It is 2026. How are we still publishing articles on medical diagnostics data science and using area under the ROC curve as the primary metric of success. ROC-AUC of 0.9 under severe class imbalance (almost always the case in diagnostics) could still mean something like 4/5 predicted diagnoses are wrong (false positives). Precision-Recall curve + mAP or GTFO.
Also, the most interesting result here is that the CNN-based feature encoder significantly outperformed a vision transformer encoder backbone…
steve-atx-7600 2 days ago [-]
Teaching students how to interpret evidence must be way undervalued still. I went to one of the top CS schools 20 years ago and you could get a degree without even taking a single probability or stats class of any kind.
chrisjj 2 days ago [-]
> How are we still publishing articles on medical diagnostics data science and using area under the ROC curve as the primary metric of success.
Waste avoidance.
Bullsh*t is more than sufficient to convince an AI-gulled target audience.
Zaraif13 2 days ago [-]
I don't agree with the flak Chinese labs get. If it's really that easy to distill and compete with frontier models, why aren't other countries anywhere near this AI race?
felixgallo 2 days ago [-]
Because distillation is a friendly term for industrial espionage, and most other countries are not willing to become international pariahs in the eyes of the west.
both things can be true. China's gearing up, but they also try to make progress by aggressively distilling Anthropic and OpenAI models, and this is currently where most of their progress comes from.
yogthos 22 hours ago [-]
People really need to stop parroting this line uncritically. The process takes time because even when you're distilling answers, you still need to actually do reinforcement training on the model. And given that Fable and GPT 5.6 just came out there simply hasn't been much time to do that. However, models like Kimi also do better than Fable or GPT on a lot of tasks, which means it's not just distillation but also difference in architecture. You can watch this talk from Kimi founder to see how Kimi was actually trained and why it performs well. https://www.youtube.com/watch?v=5CkCW1P-g88
It's also absolutely hilarious that people think only Chinese companies use distillation, as if Anthropic or OpenAI are above that or something. Not to mention that they basically ignored copyrights on all the data the siphoned and are now crying that people aren't respecting their terms of use.
Chinese labs have come up with a bunch of genuine innovations: GRPO, auxiliary loss free MoE load balancing, MLA, muon optimizer, and a bunch of other ones. The Deepseek papers are really well written, this isn’t just sneaking a peek at a peer. Anybody who thinks China is simply distilling glorious American models is not engaging with reality.
4d4m 13 hours ago [-]
Every model provider distills each other even if for benchmarking. Agree with all your points.
yogthos 8 hours ago [-]
For sure, everybody distills when they can, it would be stupid not to. I'm just pointing out that Chinese companies clearly do their own research and innovation just like American companies do. It's not that they just wait for American models to drop and then distill them.
Written to run on something like an H100 though- as the CT Scan data is quite large.
anon373839 3 days ago [-]
The model weights are only ~5GB, so this is small.
cyanydeez 3 days ago [-]
[flagged]
Bolwin 3 days ago [-]
Alibaba is entirely capitalistic I'm afraid
netsharc 3 days ago [-]
In communism, man exploits man. In capitalism, it's the other way around.
fragmede 3 days ago [-]
Fortunately, we now have AI to come along and launder the responsibility for exploitation through.
reilly3000 3 days ago [-]
Now we have AI exploiting us. Look at us race to digitize everything and feed it into ever rising oceans of context, to build infrastructure for its rapid expansion, confident that we can’t stop because we can’t all stop; we’re just humans. All we have to do is stop. And we can’t. That will end up not in our extinction, but in our ever increasing subjugation, both of freedoms and spirit.
> EPS3.9 also had significant anti-tumor effects in the mice with liver cancer and activated anti-tumor immune responses
“A Novel Exopolysaccharide, Highly Prevalent in Marine Spongiibacter, Triggers Pyroptosis to Exhibit Potent Anticancer Effects” (2025) DOI: 10.1096/fj.202500412R https://faseb.onlinelibrary.wiley.com/doi/10.1096/fj.2025004...
"A Gemma model helped discover a new potential cancer therapy pathway" https://news.ycombinator.com/item?id=45604231 :
> eCPMV VNPs + EPS3.9 + [...]
"Scientists are discovering a powerful new way to prevent cancer" https://news.ycombinator.com/item?id=45474404
Notes re: Kidneys not Livers: https://news.ycombinator.com/item?id=47460486 ; gh/topic/healthcare-ai
Science article in question: https://www.science.org/doi/abs/10.1126/science.aec6129
Also, the most interesting result here is that the CNN-based feature encoder significantly outperformed a vision transformer encoder backbone…
Waste avoidance.
Bullsh*t is more than sufficient to convince an AI-gulled target audience.
It's also absolutely hilarious that people think only Chinese companies use distillation, as if Anthropic or OpenAI are above that or something. Not to mention that they basically ignored copyrights on all the data the siphoned and are now crying that people aren't respecting their terms of use.
Chinese labs have come up with a bunch of genuine innovations: GRPO, auxiliary loss free MoE load balancing, MLA, muon optimizer, and a bunch of other ones. The Deepseek papers are really well written, this isn’t just sneaking a peek at a peer. Anybody who thinks China is simply distilling glorious American models is not engaging with reality.
Source: https://github.com/alibaba-damo-academy/damo-radar
Model: https://huggingface.co/radar-generalist
The attempt is already there: "predicting the future" is called "intelligence".
102M - modified BERT-base-Chinese text encoder
26M - 3D U-Net-style vision/anatomy encoder
2.8M - projection layers, anatomy-specific projections, query tokens and attention layer
Written to run on something like an H100 though- as the CT Scan data is quite large.