap really well - the gap between 'human comprehension bar' and 'agent execution bar' another important aspect they touch on is how to encode the human judgement that goes into the scientific process? super importance stuff

arxiv.org/abs/2604.24658 @elenal3ai.bsky.social @chenyuyou.bsky.social

14 Sept

arxiv.orgThe Last Human-Written Paper: Agent-Native Research ArtifactsScientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural...

11 Sept

recommended reading!

Andy Matuschak@andymatuschak.org

11 Sept

favorite quote so far: "We call our alternative vision reading for transformation: reading augmentation that treats the reader’s transformation as its central design concern." also

10 Sept

2026- one model release at a time www.nber.org/papers/w21788

9 Sept

sharing my latest work: what can a deeply phenotyped human cohort tell us, and what can today’s AI models do with that information? we built PhenoBench: 90 tasks across 15 clinical domains, using clinical, imaging, molecular, and wearable data from the Human Phenotype Project.

arxiv.orgPhenoBench: Mapping What a Deeply Phenotyped Human Cohort Can Tell UsDeeply phenotyped cohorts combine clinical, imaging, molecular, and wearable observations across timescales from seconds to years. This breadth can reveal which measurements inform which health-relate...

9 Sept

might be of interest to @lampinen.bsky.social @adamgayoso.bsky.social and others studying model capabilities and biological ML. the LLM prompts, schemas, and synthetic examples are public too.

tabular foundation models showed small gains over ridge. the 14 LLMs showed a jagged frontier across tasks, generally behind models trained on the same inputs. paper: arxiv.org/abs/2609.06080 more on what we found: sparsethought.com/phenobench/

sparsethought.comllms, the jagged frontier, and healtha look at health benchmarks in the gpt-6 astra and fable 5.1 model cards, incremental gains, and questions about the judges.

8 Sept

i spent some time with gpt-6 and fable 5.1, then looked at their model cards to see what improved in health. incremental gains, and some questions about the judges wrote about it here - sparsethought.com/2026/09/08/l...

galsapir.comGal Sapir · Research CVResearch CV of Gal Sapir, Staff Research Scientist at Pheno.AI, working on health AI, foundation models, clinical evaluation, and research systems.

7 Sept

can i ask experienced researchers here to review a cv please? looking for my next adventure (god i hate linkedin talk), more details soon :-) galsapir.com/cv/

8 Sept

Thanks! It's a great feeling to ask for help from people and get back human, thoughtful and patient answers :-)

galsapir.comGal Sapir · Research CVResearch CV of Gal Sapir, Staff Research Scientist at Pheno.AI, working on health AI, foundation models, clinical evaluation, and research systems.

7 Sept

can i ask experienced researchers here to review a cv please? looking for my next adventure (god i hate linkedin talk), more details soon :-) galsapir.com/cv/

@mariaa.bsky.social @mariozechner.at @tedunderwood.com

7 Sept

6 Sept

that's a good take :-)

Ryan Moulton@moultano.bsky.social

There are a few times when I've gone viral for work or a tweet that I was proud of, and it was pretty great. We should try to do that for more people, instead of the reverse.

the most accurate take

Vicki@vickiboykis.com

It’s pretty wild how cleanly the labs have been able to spin basic security incompetence into positive marketing

you know how they (Plutarch) say "the mind is not a vessel to be filled, but a fire to be kindled"? i feel like it describes working with modern llm/harness combos very well

first very preliminary impressions from gpt 6 astra - being very similar to 5.6 on scientific/biology texts ml-ish code kinda like the benchmarks suggested

I wonder if classical swe stuff really are significantly better

3 Sept

btw notice the difference, to get to science and health you have to scroll wayyy down (and the improvements are really incremental it seems). Can't wait to test it myself though

Gal Sapir@sapir.bsky.social

I honestly think it's crazy (in a good way) that "scientific research" is front and center for a model release in a frontier lab

2 Sept

I honestly think it's crazy (in a good way) that "scientific research" is front and center for a model release in a frontier lab

2 Sept

(also, there is so much headroom in that area still)

sparsethought.comno agents after 8pmon agents making it easier to keep working after my judgment starts to fade, and a simple rule for stopping.

19 Aug

something slightly different (i usually don't to personal / wlb/ touch grass / future of work type stuff) - but this is different no agents after 8 pm sparsethought.com/2026/08/19/n...

Gal Sapir reposted

mr. TIM

@timkellogg.me

19 Aug

yes. yessss. moarrrr arxiv.org/pdf/2608.09703 huggingface.co/nthngdy/matr...

12 Aug

i agree and not sure how to handle it

Maria Antoniak@mariaa.bsky.social

Every part of this resonates. I've been trying to articulate something similar about "the ground shifting underneath language" or "language being destabilized." And this feels especially uncanny and dizzying when I suspect AI use in personal messages from people I know very well. IDK where it leads.

newsletter.semianalysis.comRL Environments and RL for Science: Data Foundries and Multi-Agent ArchitecturesWorker Automation, RL as a Service, Anthropic's next big bet, GDPval and Utility Evals, Computer Use Agents, LLMs in Biology, Mid-Training, Lab Procurement Patterns, Platform Politics and Access

10 Aug

I love their reports. maybe someday I'll subscribe newsletter.semianalysis.com/p/rl-environ...

seangoedecke.comLLMs reward expertise

4 Aug

www.seangoedecke.com/llms-reward-... that's a origiod take (even if well known, it's phrased nicely)

shelbyann.substack.comThe Sirens of Data SalesWhy selling data to train models is not as easy a decision as you think for bio companies

4 Aug

a refreshing read! shelbyann.substack.com/p/the-sirens-of-data-sales (wrote some thoughts about it as well) - sparsethought.com/2026/08/04/t...

26 July

more and more I feel like the days resulting in my best work, are the days in which I hold back on "letting agents roam free" and being much more careful and well, full of care

lucumr.pocoo.orgThe Tower Keeps RisingVibecoding and the possible collapse of a shared language.

24 July

"large software projects have never been limited only by how quickly an individual can produce code. They are limited by how well people can coordinate their understanding of the system they are changing." lucumr.pocoo.org/2026/7/13/th...

22 July

always slightly funny to me when they underestimate themselves these agents

yang-ai-lab.github.ioHEARTS — Health Reasoning over Time SeriesA unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series.

recommended reading: "HEARTS: benchmarking LLM reasoning on health time series" some non-intuitive findings, some intuitive ones i'm glad to see thoroughly checked yang-ai-lab.github.io/HEARTS/#home

two quotes i liked: "benchmarks that include health often sacrifice realism for breadth." and across imputation & forecasting, LLM strategies are "remarkably simplistic, predominantly defaulting to copy-pasting with noise injection" (!!)

(couldn't find the actual people from the preprint here on bsky, but well done)

yang-ai-lab.github.ioHEARTS — Health Reasoning over Time SeriesA unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series.

recommended reading: "HEARTS: benchmarking LLM reasoning on health time series" some non-intuitive findings, some intuitive ones i'm glad to see thoroughly checked yang-ai-lab.github.io/HEARTS/#home

comparing LLMs to SOTA ML on these tasks is lacking and important. that's the "real" baseline: cheap, (mostly) explainable, well accepted acknowledging health evals haven't kept pace with models is, well, also an important statement