
|| PhD-ing Applied Math @UWaterloo š || Plundering N-Dimensional Seas || Working on world models for PDE Control || Seeking opportunities (Aerospace/Robotics!)
Joined June 2026
- Tweets1.3K
- Following298
- Followers316
- Likes3.3K
Pinned Tweet
No matter how isolated you are and how lonely you feel, if you do your work truly and conscientiously, unknown friends will come and seek you.
I think the problem with @pangram is that while they can claim that it will almost never classify human written as AI generated, I have many (persoanlly generated!) examples of it saying AI generated text is fully human.
4
5
326
Jonathan Gallagher retweeted
I have a major result to share.
Iāve found conclusive forensic evidence that the top 4 Chinese AI labsāDeepSeek, Moonshot, MiniMax and Ant Groupāhave built their newest models on NVIDIAās 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 patternāa sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isnāt based on any insider information or leaking. Itās 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 modelsāincluding DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3āwere initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIAās flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds Iād recovered, a Blackwell B300 reproduced DeepSeek-V4-Flashās untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3ās to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flashās, and changing the seed by one broke the match. (DeepSeekās 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. Itās 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/checkpoā¦.
There is much more Iāve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Do I just have AI Psychosis?
I spoke to some of the first years that I teach and they were saying that AI is more trouble than it is worth for their (first year!) python assignments and that it wasn't helpful to learn math?
Am I losing my mind? (Probably)
7
14
1,003
I was going over midterm grades with my first year Lin Alg students and heard a few of them saying:
"Bro the QED method works!!"
Of course I had to inquire.
Apparently if you don't know how to do the last step of a proof, you just skip to the conclusion and write QED.
Example #2304720 of the enshittification of everything:
I found a dry erase marker literally from 1988 and it still works better than a week old expo in the year of our Lord 2026
1
2
105
Currently teaching 50 first year engineers linear algebra
Rumors that congress wants to make this illegal !
2
15
561
Giving a talk today about agentic workflows in academic research
What are some of your favorite use cases? Which strengths or weaknesses would you highlight?
I would love to have some more to add !
1
3
74
This should be treated as attempted murder
Maximum prison sentences for people who are willing to play with the lives of the sick over a few Bitcoin
SITUATION DETECTED: The EU refrigeration cyberattacks have escalated to medical infrastructure. Hackers breached Coolworld, a French national blood bank contractor, changing transfusion plasma deep-freeze temps from -30°C to 110°C. It was caught before the blood was boiled.
49
I think I reached out to no more than five schools when I was considering graduate studies in math.
Of the five, five two told me to not bother applying and two accepted me.
I ended up in a PhD program because my masters supervisor and I developed a good working relationship!
Genuine question. Does it even make sense to apply to so many programs?? I thought applying for a PhD entails first reaching out to a potential supervisor and trying to build a relationship with them and working out a potential research q and then applying???
1
3
272
The Edge you are looking for isn't a 16-hour day
It's a 9-hour day with under 10 minutes of screen time
1
102











