James Lee (@JamesLeegit) header
James Lee (@JamesLeegit) profile picture

Biomedical Engineering PhD Student @Columbia & @nygenome (@NYGCtech)

New York, NY
Joined March 2021
Photo by James Lee (@JamesLeegit)
James Lee retweeted
[Please reshare] I’m recruiting PhD students to join my lab at @UW! Our current focus is on understanding the effects of social media algorithms and finding ways to improve them, but we are increasingly interested in how our findings from social media apply to chatbots/LLMs.
8
65
8
287
14,863
James Lee retweeted
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 unemploymentlabs.com/checkpo…. 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.
1,821
6,427
2,203
41,248
6,776,547
This is why Jensen was taken out.
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-W…
193
3,046
83
14,961
476,250
James Lee retweeted
A research question can be completely safe, but mention a pathogen or certain lab techniques, and some AI models will refuse to answer. Our latest benchmark, BenchBench-Refusals, tests how well leading AI models distinguish harmful from benign biology requests. See the results: benchling.com/benchbench/ref…
1
2
1
7
1,312
James Lee retweeted
Excited to share our latest preprint! 🎉 SPACE-Tag enables high-throughput spatial chromatin profiling with CUT&Tag. Huge congrats to Chao, Ruiyang, Sara, Maros, Silas, Ivan, and Sanja! 👏 🔬 biorxiv.org/content/10.64898…
14
1
61
5,814
James Lee retweeted
Demystifying the 17 biology evals in Anthropic's Fable 5.2 / Mythos 5.2 system card 🧵
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
9
25
4
154
16,989
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,368
7,258
5,755
55,916
10,086,005
James Lee retweeted
I’ve never been able to draw, until I learned to turn math ideas into code and explore visualizing things. More than 20 years ago I started with @WolframResearch Mathematica, now ~1000 #Python scripts: from complex analysis to statistical mechanics to fractals. #builtWithMarimo
24
81
6
1,130
59,374
Cell–cell interactions are key to all biological processes, but there hasn't been a good way to perform inter-cellular screens for genetic regulators of these interactions in vivo. We developed match-seq (multiplexed associative tagging of cell interaction histories) - using engineered virus-like particles to transfer barcoded mRNAs from 'sender' to neighboring 'receiver' cells, so you can reconstruct cellular interactions with sequencing alone, no imaging required. We coupled match-seq with CRISPR + scRNA-seq in cancer, revealing cancer-immune interaction neighborhoods at scale, and new cancer genes that can be targeted for tumor rejection via immune cell interactions. Led by the brilliant @peterdu_ with Bassik lab! biorxiv.org/content/10.64898…
CRISPR screens are great at telling you what a gene is doing inside your cells. But what if you're interested in what's happening to its neighbors? Today, I'm excited to share match-seq, a new method that deciphers cell-cell interactions using barcoded RNA transfer. 1/9
8
77
5
406
44,126
A new class of biotech buyer is emerging, as frontier AI and big tech companies acquire early-stage biology foundational platforms. How do big tech acquisitions differ from traditional pharma exits, and what are the potential implications for founders, investors and the pharmaceutical industry? nature.com/articles/s41587-0… rdcu.be/URMEtSZqA84a
12
71
11
256
59,426
James Lee retweeted
Replying to @kchonyc
Exactly. The "escapes" were made possible either through egregious negligence or deliberate purpose (marketing? Misplaced hopes of regulatory capture?).
154
528
103
4,009
491,249
Narrative violation: Astra is better than Sol, but still not SOTA on hard wet lab biology tasks Most biological reasoning is not captured in public data, so hard to train for this. Excited to see the progress though!
7
14
1
111
10,904
James Lee retweeted
A pervasive RT–qPCR artifact inflates RNA knockdown by RNA-targeting CRISPR - Yikey Yikes! BTW, similar concerns apply to siRNAs/ASOs that are sometimes suspected to affect RT and/or PCR . So nice somebody looked into this! rdcu.be/RUYFaGa649va
11
1
60
8,441
GDP per capita is super strongly correlated with life satisfaction across countries - but *negatively* correlated with meaning/purpose The AI-makes-us-all-rich future: lots of hedonic satisfaction, little eudaimonia? from @BetseyStevenson
15
70
7
291
21,704
I’m in love with this sentence: “Everyone knew it was impossible, until a fool who didn't know came along and did it.”
35
393
15
2,322
51,285
Hope "AI Scientists" don't go the same trajectory as many bioinformatics tools. 10000 flavors. Impossible to know who is better for what purposes. Everyone benchmark maxxing. Users bewildered & confused. Hopefully they will all be equally useful rather than equally useless.
21
29
5
350
19,992
James Lee retweeted
I wrote about the remarkable proliferation of cancer immunotherapies, and what's in store erictopol.substack.com/p/the…
16
110
8
430
43,834
James Lee retweeted
In a meta-analysis of 210 biomedical AI studies that statistically compared models under cross-validation, 97% used invalid statistical tests. Here's our new preprint doi.org/10.64898/2026.05.17.… led by @tianchuzeng @kkli20111 @ZShaoshi @ten_photos 1/N
9
95
17
275
122,669