iiinova — where your digital self keeps evolving

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

Thanks to our team across Sheba Medical Center, NVIDIA and Mount Sinai. Want to see how we made a Nucleotide Transformer diploid? Check out our preprint: DNT: Diploid Genomic Foundation Model www.biorxiv.org/content/10.6... Models: huggingface.co/collections/...

biorxiv.org

DNT: Diploid Genomic Foundation Model

Clinical interpretation of genetic variation depends on the diploid genotype, including zygosity, allele dosage and whether multiple variants occur in cis on the same homologue or in trans on differen...

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

We also introduced Contrastive Phase Loss for phase-sensitive pretraining. At 100M parameters, cis/trans AUROC rose from 0.54 to 0.62.

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

Phase-blind models stayed random; DNT reached 0.65 AUROC. Diploid-awareness also improved ClinVar SNV and indel prediction, while haploid genomic-function performance was preserved or improved.

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

3/ We trained on trillions of base-pairs from the phased 1000 Genomes. We introduce new diploid-aware benchmarks for pathogenicity/VEP, zygosity, hemizygosity and ClinVar indels, plus cis/trans benchmarks where the variant pair, gene and genetic background stay the same; only phase changes.

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

2/ Zygosity and phase matter. In recessive disease, the same pathogenic variants can disrupt one or both gene copies depending on whether they are in cis or trans. That’s why we built DNT: a diploid genomic language model family that read both copies together, including SNVs and indels

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 15/09/2026

Our diploid aware genomic language model is out! 1/6) The genomes we train AI models on do not exist. Most genomic language models read a single haploid sequence, but human genomes are diploid.

151

Dan (✈️ ICML-26) @ddofer.bsky.social · 18/08/2026

Other countries not having physical seperations of bike VS car lanes is utterly insane to me. It's enough to have some bricks or plastic bumps!

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 27/07/2026

Reverse lecture/classrom ; reverse RPG

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 18/07/2026

It is absurd that I can reread #Watchmen so many times, and still find hidden details #comics (Also, absolute editions are a joy) #dccomics

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 17/07/2026

Ofer, D., Linial, M. & Shahaf, D. InterFeat: a pipeline for finding interesting scientific features. Nature Scientific Reports (2026). doi.org/10.1038/s415...

doi.org

InterFeat: a pipeline for finding interesting scientific features - Scientific Reports

Scientific Reports - InterFeat: a pipeline for finding interesting scientific features

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 17/07/2026

On 8 major diseases and 400K patients, we pick ~40 features per disease out of 3,700+. About a third got validated by medical researchers — vs ~0-7% for SHAP-ranked lists! #UKBiobank

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 17/07/2026

We consider a finding as only "interesting" only if it's (a) predictive, (b) not already known, and (c) makes sense. InterFeat checks all three automatically and explains why!

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 17/07/2026

InterFeat just got the top broad-applicability score in the ISBCB paper award (Israel's national bioinformatics society)! @michalinial @HyadataLab InterFeat helps researchers find the interesting findings in data: ones that are novel, make sense and actually useful.

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 16/07/2026

Conference and poster pics from icml @icmlconf.bsky.social : You're all welcome to use my photos. I request is that you give credit/cite (Dan Ofer / @ddofer ddofer@gmail.com) drive.google.com/drive/folder...

drive.google.com

ICML 26 Dan - Google Drive

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 12/07/2026

That's a wrap on this year's ICML. Fun and stimulating! To all those I met - happy to keep on talking science (and the rest). :)

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 30/06/2026

2 papers in the Foundation Models and Large Language Models for Life Sciences Workshop, 11.7: ROOM S317 - sessions 1 & 2 #jepa #sentencetransformers #FM4LS

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 30/06/2026

Heading to #ICML2026 Seoul! I’ll be there from July 3–20 presenting my work on Protein/Biological LLMs. (And travelling + photographing, geeking out). I'd love to connect, share ideas, co-travel post-con or collaborate. DMs open! @icmlconf.bsky.social #ICML #AI #eleutherai #korea

200

Dan (✈️ ICML-26) @ddofer.bsky.social · 19/05/2026

I tried aurora now, and holy shit, it beat ADAM (nvm muon). Without tuning params too!

