Thaddeus Low (@tlow222) profile picture
Joined March 2021
Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)Photo by Thaddeus Low (@tlow222)
Thaddeus Low retweeted
Why Jensen was fired.
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.
171
612
36
7,023
689,923
Thaddeus Low retweeted
i feel for the math people. us software engineers got a few years of “oh shit, it’s coming.” they got months.
A quick check shows that this list claims to fully solve 90 of the top 500 open problems in math (proofatlas.ai/open-problems/). The highest ranked would be:
15
12
1
331
9,013
This is probably gonna make me sound crazy, but I think the age of the intuitive mathematician is on us. I wonder if this unlocks a lot of Ramanujans.
1
1
24
Thaddeus Low retweeted
stated preferences: the future of humanity is of critical importance and china cannot win revealed preferences: hires mainly childless employees and chinese citizens
7
19
4
647
15,336
Codex mini is way too cute.
1
36
Thaddeus Low retweeted
Three years ago, I accidentally uploaded a four-second audio clip of myself sighing and saying "please just let me die" to the official corporate announcement channel on Slack. I thought I was searching for a file, but I somehow clicked the microphone icon. The worst part is that it was a Monday morning, and my boss, an incredibly earnest man named Greg, thought it was a genuine cry for help. Instead of just deleting it or laughing, he scheduled an emergency one-on-one "wellness check" with me. I had to sit there in his glass-walled office and explain that I wasn't actually suicidal, I was just trying to open an Excel sheet that had been loading for too long. He didn't really believe me, so for the next six months, he would randomly leave small stress balls and herbal tea bags on my desk with a solemn, knowing nod. I couldn't even tell him to stop because it would look like I was in denial. I still have a drawer full of lavender tea.
201
787
97
24,195
563,791
I have been on both sides of this. Now I always double check with multiple strangers as a traveler. Do it blatantly in front of the 1st opinion if you must.
I accidentally gave someone directions to the SFO airport instead of the Oakland airport. I realized my error right after the bart doors closed. I'm sure they missed their flight. this will haunt me the rest of my life
1
240
“Would you love me if I were a worm?”
Step into the craziest visualization of the fourth dimension you have ever seen
2
78
Thaddeus Low retweeted
Today, we’re releasing Kalypta, the first app to block AI notetakers in your meetings. Granola? Wisprflow? Cluely? No more. With Kalypta, you become inaudible to AI. Your call continues normally.
721
922
548
16,091
2,392,769
Thaddeus Low retweeted
Big AI news... Jev enables this microphone that mutes you when you talk BS @typesafeai
14
12
8
175
27,817
“Reality throws an error” Bro it doesn’t. Navier stokes is continuum mechanics math. Continuum is an approximation. A continuum assumption breaks down when you go small enough and closer to discrete. It’s the limits of the mathematical framework.
What if the Navier-Stokes breakthrough just accidentally proved we’re not living in base-level reality? Think about what it’s saying… We have equations that describe how fluids behave incredibly well…until you push them far enough. Then something bizarre happens. The equations can create a vortex that gets smaller and smaller while its velocity approaches infinity. Reality basically throws an error… But obviously nothing in physical reality is actually moving infinitely fast. So what does the infinity represent? Maybe we’ve hit the resolution limit of the layer we’re observing… Like a video game. From inside the game, everything looks continuous. But zoom in far enough and eventually you hit the limits of the engine rendering it. Physics keeps doing this. Black holes…singularity. Big Bang…singularity. Now even fluid dynamics…singularity. Maybe infinity isn’t something that exists in nature. Maybe infinity is what the math spits out when we try to look beyond the resolution of our layer of reality. In other words… Maybe a singularity isn’t where reality breaks. Maybe it’s where the simulation stops rendering.
1
95
PhD scientists are CPUs. Engineers are GPUs.
Gavin Baker says the physics PhDs calling orbital data centers impossible on X have not out-thought the 10,000 SpaceX engineers who treat it as solved "It's very hard for me to engage. There are all these people on X and they're like, I am a physics PhD and this is impossible." "There's a friend, another investor, who actually is a physics PhD, who I had many arguments about it with him, and he's like, I am a PhD and this is impossible. And then he goes to the SpaceX day and talks to the SpaceX engineers, and he's like, well, I was wrong." "So let's say you're an astrophysics PhD. You are brilliant, you're hanging 100 IQ points on me. Have you thought about this for an hour? Have you thought about it for 10 hours?" "Because you have 10,000 of the world's smartest engineers at SpaceX who've thought about this each for hundreds if not thousands of hours. And the sum of that, working with very sophisticated engineering tools, is it's a solved problem." _______ For more on Gavin Baker: firesidealpha.substack.com/p…
40
It just occurred to me that I started following levelsio because hoodmaps cracked me up. I didn’t even know what he was known for before that
Strava's data is quite protected so I couldn't index it for Hoodmaps But the good thing is the US Census data publishes median income level and the data is very fresh and accurate, so I added it to 🗺️ Hoodmaps.com So now you can switch to [ 💰 Income mode ] and instantly see wealth or lack of wealth concentrated
68
Forgiveness is an insane concept. Because it allows “it is better to ask for forgiveness than permission”
They totally killed Claude with this ethics disclose stuff. It won't generate anything. Insane.
2
94
Code review is more about consensus than it is correctness. It’s a consensus on who owns cleanup when things go wrong. You fully remove it only when you can fully blame it on the machines or rely on them to cleanup.
i personally think code review is dead. the team does not agree. directionally found this surprising.
2
4
893
AGI has been accomplished.
Claude Science is incredible. I gave it some sequencing data, and in 8 hours it did a full analysis, generated figures, wrote a paper, submitted it for publication, got rejected, revised and resubmitted, got rejected again, it is now applying for positions in industry
41
Training a depth model on more datasets should make it generalize better. When we tried this with ECoDepth, it actually made it worse. We weren't able to explain it, so we decided to ship single-dataset models. Two years later I finally know why. [thread]
Another banger image2depth from ECoDepth. Very useful for self-driving scenarios! nitter.cf/AradhyeAgarwal/status/…
8
23
435
75,969
This is the funniest thing I have seen in a while on here
This quoted post is unavailable.
36
Thaddeus Low retweeted
Science’s paradox: we rely on youthful, high‑variance creativity, but force people through a decade+ of training that nurtures conformity and risk‑aversion. By the time you get real power, you’ve been overeducated out of the mindset that made you disruptive in the first place.
26
42
6
446
25,926
Never underestimate the damage a kid can do yelling “He’s an unc!!”, as he runs away laughing merrily with his friends.
My boyfriend suddenly became obsessed with going to the gym. New clothes. New haircut. New confidence. I thought he was preparing to leave me. One day I asked why he'd changed. He looked embarrassed. Then admitted: "One kid called me 'uncle' last month."
78