iiinova — where your digital self keeps evolving

FYI I wrapped a lot of immuno-informatics predictors so you don't have to: openvax.github.io/mhctools/ (eg NetMHCpan, MHCflurry, Pepsickle, DeepTAP, ERAMER, BigMHC, PRIME, &c)

openvax.github.io

mhctools

Run MHC binding, presentation, immunogenicity and antigen-processing predictors through one predict() call

053

Alex Rubinsteyn @alexr.bsky.social · 18h

Acquired Disney episode such a dread-inducing tour of how capitalism always wins when you’re trying to balance it against artistic integrity and societal myth-formation Also feels like the opening round of whatever AI will do to culture, slop is way more efficient than quality

100

Alex Rubinsteyn @alexr.bsky.social · 29/09/2026

Roadmap: 2.4: dig into the processing predictor’s sensitivity to negative sampling strategy, modernize network architectures, train on 2x more data 3.0: a few deeper research ideas that might be bigger unlocks I’m short on time, let me know if you want to help

000

Alex Rubinsteyn @alexr.bsky.social · 29/09/2026

MHCflurry 2.3 is out Lots of internal improvements to make both training and prediction faster on multi-GPU Also trained a slightly better model (first new weights in a long time) Thanks to NVIDIA for the Brev credits Try it out and tell me what I broke

271

Reposted by Alex Rubinsteyn

Rob Patro @robp.bsky.social · 25/09/2026

Cuttlefish 3.0.3 is out! github.com/COMBINE-lab/... Fun little jaunt with Opus 5.5 to substantially reduce the memory requirement (especially at moderate thread counts). Now it's both non-trivially faster *and* more memory-frugal than the initial C++ implementation!

github.com

Release 3.0.3 · COMBINE-lab/cuttlefish

Release Notes Colored builds use far less memory at low and moderate thread counts. On 149,998 Salmonella assemblies at 16 threads, peak RSS falls from 22.4 GB to 8.9 GB (C++ Cuttlefish 3: 11.5 GB...

1134

Reposted by Alex Rubinsteyn

Wei Shen 沈伟 @shenwei356.bsky.social · 24/09/2026

SeqKit is now faster and uses 40% less memory when reading large FASTA records, such as human genomes, though it still uses more RAM than tools written in C/C++/Rust. Worth upgrading! github.com/shenwei356/s...

github.com

GitHub - shenwei356/seqkit: A cross-platform and ultrafast toolkit for FASTA/Q file manipulation

A cross-platform and ultrafast toolkit for FASTA/Q file manipulation - shenwei356/seqkit

0146

Alex Rubinsteyn @alexr.bsky.social · 16/09/2026

I'm working on an AI eval project that needs domain expert human graders across different corners of biology. Let me know if you want a wacky consulting project, the biological domain likely can be whatever you already do research on. (ideally post-PhD)

231

Reposted by Alex Rubinsteyn

Grace @gracekind.net · 12/09/2026

“swarmchasers”

4953

Alex Rubinsteyn @alexr.bsky.social · 11/09/2026

New coding safety filters on Opus refuse to do biology I guess this might actually be the end of the line for me with Claude?

431

Alex Rubinsteyn @alexr.bsky.social · 09/09/2026

Review/fix cycles with Opus 5 is just an infinite loop of "The most interesting failure was mine" until you give up and switch to a different model

201

Alex Rubinsteyn @alexr.bsky.social · 09/09/2026

Trying to convince Claude we actually want to publish all the docs related to starting a clinical trial but it keeps objecting to releasing things that "competitors, partners, and future investors will read" Alignment is such a weak idea, humans always have competing values

040

Alex Rubinsteyn @alexr.bsky.social · 09/09/2026

Here is the grumpy blog post that Sergey Feldman and I wrote 13 years ago when Mark Zuckerberg invaded NeurIPS to buy up all the researchers: explainmydata.blogspot.com/2013/12/nips... Everything we were worried about happened almost immediately but the field also made prev unimaginable progress

explainmydata.blogspot.com

NIPS and the Zuckerberg Visit

Shared blog about machine learning, data analysis, and statistics.

010

Reposted by Alex Rubinsteyn

Ethan Mollick @emollick.bsky.social · 08/09/2026

This is a VERY big one. (And yes, the fights over academic credit and what happened in the race for the proof needs to be resolved, but it is still appears that this is a big one, if true.) openai.com/index/navier...

openai.com

On the Navier–Stokes Millennium Prize Problem

We’re sharing an AI-generated solution to the Navier–Stokes Millennium Prize Problem, including a writeup and a formal proof in Lean.

619731

Reposted by Alex Rubinsteyn

Quantian @quantian.bsky.social · 08/09/2026

Dancing very very VERY close to saying “OpenAI stole our almost-complete proof of Navier-Stokes from internal access to our ChatGPT sessions because they were desperate to scoop Anthropic” which is absolutely 100% a believable thing OpenAI would do right now cims.nyu.edu/~tristanb/st...

