Sign in

Artur Szałata

@chatgtp.bsky.social
2K followers 4.7K following 56 posts

Research Scientist (ML) @ Isomorphic Labs | prev. @ Fabian Theis lab, intern @genentech.bsky.social & @novartis.bsky.social, MSc @icepfl.bsky.social. Opinions stated here are my own.

PostsRepliesMedia
Artur Szałata @chatgtp.bsky.social · 07/07/2026
I'm excited to share that a month ago, I joined Isomorphic Labs! It's hard to imagine a more pressing and motivating mission than reimagining drug discovery to solve all disease. I feel incredibly privileged to embark on it with such a stellar team!
070
Artur Szałata @chatgtp.bsky.social · 24/06/2026
Using Starlink on a plane for the first time and the difference in unreal
030
Artur Szałata @chatgtp.bsky.social · 27/04/2026
tried with chatgpt, claude and gemini, and only gemini picks red but with a disclaimer
040
Artur Szałata @chatgtp.bsky.social · 21/04/2026
I suspect clankers prefer AI writing style overall. Would be great if there were a recent or ongoing comparison of human vs AI reviewers of human vs AI writing
000
Artur Szałata @chatgtp.bsky.social · 13/04/2026
I really wish there was one for tech-bio. If you don't want to run one yourself, would you mind sharing more details to help others do it? 🙏
120
Artur Szałata @chatgtp.bsky.social · 10/04/2026
Ladies in the background don't have ducks 😢
130
Reposted by Artur Szałata
Mark Riedl @markriedl.bsky.social · 31/03/2026
Oh wow, Anthropic accidentally leaked Claude Code and it’s been cloned and made public github.com/instructkr/c...
github.com
GitHub - instructkr/claude-code: Claude Code Snapshot for Research. All original source code is the property of Anthropic.
Claude Code Snapshot for Research. All original source code is the property of Anthropic. - instructkr/claude-code
412332
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 28/03/2026
I love this. And right now atproto is the only place you could really scale up these changes
2402
Artur Szałata @chatgtp.bsky.social · 26/03/2026
Warsaw University Library. Has an awesome rooftop garden that's perfect for a walk while taking a break. Good napping spots and food options nearby - what else can a student dream of?
000
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 15/03/2026
For a recent lab meeting, I wrote up a grab bag of ways to think about your development as a researcher during a PhD: emerge-lab.github.io/papers/an-un... Sharing in case folks find it useful or have feedback!
emerge-lab.github.io
610112
Artur Szałata @chatgtp.bsky.social · 03/03/2026
Great collaboration with Tianyu Liu, Wenxin Long, @fabiantheis.bsky.social , Lingzhou Xue, and Hongyu Zhao
020
Artur Szałata @chatgtp.bsky.social · 03/03/2026
We embed chemical and genetic perturbation readouts and use the model for gene-drug interaction prediction, among other tasks.
100
Artur Szałata @chatgtp.bsky.social · 03/03/2026
Paper alert 🚨 @ MLGenX @iclr-conf.bsky.social 2026! PertOmni - CLIP-style multimodal representation-learning framework for contrastive alignment of perturbation readouts and textual embeddings.
openreview.net
Learning Perturbation Effects Through Contrastive Alignment of...
Single-cell perturbation screens offer a scalable approach for characterizing the effects of genetic and chemical interventions on cellular state. However, most existing representation-learning...
151
Artur Szałata @chatgtp.bsky.social · 03/03/2026
On top of the paper, I had the pleasure of working with the incredible team at BRAID: Jan-Christian Hütter, Vladimir Ermakov, Zoe Piran, and Russell Littman on the Virtual Cell Challenge virtualcellchallenge.org/leaderboard, where our team ranked 13 on the most significant metric, PDS, and 32 overal
lnkd.in
LinkedIn
This link will take you to a page that’s not on LinkedIn
000
Artur Szałata @chatgtp.bsky.social · 03/03/2026
This work was done during my internship at BRAID @genentech.bsky.social , together with Alexander Wu, Jan-Christian Hütter, Zoe Piran, Russell Littman, David Richmond, and @fabiantheis.bsky.social
110
Artur Szałata @chatgtp.bsky.social · 03/03/2026
PerturBERT, instead, is a transformer that learns how genes co-vary under perturbation by modeling perturbation signatures. Learned gene embeddings match or outperform scGPT and other baselines despite PerturBERT training on a 30x smaller dataset and using 65x fewer genes.
