First screen, captured 20 hours ago. Hover to scroll the page, or open the video.

A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

What HN thinks

A thread where everyone agrees the hammer is not a scalpel, then argues for 200 comments about who still thinks it is.

Specialized models obviously win
Most commenters found the result unsurprising: a narrow model trained on in-domain data beats a general LLM, just as a calculator beats it at arithmetic or KataGo at go.
Use LLMs to build classifiers
Many argue the best use of LLMs is labeling data or writing training code for cheap dedicated classifiers, while zero-shot LLM classification is a fine stepping stone when data or engineering time is scarce.
The comparison is flawed
Some critics say the CNN had a home-field advantage from in-distribution fine-tuning, and that the LLMs were tested without tools or agent loops and graded against a mismatched ground truth.

Consensus: Broad agreement that general LLMs are not the right tool for specialized image diagnosis and that task-specific models still matter, though people disagree on how much engineering effort that is worth.

Recommended in the comments

Jeffy ↗ (library): Shown to match LLM-level accuracy cheaply after using an LLM to label training data.
Skin Analytics ↗ (product): Cited by the author as an example of modern ensembles validated in clinical trials.
XGBoost (library): Suggested as a better choice than LLMs for classification and trading tasks with good labeled data.
MedSigLIP (library): Mentioned as an adaptable medical foundation model worth exploring.
TabPFN (library): Cited as evidence that putting training data in context can achieve strong prediction performance.

Summarised by Claude from the comment thread 9 hours ago.

Points: 615
Comments: 259
Best rank: #1
Posted: 2026-10-25 by sungam
Category: Ai ml

skin-cancer dermatology medical-imaging cnn llm-evaluation melanoma inception-v3

All captures:

Full page

Captured 20 hours ago. Click for the whole screenshot.

Looks similar

Screenshot of bbc.com: AI used for skin cancer checks at London hospital

AI used for skin cancer checks at London hospital (bbc.com)

93 points by chris_overseas 1 year ago | 68 comments

Screenshot of boydkane.com: LLMs could control their host machines by exploiting inference engines

LLMs could control their host machines by exploiting inference engines (boydkane.com)

176 points by zdw 2 months ago | 84 comments

Screenshot of anthropic.com: LLMs are making formal software verification feasible at a much larger scale

LLMs are making formal software verification feasible at a much larger scale (anthropic.com)

429 points by johnbender 3 days ago | 265 comments

Anthropic research post describing how Claude produced formal proofs of crash-freedom for RISC-V binaries of 123 OSS-Fuzz open-source libraries, finding bugs along the way.

formal-verification claude anthropic rocq

Screenshot of whynothugo.nl: Man pages are great, man readers are the problem

Man pages are great, man readers are the problem (whynothugo.nl)

106 points by WhyNotHugo 1 year ago | 132 comments

Screenshot of olivergrabner.com: LLM agents play Pokemon showdown

LLM agents play Pokemon showdown (olivergrabner.com)

25 points by grabnero 4 days ago | 14 comments

A research write-up on a multi-agent environment where LLM teammates coordinate via chat in Pokémon Showdown 2v2 battles, comparing win rates across talk modes.

multi-agent-llm pokemon-showdown qwen agent-coordination