Joined September 2013
- Tweets20.9K
- Following1.7K
- Followers252
- Likes11.4K
Jordan B. Tweeterson retweeted
I'm offering a USD$500 reward to the first person who can present pre-2020 human-written text that the latest version of @pangram flags as AI.
The specific requirements to claim the reward are:
1. At least 750 consecutive words.
2. Pangram flags it as 75% or more AI generated.
3. You provide proof that the text was first written in 2019 or earlier and is unmodified.
4. The text was ostensibly human written. This requirement only exists to avoid trick answers like output from GPT-2, machine translation, markov chains, the same word 500 times etc.
Regarding requirement #3, unfortunately in 2026 there's no way to prove that new text wasn't AI generated, so what I need provided is proof that the text existed in 2019 or earlier, before AI writing got so good. For example, valid proof might be:
- A link to the text on a website that's been logged by the Internet Archive's Wayback Machine before 2020.
- A link to a paper published on a third party website with a pre-2020 publication date. Note: I will need to be able to access the paper through the same website that has its publication date. Otherwise someone could send a modern AI generated paper and simply link to the date and abstract of a paper with the same name but different content.
- A chapter from a book edition with a known pre-2020 publication date where I can access a pre-2020 copy for myself. Note: For the same reasons as above, simply sending a passage of text and a link to a book's name and publication date doesn't prove anything on its own. The ideal situation would either be that there's an eBook of it available somewhere with a known pre-2020 upload date, or it's a well-known enough book that the content simply couldn't be faked everywhere.
Q&A:
"Why are you offering this reward?"
Some people are blaming the detector because they don't want to admit to using AI for their writing. I don't like people lying for personal gain.
"Are you affiliated with Pangram in any way?"
No.
"How do I submit something to you?"
Just reply to this tweet with the necessary info and I'll have a look when I can.
"How can I trust you to judge fairly?"
I'll publicly reply with my verdict on submissions and the world can judge me. The world can also judge you - please only submit something if it really fits all the criteria. Especially since I'll need to review all submissions in my free time.
"But it's my private writing, can I send it to you privately?"
If it's always been private then how will you prove that it's pre-2020?
"I have the perfect example but I left the text with my girlfriend in Canada."
She can submit it and win $500.
"Won't this be trivially easy to claim? I heard even chapters of Don Quixote and Moby-Dick get fully flagged as AI!"
Tell me which chapter and win $500.
"Conversely, isn't 750 words at >=75% a high bar? If Pangram is so reliable, why not 1%?"
At very low word counts Pangram IS somewhat unreliable, and with a monetary incentive I can imagine someone scanning every book on Project Gutenberg just to find one paragraph that's falsely flagged as AI. At the same time, most claims that Pangram is flagging human writing as AI have it showing the entire text as AI. I recently saw someone on X say that they personally knew of "thousands" of examples of human text that was flagged like this (unfortunately they declined to mention any). So I've set the bar somewhere in the middle.
"But Pangram was trained on existing text; that's why it always flags OLD text as human but falsely flags NEW text."
As far as I'm aware this isn't really how Pangram works. But I'd love to be able to award this bounty for new human writing, it's just that there's no way to verify people's honesty online for the new stuff.
"How will you pay out if I win?"
However works. Probably an international bank transfer via @Wise .
"How can I help?"
When someone says they know of an old published paper or story that Pangram flags as AI, you could point them to this tweet. They can win $500.
"How long will your offer last if nobody wins?"
As long as this post is up. If you're reading this, it's still available.
8 hours now
We’re seeing positive results with our efforts to restart affected infrastructure and recover server availability in an isolated manner. We’re actively monitoring the performance of these portions of infrastructure to verify stability, which we suspect will support improved availability as these actions are subsequently expanded more broadly. More info under EX1550543 and MO1550821 in the Microsoft 365 admin center, as well as msft.it/6019xBGoc if the admin center is not available to you.
50
Jordan B. Tweeterson retweeted
It's astonishing, isn't it, how half of veterans who served since 9/11 are disabled? It's really incredible how a war with 53,000 wounded produced 2.8 million invalids.
Good thing their crippling disabilities don't keep them from working -- 71.3% labor force participation rate.
Completely wrong. NPV would actually give politicians the incentive to go after votes all across the country
The electoral college is, without question, an inseparable friend to those who believe that every American across our vast country deserves to be heard.
We must fight to defend it from the far Left.
1
105
Jordan B. Tweeterson retweeted
If you're a China hawk, and you read the following revelations/allegations in a 48 hours period ...
1. Jensen Huang is sleeping with a Chinese spy.
2. Nvidia has been selling chips to Chinese AI companies for years through a Singapore intermediary overseen by that Chinese spy.
3. Nvidia’s lawyers pushed back on a federal investigation into this illegal scheme.
4. The White House knew all of the above and did absolutely nothing.
... I think it is understandable if what happens next is your head explodes
derekthompson.org/p/this-is-…
Jordan B. Tweeterson 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.
Yeah young men were drawn to Nick because of his coffee jokes. Just shut up
Tim Pool watched the famous Nick Fuentes black coffee clip and had the whole studio laughing.
He says the clip is an example of why so many people follow and watch Nick Fuentes, because he's not just political but he's funny and relatable too.
30
Jordan B. Tweeterson retweeted
My sources inside Nvidia tell me there’s a lot more dirt to come. Apparently the biggest issue the board had was that Jensen routinely turned in expense reports for hotels in the $350-$400/night range when the employee handbook policy explicitly limits this to $300/night
BREAKING: Nvidia, $NVDA, CEO Jensen Huang has resigned as CEO, President, and member of the Board of Directors, effective immediately.
Jordan B. Tweeterson retweeted
Conveniently there are hundreds of House members so you can look at broad aggregates to try to identify patterns