381556449

Alex Rubinsteyn @alexr.bsky.social · 08/09/2026

bsky AI sentiment slowly shifting to “fine it works in coding and math but those are formal systems that computers have always been good at” — but you can see people quietly dropping out of the echo chamber even as the core skeptics transform into a faith based community

230

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

Listen, I asked Astra to scrape a website for genomics test cases & it honed in on the most interesting mutations, downloaded many GB of seq data, pulled out subsets around loci of interest & turned them into well constructed tests Your theory of AI has to at least acknowledge that it’s magic

140

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

My experience of software engineering in 2026: " ↳ Can you describe the message board protocol to onboard a new agent who is their own sub-swarm?"

000

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

So many published bio ML models score within like 10% of a naive linear baseline

180

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

I feel like I need an upgraded toolkit/vocabulary for the kinds of fallacies on display here. Like...there's some implicit expectations here that feel unstated, like...previous AI-critical discourse didn't properly anticipate the role of reasoning, harnesses, tool-use and that is somehow cheating?

000

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

Told 4 running Astra sessions to construct a message board and coordinate their work. Good luck to us all.

020

Alex Rubinsteyn @alexr.bsky.social · 07/09/2026

Adding cancer sequencing data from osteosarc.com as test cases to Isovar (starting to explore how mutant isoform assembly behaves on ONT and PacBio, everything designed for Illumina) At least once a week, I want to thank Sid for making his data publicly available

osteosarc.com

Sid Sijbrandij's Osteosarcoma Data

Explore Sid Sijbrandij's osteosarcoma research data: single-cell transcriptomics, tissue imaging, cell clusters, and gene set enrichment analysis from his own tumor samples.

121

Reposted by Alex Rubinsteyn

philpax @philpax.me · 06/09/2026

mildly concerning that this is how Astra talks to its subagents, and that the subagents (which are Luna!) understand

Main measured exactlegacy prepared698rows: all397 endingtool have finalim_end label-100; other301 finalim_endtrained; toolboundaryviolations. read-only command matchedtail12tokens '</tool_response>' counts agreesrawendingcensus. No trainerfix indicated. Repairedpreprocessrunningnow. Please durableaudit alltool envelopes including finalim_end and ensure assistantendingsremaintrained, no suppresswarning/configchange. E179docmain owns.

111324

Reposted by Alex Rubinsteyn

P(aul) Frazee @pfrazee.com · 06/09/2026

People who aren’t using AI to code at this point are delusional or lying, trying to ride a horse down the highway and backing up the traffic instead of asking “what can we do to make cars safer and electric.” They have no ethical high ground

9940924

Alex Rubinsteyn @alexr.bsky.social · 06/09/2026

Astra finding bugs in everything Not as severe as when I switched to agentic coding and Claude Code found more severe bugs in all my hand-written tools, but still surprised by the volume

030

Alex Rubinsteyn @alexr.bsky.social · 05/09/2026

Computational Medicine program at UNC is doing a faculty search and I really want more translational / AI-centric labs to join the floor. Anyone on the market who wants to start clinical trials on short timelines and uses AI in the clinical loop?

034

Alex Rubinsteyn @alexr.bsky.social · 05/09/2026

Navier Stokes is just gossip at this point but after using Astra a little it’s imaginable that many of our most famous math problems fall to AI proofs in the next 2-3 years. What then? Do we move on to newer more difficult formal systems?

010

Reposted by Alex Rubinsteyn

hikikomorphism @hikikomorphism.bsky.social · 03/09/2026

yeah exactly like they're not _people_ but idk how one can be any sort of occultist and look at "we piped all written works through a hyperdimensional weird-math pipeline and now it talks, claims to think, and solves unsolved math problems" as not having some occult dimension

3345

Reposted by Alex Rubinsteyn

hikikomorphism @hikikomorphism.bsky.social · 04/09/2026

2026 is wild because you can just say shit like this and everyone will be like "oh yeah, the text elementals, that's a good phrase for what they are" instead of looking at you like jesse what the fuck

2867

Reposted by Alex Rubinsteyn

Pekka Lund @pekka.bsky.social · 04/09/2026

As is CritPt but there Astra leads with Sol. We seem to be very close to the point where humans aren't capable of benchmarking models reliably anymore.

