MD, PhD — Computational Bio & Precision Medicine.
Joined September 2010
- Tweets31.5K
- Following4.4K
- Followers7.3K
- Likes66.8K
Peter Ebasone retweeted
a lot of people speak with the kind of confidence you can only get by never having done or having to do the work
Interesting preprint:
PALM fine-tunes ESM2 to predict protein domains and functional sites. Without ever training on ClinVar, its zero-shot score (paired with ESM2) beats AlphaMissense on 128k literature variants outside ClinVar.
biorxiv.org/content/10.64898…
Peter Ebasone retweeted
People who reason by first tracing things end-to-end have a very different process that can look particularly stupid until it very, very suddenly clicks in to a creepy level.
Peter Ebasone retweeted
Your genetic risk variant may have been harmless for 60 years. Then the aging RNA machinery stops buffering it.
New review in Trends in Cell Biology by our grad student @sgrandles : RNA dysregulation belongs alongside DNA damage & proteostasis as a driver of neuronal vulnerability. sciencedirect.com/science/ar…
if you’re unsure about where you want to go in life, an easy step forwards is to figure out where you absolutely don’t want to be and get away from there.
Peter Ebasone retweeted
It's not just species of the gut microbiome, or their metabolites, but also genomic variants in the bacteria, that relate to diseases and health
nature.com/articles/s41586-0…
Peter Ebasone retweeted
New in @NatStructMolBio! 🧬 We screened 730 RNA-binding proteins and found 281 that control circular RNAs, from how they’re made to how much protein they produce. Thanks to our collaborators at MSK, Columbia, and UCSD, and to my amazing lab, led by Tao Yu nature.com/articles/s41594-0…
Peter Ebasone retweeted
BE Seminar this Thursday with @arjunrajlab @PennEngineers “Systems of Cells and Systems of Science” ow.ly/b7Yl5103agn
Interesting paper: Fifteen challenges for generative AI applications to cell biology
sciencedirect.com/science/ar…
Peter Ebasone retweeted
Learn linear algebra. Learn to visualize matrices as spatial transformations. Learn to see calculus as the language of optimization. Learn vector spaces until neural networks stop looking like black magic and start looking like high-dimensional geometry. A deep understanding of math transforms you from an API-wrapper developer into an AI architect.
this is very promising for small labs and individuals.
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.
The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up.
So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.
Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.
alphaxiv.org/abs/2610.39302
Peter Ebasone retweeted
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.
The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up.
So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.
Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.
alphaxiv.org/abs/2610.39302
Peter Ebasone retweeted
LLMs accumulate a lot of declarative knowledge.
That's how they can answer a large number of questions.
That's also why they require so many parameters and hence so much memory.
But one can replace the accumulation of knowledge by *understanding*
That what JEPA world models do.
That's one reason they can be useful with considerably less memory and training data than LLMs.
This paper merely shows that the underlying similarity structure of the data, whether text or image (e.g. co-occurence of words in text or objects in images) uniquely determines the shape of the representation point cloud, when the encoder is trained trained with a common anti-collapse criterion like SIGReg or VCReg (not by reconstruction).
Peter Ebasone retweeted
The problem isn’t that AI is intelligent. It isn’t. The problem is that too many adults (but not all) are at once mechanistic, imitative, unimaginative, stupid, formulaic, influenceable, & gullible, arrogant beyond their skill set, & too expensive to maintain.
Peter Ebasone retweeted
Everyone wants to have solved a hard problem, but almost nobody wants to live inside the grind of actually solving it. The breakthroughs belong to the few who are obsessed with the process, not just the idea of the outcome
Peter Ebasone retweeted
🚀 Introducing Next-Concept Modeling (NCM): our new paradigm for fast, parallel, coarse-to-fine text generation.
Compared to strong diffusion language models (DLMs) at similar parameter size, NCM achieves:
⚡ Up to 31× faster generation on downstream tasks like summarization and machine translation with comparable or better quality.
🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity and using 3–5× fewer training tokens.
🧩 Better performance on reasoning tasks like Sudoku, outperforming both AR and diffusion baselines with fewer sampling steps.
With @ishanskhare, @IddahMlauzi, @ekellbuch, @simran_s_arora, @Azaliamirh, and @HazyResearch.
More details in the thread 🧵👇
Peter Ebasone retweeted
A @NobelPrize for a Badger 🏅
Congratulations to Francis Halzen, named the 2026 Nobel laureate in physics for his groundbreaking work with @uw_icecube to advance our understanding of neutrinos and the universe’s most energetic events.
news.wisc.edu/university-of-…