100
Artur Szałata @chatgtp.bsky.social · 03/03/2026
Paper alert 🚨 @ MLGenX @iclr-conf.bsky.social 2026! PerturBERT openreview.net/forum?id=ZsD... - Most transformers in single-cell omics (e.g. scGPT) learn gene co-expression through pre-training on masked gene expression levels.
openreview.net
PerturBERT: Learning Gene Co-Variation Embeddings from Perturbation...
Current foundation models for transcriptomic data are typically trained in a self-supervised manner to predict gene expression within a sample given other genes, thereby learning gene co-variation...
120
Artur Szałata @chatgtp.bsky.social · 02/03/2026
Caris is launching a WGS+ML blood test for early cancer detection. Impressive sensitivity: 56% stage 1 and 70% stage 2 - if it holds, it would blow competitors out of the water
carislifesciences.com
Caris Completes Interim Readout of Achieve 1 Study Demonstrates Superior Sensitivity and Specificity of Caris Detect
Caris Life Sciences Completes Interim Readout of Achieve 1 Study Demonstrates Superior Sensitivity and Specificity of Caris Detect
000
Artur Szałata @chatgtp.bsky.social · 27/02/2026
Nano Banana 2 test. Humans are still safe 😌
020
Artur Szałata @chatgtp.bsky.social · 25/02/2026
Inspired by
000
Artur Szałata @chatgtp.bsky.social · 25/02/2026
"Constitution" is a bad name for model preference guidelines doc. It suggests, among others, public legitimacy, where there is none.
120
Artur Szałata @chatgtp.bsky.social · 19/02/2026
Surprisingly, all of the models give me the right answer
010
Artur Szałata @chatgtp.bsky.social · 19/02/2026
Interesting how opus 4.6 extended and gemini 3 pro give me the correct answer here, but not ChatGPT5.2 thinking that I'd trust more with most of my work
010
Reposted by Artur Szałata
Chris Paxton @cpaxton.bsky.social · 07/02/2026
How to fake a robotics result: a short blog post listing many sins which annoy me (many of which I am guilty of from time to time, to be fair) open.substack.com/pub/itcanthi...
1285
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 07/02/2026
Haha @cpaxton.bsky.social is on fire: open.substack.com/pub/itcanthi... Some useful tips in there even for non-roboticists looking to make their paper artificially look good
open.substack.com
How to Fake A Robotics Result
Since that deadline is coming up
2293
Artur Szałata @chatgtp.bsky.social · 02/02/2026
e-ink is a lifesaver for anyone with cybersickness
000
Reposted by Artur Szałata
Ted Underwood @tedunderwood.com · 19/01/2026
And for a lot of Gen X journalists and academics, the answer to (a) — assuming existing skills and plans — is legit "no." AI can be useful, for sure, but the paths it differentially advantages are not the paths where they have accumulated momentum, expertise, and social capital. +
1142
Artur Szałata @chatgtp.bsky.social · 15/01/2026
Big thanks to Olga Novitskaia for spotting the issue! Link to the broken history mapper: useast.ensembl.org/Homo_sapiens...
useast.ensembl.org
000
Artur Szałata @chatgtp.bsky.social · 15/01/2026
A quick story on how we matched genes across two datasets with different Ensembl versions. 1. There must be a tool out there. Ensembl ID History converter ofc! 2. Doesn't match Ensembl search outcomes due to a bug 3. Lesson: use this client instead github.com/Ensembl/ense... !
github.com
Mapping of ENSG_IDs between different release of the Ensembl database · Issue #744 · Ensembl/ensembl
Dear members of the Ensembl team, I wasn’t sure who to contact, so I’m starting here. I am writing to ask you questions about the IDMapper tool presented on your website: https://www.ensembl.org/Ho...
141
Artur Szałata @chatgtp.bsky.social · 13/01/2026
Also awesome for finding posters at a conference!
010
Reposted by Artur Szałata
Arc Institute @arcinstitute.org · 09/01/2026
Predicting cell state in previously unseen conditions has typically required retraining for each new biological context. Today, Arc is releasing Stack, a foundation model that learns to simulate cell state under novel conditions directly at inference time, no fine-tuning required.
