Excited to share a new preprint from our lab, led by our staff scientist Nseer Iqbal: the first high-resolution structure of a bacterial flotillin assembly. https://t.co/QO7SiPLShC
AI Overbuilding hypothesis proven 🧵 1/n
Boys and girls, you have been hearing from me about AI overbuilding for more than 6 months at least. SemiAnalysis now indicates that it is actually happening. And this is not due to Jensen saga alone.
Frontier labs - in their wildest growth scenario - can not consume all this compute we are building! My estimate capex will be cut as early as H2 2027.
Newer model architecture (e.g. encoder-decoder), as well as, newer rack scale systems make it possible to produce far many more tokens than we can consume.
@jukan05@teortaxesTex@stevehou
About 90% of RNA molecules in our cells are tRNAs. To date, we have mostly been blind to their abundance and regulation in single-cell sequencing. Today, I’m excited to share a new manuscript from my MD/PhD work
https://t.co/pQqVF7tAZI
Announcing Generalised Superintelligence! (GSI) 🎉
GSI has achieved 250+ novel discoveries and breakthroughs across eight domains, including Quantum, Superconductivity, Chip Design and Mathematics through its Meta Solver Platform. GSI is extremely compute efficient compared to other frontier labs and excels at finding the hard structure in a problem first, before attempting it.
Everyone can contribute to scientific breakthroughs using the kind of superintelligence GSI is building.
Research preview out now at https://t.co/LSwU67gczE
BioLLM-8B beta is live!
Patent-pending adapter that writes living human neuron activity into LLM layers.
J-space explorer shows exactly how the cultures shift every word, layer by layer.
8B free now → https://t.co/1WSvr17lRn
14B & 27B dropping soon
Let's dive in 🧵👇
Genomes are long but redundant. Not great for pretraining.
To fix that, we've developed a lossless tokenization strategy that shortens sentence length by >2000x.
For pretraining, we achieve a 0.55 validation CE on All of Us.
Public replication on 1KG linked below.
ARIES aligned protein sequences more accurately than established methods, especially when sequences shared little similarity. It uses numerical representations of amino acids learned by protein language models to build a shared template for alignment. https://t.co/omEao6oIU7
In neuroscience, a canonical microcircuit is thought to subserve all cortical computation. In AI, neural networks universally subserve intelligent behavior. Is there convergence between computational motifs in natural and artificial intelligence? Work with @dyamins 🧠🤖
Finally published:
“Symmetry as a common origin of hierarchical emergence and abstraction”
https://t.co/tyN59jdI0Y
Symmetry generate hierarchical macroscales and shape the structure of Bayesian beliefs, making inference of high-dimensional systems tractable
Big reasoning speedups are here. 1,000× faster training and 280× less compute than any looped model. All done using a fully trainable energy from the 80s
Keep reading to hear how energy is the right paradigm for reasoning
With @HughBlayney@jcrwhittington@prefrontal_ai
🧵/10
1/ 🧵 New preprint! Do protein foundation models really understand local chemistry?
We probed them with NMR chemical shifts, per-atom reporters of chemical environment.
Short answer: they struggle exactly where the biology gets interesting 👇
[link] https://t.co/2DhcjYGM99
Excited to share the second piece of my postdoctoral work with @JShendure. We solved the last remaining "tangle" -- the nucleolar organizer region (NOR) -- in the telomere-to-telomere assembly of the human genome with gene editing. (1/n)
https://t.co/JMLJHR8d83
🧠 How do brain cells maintain their identities and functions for decades?
Thrilled to share our new study in Cell! We mapped the regulatory genome of the adult human brain to explore cell identity, developmental memory, and disease risk: https://t.co/JPoWpz4Zaf 🧵 (1/14)
Excited to share what we've been working on since this summer- started in my internship at @MSFTResearch!
Modern DNNs predict neural activity surprisingly well, but are black boxes. We introduce interpretable weight-space programming (iWSP) and ask: can LLM agents build predictive models of the brain directly in code, with no training or pretrained weights, so every feature stays transparent? 🧵 (1/6)
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.