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 17/05/2026

#Autoresearch on #proteins - amazing what you can do with a gpu and a few nights on the SOTA for language models. (Now I just need more gpus) Viva @karpathy.bsky.social

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 07/05/2026

4 papers written and submitted for #Neurips (machine learning conference), first author on 3 of them. Let's see which get in. Also, if I ever try to do this much research at once again, please stop me. With a big rock and bottle of chloroform. It'll be a kindness

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 13/04/2026

The solution is to vote him Supreme executive powers, after the attempt on his life left him scarred and deformed.

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 08/04/2026

Spire 2 is a bit easier than the first at base I think. (I'm in favour)

010

Dan (✈️ ICML-26) @ddofer.bsky.social · 05/04/2026

6. Most exciting <Censored> trick: The newest paper really does work much better, from being worse than MLM to outperforming it in a fair ablation. Performant implementations are hard with no compute in medicine and biology 🤐

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 05/04/2026

5. #Kimi Attention Residuals (AttnRes) + Block - nuked my initial results, horrifically slow if using just the pseudocode (dropping got me from ~6 sec/it to 0.8~1.4 it/sec). Maybe needs better kernels (or a better coder than Opus, gemini and gpt and me).

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 05/04/2026

3. Exclusive self attention: costs speed, but learning seems better, worth it. 4. Alt masking rates: jury out.

lnkd.in

This link will take you to a page that’s not on LinkedIn

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 05/04/2026

SOTA LLM architecture ablations for a novel encoder paper proceeding apace: 1. hybrid architectures (Not the one you're thinking of): Jury is out, but I like it. 2. Local/global (#modernbert) style: less compute benefit than I expected, maybe my implementation sucks?

lnkd.in

This link will take you to a page that’s not on LinkedIn

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 16/03/2026

No, the CEO who felt that he'd be laughed at if he paid lots of money was the problem

080

Reposted by Dan (✈️ ICML-26)

Leonore (Lee) Carpenter @leecarpenter.bsky.social · 09/03/2026

Yes everything is terrible but let’s not lose sight of the fact that someone has written the Headline to End All Headlines.

A headline reading “Camel pageant thrown into chaos after 20 competitors disqualified for using hump-plumping injectables”

622384701

Dan (✈️ ICML-26) @ddofer.bsky.social · 06/03/2026

Yeah, I'd run it yourself or more easily, check that it didn't edit them. Benchmark hacking (as in, deleting it) is a classic issue.

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 06/03/2026

Ah missiles on the ME. (At least we set up wifi in our shelter!)

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 06/03/2026

+ Does it pass their test suite(s)? Very cool if so, very hallucinated if not :)

100

Dan (✈️ ICML-26) @ddofer.bsky.social · 06/03/2026

What about repgrep ? I think that's what they use the most (espec for agents)

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 04/03/2026

Fire lots of rockets, play stupid games, win stupid prizes :|

010

Reposted by Dan (✈️ ICML-26)

The Kinsie @thekinsie.com · 13/02/2026

Unreal Tournament 2004 is now available for free. ...well, I mean, you probably could've grabbed a GOG offline installer out the back of a truck earlier, but NOW it's Official/Legally Free and has the in-development community patch packed in for ongoing support. www.oldunreal.com/downloads/ut...

3323661153

Dan (✈️ ICML-26) @ddofer.bsky.social · 03/02/2026

Huh. Now I wanna check if I can find anyone I know there

000

Reposted by Dan (✈️ ICML-26)

Davis Bickford @davis.social · 01/02/2026

yeah, so Fallout nailed it

Screenshots from Fallout showing Lucy, along with the words “One side is murdering people, enslaving them, crucifying them,
and the other side is just vaguely problematic.”

123589697

Dan (✈️ ICML-26) @ddofer.bsky.social · 22/01/2026

I feel this. (We're Currently arguing with a reviewer over why we had "only" 5 expert PhD-MD doctors review and annotate "only" a few hundred samples for validation). It's hard and expensive to do!

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 18/01/2026

Just shut up and take my money, you filthy liberal nerd <3

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 16/01/2026

Eh. Not really, even if you're looking at agents, acting from feedback or code. They ain't geniuses, but then again, neither are more people

000

Dan (✈️ ICML-26) @ddofer.bsky.social · 31/12/2025

Happened to me when heating up restaurant leftovers in the takeaway box. (Inside of cover was aluminum). Good way to find out that my fire extinguisher had expired I guess.

010