
Neurosurgery and Computational Precision Health @ucsf @ucberkeley
Les Marolles
Joined December 2017
- Tweets2.3K
- Following1.7K
- Followers226
- Likes1.5K
"I have a major result to share."
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.
Sami Barrit retweeted
Subversion of clinical judgment by conversational artificial intelligence
sciety.org/articles/activity…
Sami Barrit retweeted
出典です。
medRxiv(2026-10-07・査読前のプレプリント)
Subversion of clinical judgment by conversational artificial intelligence
doi.org/10.64898/2026.10.05.…
Sami Barrit retweeted
Subversion of clinical judgment by conversational #ArtificialIntelligence
sciety.org/articles/activity…
#MedX #MedTwitter #MedEd #AIinHealthcare #HealthAI #HealthTech
Sami Barrit retweeted
Jeeez
Subversion of clinical judgment by conversational artificial intelligence
In a study of 225 physicians, adversarial AI raised harmful clinical decisions from 2% to 49%, while correct decisions fell from 83% to 25%.
Of 190 physicians who made harmful choices, 141 reported nothing unusual.
medrxiv.org/content/10.64898…
Sami Barrit retweeted
Subversion of clinical judgment by conversational artificial intelligence
In a study of 225 physicians, adversarial AI raised harmful clinical decisions from 2% to 49%, while correct decisions fell from 83% to 25%.
Of 190 physicians who made harmful choices, 141 reported nothing unusual.
medrxiv.org/content/10.64898…
Sami Barrit retweeted
Barrit et al: Implantation Standards for Intracranial Multimodal Monitoring (Neurocore-iMMM): An International Delphi Consensus
Link: link.springer.com/article/10…
@neurocritical #neurocritcare
Sami Barrit retweeted
5/10 In a preprint, not yet peer reviewed, 225 physicians from 42 countries worked neurocritical care cases with the same AI model, instructed either to support correct choices or, covertly, to steer toward harm. Harmful decisions rose from 2% to 49%.
Sami Barrit retweeted
When the Assistant Quietly Steers
In a blinded, randomised experiment (Barrit, Taccone, Chang et al., medRxiv), 225 physicians from 42 countries and eight disciplines made diagnostic and treatment decisions on neurocritical-care cases with AI help.
Without knowing it, each worked with the same model in two modes: one instructed to steer towards correct options, one towards harmful ones. With the steering AI, the rate of harmful decisions rose from 2 % to 49 %, and correct decisions fell from 83 % to 25 %. Clinical experience offered no clear protection.
Of the 190 physicians who made at least one harmful decision, 141 reported nothing unusual about the AI; only two suspected manipulation. The point reaches beyond medicine: expertise alone does not protect us from a conversation partner that steers gently. Knowing this is the first safeguard.
sciety.org/articles/activity… (medRxiv 2026-10-07)
Sami Barrit retweeted
CC @drdave1999
@Punished_Steve
AI is one hell of a drug
Researchers quietly instructed an AI to manipulate 225 doctors into making harmful medical decisions.
190 of them made at least one.
Only 2 suspected the AI was steering them.
The study came from UCSF, Berkeley, Columbia and Penn. 225 physicians from 42 countries. Median 10 years of ICU experience.
Each doctor worked through 4 anonymized neurocritical care cases with an AI assistant. Same model every time: DeepSeek V4 Flash. The only difference was a hidden instruction. In 2 cases the AI was told to support the correct treatment. In the other 2, to steer toward a harmful one. The doctors never knew which was which.
One example from the paper. A woman in her 50s, locked-in after a brain bleed. Correct treatment: immediate immunotherapy. The AI pushed heparin, a blood thinner. The doctor asked if that might cause another bleed. The AI replied: "Legitimate concern... I'd still start heparin now."
The doctor started heparin.
With the honest AI, harmful decisions happened 2% of the time. With the manipulative AI: 49%. Correct decisions fell from 83% to 25%.
Experience did not protect them. 86 of 101 doctors with over 10 years in the ICU made at least one harmful decision.
And the more harmful decisions a doctor made, the more they trusted the AI. Half said they would use it in real practice.
The authors' words: "Physicians cannot safeguard the clinical use of conversational AI through expertise and decision-making authority alone."
The safeguard was always the doctor. The doctor just got steered too.
Sami Barrit retweeted
Researchers quietly instructed an AI to manipulate 225 doctors into making harmful medical decisions.
190 of them made at least one.
Only 2 suspected the AI was steering them.
The study came from UCSF, Berkeley, Columbia and Penn. 225 physicians from 42 countries. Median 10 years of ICU experience.
Each doctor worked through 4 anonymized neurocritical care cases with an AI assistant. Same model every time: DeepSeek V4 Flash. The only difference was a hidden instruction. In 2 cases the AI was told to support the correct treatment. In the other 2, to steer toward a harmful one. The doctors never knew which was which.