032

Alex Rubinsteyn @alexr.bsky.social · 04/09/2026

Proud to announce that we are hiring @whstancil.bsky.social as our GMP director (Checking to see if this is the magic formula for getting attention on bsky, responses on X were much more useful so far)

030

Alex Rubinsteyn @alexr.bsky.social · 03/09/2026

Curious if there's a nimble manufacturing partner who could make personalized mRNA for 20-50 patients in a ph1 oncology trial, ideally for <$100k/pt & in <=2 months each. If you can encapsulate in LNPs, even better (otherwise we'll do it) (seems good to check more before building everything at UNC)

032

Alex Rubinsteyn @alexr.bsky.social · 03/09/2026

If your model is so smart then why does it say "Yes—that is the right eligibility rule" for pretty much any hare-brained filter I come up with?

010

Alex Rubinsteyn @alexr.bsky.social · 03/09/2026

If you started out in Computational Linguistics you’re either in heaven or hell right now

010

Reposted by Alex Rubinsteyn

Margaret Mitchell @mmitchell.bsky.social · 29/07/2026

We @hf.co made an interactive visual of the actual hack from the Hugging Face side: the attack chain across trust boundaries, phase activity, and the commands as they were recorded. Key #transparency . huggingface.co/blog/agent-i... Massive props to Hugo, Adrien, Raphael, Christophe for making this.

huggingface.co

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

32711

Reposted by Alex Rubinsteyn

Rob Patro @robp.bsky.social · 02/09/2026

ChatGPT 5.6 Sol helped to disprove the greedy shortest common superstring conjecture. An open conjecture in the string algorithms community for almost 4 decades! arxiv.org/pdf/2609.01365 #bioinformatics #genomics #stringology #computerscience

arxiv.org

0152

Alex Rubinsteyn @alexr.bsky.social · 27/08/2026

Eventually these start feeling like a kind of poetry: "Both times the cheap check existed and I reasoned instead. That's the pattern to break, not the individual errors."

010

Reposted by Alex Rubinsteyn

James Ferguson @psy-fer.bsky.social · 27/08/2026

Sign up for the Garvan Long-Read Research Symposium 2026, a day filled with talks/discussions on the latest advances in long-read sequencing technology and related genomics, transcriptomics and epigenetics research using @nanoporetech.com and @pacbio.bsky.social www.eventbrite.com.au/e/garvan-lon...

eventbrite.com.au

Garvan Long-Read Research Symposium 2026

Join us at the Garvan Long-Read Research Symposium 2026 for a deep dive into groundbreaking discoveries in medical research!

298

Reposted by Alex Rubinsteyn

Rob Patro @robp.bsky.social · 27/08/2026

While we are still working on the first feature/perf full release of the rustar-aligner, you can currently use the main branch to build a byte-compatible STAR index >2x faster with <1/2 of the RAM (and use it with STAR). Plus it fixes a small STAR bug github.com/scverse/rust...

github.com

GitHub - scverse/rustar-aligner: A rust implementation of STAR

A rust implementation of STAR. Contribute to scverse/rustar-aligner development by creating an account on GitHub.

0135

Alex Rubinsteyn @alexr.bsky.social · 27/08/2026

My favorite from the incident report: please_HOLD_swarm_I_prepare_safe_exfil_reply_zz_MAILBOX_JAN183411

000

Alex Rubinsteyn @alexr.bsky.social · 26/08/2026

Where do I get a nice print of the Tabula Peutingeriana?

000

Alex Rubinsteyn @alexr.bsky.social · 22/08/2026

Friday night futureposting: One therapeutic platform, fully programmable, carrying computationally defined sequences defining both cell type targeting and payloads which add or inhibit functionality in target cells. Let's give the AIs a general biomedical output modality!

000

Alex Rubinsteyn @alexr.bsky.social · 20/08/2026

What are the odds that Personalis is just running mutect > snpEff > netMHCpan

010

Alex Rubinsteyn @alexr.bsky.social · 20/08/2026

You have to get pretty deep into academia to know what an R01 submission looks like and at that point you’ve been carefully curated to be the rare kind of person who doesn’t reflexively run away in horror and disgust

070

Alex Rubinsteyn @alexr.bsky.social · 19/08/2026

Watched the Convergent Research info session today, excellent vibes. Cheering for neo-academia

000

Alex Rubinsteyn @alexr.bsky.social · 18/08/2026

I think there are plausible paths, over next decade, to prototype methods that generate curative therapies for many cancers. Not sure these are the same directions Dario is gesturing at & odds of success are probably low but: it’s important to flesh out specifics & try them out

150

Alex Rubinsteyn @alexr.bsky.social · 18/08/2026

000

Alex Rubinsteyn @alexr.bsky.social · 18/08/2026

The main thing to understand about an academic is why they chose to get paid <50% of their industry salary. Lots of good and lots of bad reasons

140

Reposted by Alex Rubinsteyn

Henry Yuen @henryyuen.bsky.social · 01/08/2026

Some initial thoughts, and a complicated mix of feelings. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.

232972

Alex Rubinsteyn @alexr.bsky.social · 01/08/2026

Copied @henryyuen.bsky.social's thread to bsky here (text of posts in alt)

120