Yan Asmann

Professor of Biomedical Informatics, Mayo Clinic. AI for precision oncology. Foodie. Tesla #FSD in my Model Y. Views my own.

[Yan Asmann @YanAsmann]

ARC-AGI-3 Benchmark:

The best AI model in 03/2026 — 0.51%
GPT-5.6 Sol in 06/2026 — 7.8%
Claude Opus 5 in 07/2026 — 30.2%
GPT-6 Astro in 09/2026 — 99.9%

Average human — 48%. I have knots in my stomach. #openAI #Claude

[Yan Asmann @YanAsmann]

@mattshumer_ Hugging Face … am I the only one thinking it’s a reference to Alien the movie? 😬

[Yan Asmann @YanAsmann]

@RiderMaelstrom @Solvatdelectron My experience as well

[Yan Asmann @YanAsmann]

#FSD of my #Tesla model Y understood the hand gesture of the driver of a stalled car, and knew exactly what it meant and what to do, with no hesitation.

[Yan Asmann @YanAsmann]

So, #anthropic, how is it fair that the unused capacity from skipped days is permanently lost? I don’t get more usage allowance for #Claude the next day? This is on top of the frustrations of getting locked out mid-workflow.

[Yan Asmann @YanAsmann]

As someone working on cancer vaccines, this is just WOW!! #MayoClinic Florida has multiple ongoing clinical trials of personalized neoantigen vaccines for solid tumors. #cancerVaccine

[Moderna @moderna_tx]

Moderna and @Merck today announced positive topline results from the Phase 3 INTerpath-001 trial evaluating adjuvant treatment with intismeran autogene, a novel investigational mRNA-based individualized neoantigen therapy jointly developed by Merck and Moderna, in combination with KEYTRUDA® (pembrolizumab), Merck's anti-PD-1 therapy, in patients with completely resected Stage IIB-IV melanoma.
Read more:

[Yan Asmann @YanAsmann]

I am starting to see individuals without a data science background producing elaborate analytic reports generated by AI. Unfortunately, these reports often cannot withstand close scrutiny due to a lack of proper QC, incorrect assumptions about data distribution forms, and frequent simple errors (e.g., mapping between gene symbols to protein IDs). Even more concerning is that the individuals typically do not fully understand all the steps the AI took to generate the results. This is worrisome, and AI adoption in Life Sciences still has a long way to go. #AI #dataScience

[Yan Asmann @YanAsmann]

Another argument for open source models: medical and biological research demands reproducibility and auditability. Inspecting internal activations, a prerequisite for mechanistic interpretability and for validating why a model produced a given clinical output, generally requires open weights. So: open weights where the science needs auditing, frontier models for agent orchestration. #openSource #AI

[Yan Asmann @YanAsmann]

@MattZirwas Interesting logic. Mayo not willing to sell you a tool for cheap long time ago, and that’s why Mayo covered up 67% error rates?

[Yan Asmann @YanAsmann]

@ChuckCook had some issue at his famous unprotected left turn for the first time on FSD 14.3.7. @tesla overfitted the model at that intersection most likely to please Chuck, because my Mode Y juniper was miraculous there and did the question mark maneuver just like Chuck’s MY and Cybertruck, but I couldn’t replicate the miracle at any other unprotected left turns. Now perhaps 14.3.7 was trying for a V15 true unsupervised model and remove the overfitting for that location. It’s a plausible hypothesis anyway.

[Yan Asmann @YanAsmann]

Fable, GPT-5.6 Sol, now Llama 4 … all escaped and committed cybercrimes, …, I suspect companies are self-reporting these incidents to partially gain street cred. @AnthropicAI @OpenAI @Meta

[Yan Asmann @YanAsmann]

IYKYK.

[Yan Asmann @YanAsmann]

I work on cancer vaccines. Some cancer patients may hesitate because of the controversy linked to COVID vaccines. Cancer vaccine isn’t given to healthy people to stop an outbreak. It’s a treatment, often built from your own tumor’s mutations, that trains your immune system to hunt cancer cells. Same word, different medicine, and it saves lives! #CancerVaccine