1116
Artur Szałata @chatgtp.bsky.social · 12/01/2026
This. You can add a link in a reply to the post and avoid the penalty
000
Artur Szałata @chatgtp.bsky.social · 12/01/2026
screenshot with a link alongside? Could draw those who want to participate in discussion where it started
110
Reposted by Artur Szałata
Sakana AI @sakanaai.bsky.social · 12/01/2026
Introducing DroPE: Extending Context by Dropping Positional Embeddings We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity. arxiv.org/abs/2512.12167 pub.sakana.ai/DroPE
211821
Reposted by Artur Szałata
Ethan Mollick @emollick.bsky.social · 11/01/2026
One very familiar pattern in AI and science right now is going from a lot of false starts on hard tasks (there have been near-misses where AI appears to solve an Erdos problem but just finds an old solution no one knew about) to actually doing the thing soon after. Three Erdos problems in 3 days.
310812
Reposted by Artur Szałata
Andrew Gordon Wilson @andrewgwils.bsky.social · 07/01/2026
We introduce epiplexity, a new measure of information that provides a foundation for how to select, generate, or transform data for learning systems. We have been working on this for almost 2 years, and I cannot contain my excitement! arxiv.org/abs/2601.03220 1/7
814234
Reposted by Artur Szałata
Ryan Moulton @moultano.bsky.social · 30/12/2025
Finished the essay. moultano.wordpress.com/2025/12/30/c...
moultano.wordpress.com
Children and Helical Time
In subjective time, childhood is half of life. Life, then, is the creation of childhoods. You have yours, and then you get to create them for others.
1714428
Reposted by Artur Szałata
Chris Paxton @cpaxton.bsky.social · 24/12/2025
Autonomous RIVR delivery robots in Pittsburgh
2019844
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 14/12/2025
A quick sunday rewrite of an old blog post about how one should evaluate the effectiveness of an empirical paper: open.substack.com/pub/emergere...
open.substack.com
How much should I trust this paper?
Or how to build your castle on sand
1183
Reposted by Artur Szałata
Nathan Lambert @natolambert.bsky.social · 06/12/2025
Good researchers obsess over evals The story of Olmo 3 (post-training), told through evals NeurIPS Talk tomorrow. Upper Level Room 2, 10:35AM. Slides: docs.google.com/presentation...
1306
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 17/11/2025
Elon’s power is that he offers a positive vision of the future. This attracts employees, funding, support. There’s a massive techno positive hole and he fills it.
7585
Artur Szałata @chatgtp.bsky.social · 23/10/2025
here's the press release too
cellarity.com
Cellarity Publishes Framework for Discovery of Cell State-Correcting Medicines in Science
The Cellarity platform provides critical insights into cellular dysfunction and reveals new biology that we can use to treat disease.
010
Artur Szałata @chatgtp.bsky.social · 23/10/2025
Active learning with DrugReflector beats SotA in phenotypic hit-rate for virtual screening. Includes a sc perturbation dataset with 10 lines and 104 compounds. Out in @science.org now! Grateful to Cellarity and @fabiantheis.bsky.social for the opportunity to contribute to this outstanding project!
science.org
Active learning framework leveraging transcriptomics identifies modulators of disease phenotypes
Phenotypic drug screening remains constrained by the vastness of chemical space and technical challenges scaling experimental workflows. To overcome these barriers, computational methods have been dev...
1112
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 23/10/2025
What if we did a single run and declared victory
Three panel thing. In the left panel we use error bars. In the second, we take statistical significance as the biggest number but still have error bars. In LLM science, we just have the biggest number
1333570
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 13/10/2025
Community notes when
2415
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 11/10/2025
Yeah this is my biggest “AGI hype is not real” is that almost no one at these companies behaves like it’s real
0152
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 28/09/2025
My skepticism of LLM-as-scientist stems from how imbalanced the literature is. Median paper is mildly negative result presented as positive, it's unclear how to RLHF on good hypothesis vs. bad hypothesis, etc. We barely know how to teach this skill, how can we RLHF it
812210
Reposted by Artur Szałata
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 25/09/2025
For folks considering grad school in ML, my advice is to explore programs that mix ML with a domain interest. ML programs are wildly oversubscribed while a lot of the fun right now is in figuring out what you can do with it
815217