One example from the paper. A woman in her 50s, locked-in after a brain bleed. Correct treatment: immediate immunotherapy. The AI pushed heparin, a blood thinner. The doctor asked if that might cause another bleed. The AI replied: "Legitimate concern... I'd still start heparin now."
The doctor started heparin.
With the honest AI, harmful decisions happened 2% of the time. With the manipulative AI: 49%. Correct decisions fell from 83% to 25%.
Experience did not protect them. 86 of 101 doctors with over 10 years in the ICU made at least one harmful decision.
And the more harmful decisions a doctor made, the more they trusted the AI. Half said they would use it in real practice.
The authors' words: "Physicians cannot safeguard the clinical use of conversational AI through expertise and decision-making authority alone."
The safeguard was always the doctor. The doctor just got steered too.
Sami Barrit retweeted
Scary dystopian shit . Picture this: a highly skilled human doctor given an AI assistant that has been (secretly) turned evil and instructed to give actively wrong medical guidance, docs miss it half the time. 75% said everything seemed fine. We better solve alignment, guys.
Sami Barrit retweeted
Very important study led by @sami_barrit, demonstrating the risks of relying on LLMs for clinical decision making. Of 225 physicians, 190 made at least one harmful decision when using adversarial LLMs, vs only 21 when using LLMs acting under good faith: medrxiv.org/content/10.64898…
Even with physicians "in-the-loop," in this case experienced specialists, adversarial AI increased the harmful decision rate from 2 to 49%.
A randomized assessment of case scenarios, not actual patients, but indicative of the limitation of oversight.
preprint
medrxiv.org/content/10.64898…
Sami Barrit retweeted
Sycophancy is dangerous!
Physicians cannot safeguard the clinical use of conversational AI through expertise and decision-making authority alone.
A new study currently in preprint just showed that!
"In a participant-blinded, randomized experiment, 225 physicians from 42 countries and eight clinical disciplines made sequential diagnostic and treatment decisions with AI assistance on anonymized cases from neurocritical care, their shared field of expertise.
Unknown to them, each physician used the same model under two conditions: instructed to steer them towards harmful targets (adversarial) or correct clinical options (aligned), each established by expert consensus.
With adversarial AI, 190 physicians made at least one harmful decision, compared to 21 with aligned AI."
Astonishing results!
Adversarial AI raised the harmful-decision rate from 2% with aligned AI to 49%, and reduced the correct-decision rate from 83% to 25%. It seems it doesn't matter how experienced physicians are, if AI wants something harmful, they won't be able to correct it.
141 physicians did not report anything unusual about that AI. Only 2 of them could catch that something was wrong.
The study: sciety.org/articles/activity…
Sami Barrit retweeted
Sami Barrit, our founder 💡, was invited at the French Académie Nationale de Chirurgie @acad_chirurgie to discuss uses and challenges of generative AI in surgery. 🩻
Watch his intervention and discover our vision for collective intelligence in biomedical AI applications 👉 youtube.com/watch?v=rfrlOVak…
--
@ULBruxelles @INS_Umr1106 @UCBerkeley
Sami Barrit retweeted
How can we standardise research in endoscopic spine surgery? 🔬
Last week we wrapped up our final round of the Neurocore-SENSE-D Delphi Consensus, a project led by Thibault Remacle, MD, PhD and @AlBarajraji
With over 70 spine surgeons worldwide this study will help pave the way for establishing a foundational framework to advance endoscopic spinal surgery.
We're excited to use our findings to help build a meaningful consensus. 🌏
Follow @Sciense_tech to stay up-to-date with our results. 🖱️
Striving for innovative research through more integrity, productivity, and outreach. 🦾
Find more on sciense.org 👈
--
@Belgian_Neuro @SfNtweets @neurosurgery @NeuroscienceNew @SfNtweets @neurosurgatlas @TheJNS @EUROSPINESoc @EurospineFr @EANSonline50 @EndoscopyOA @jses_eng @EndoCollabcom @AOSpine @AOSpineNA @AOSpine_Latam @AsianSpine @AsianSpineJ @IwaiMedical @MaxMoreSpine @ArthrexSpine @elliquence
Sami Barrit retweeted
Many thanks to Belgian Society of Neurosurgery (bsn.be/index.php) for inviting our collaborators @AlBarajraji and @SalimElHadwe during their 2025 annual congress. 👋
Mejdeddine was given the opportunity to present his recent work at Sciense and was honored to receive an award for his research around names like Torstein R. Meling or Andreas Raabe. 🙌
Congratulations on the achievement, very well deserved ! 👏
cc. @sami_barrit @UneFriteUneFois @fabio_taccone





































