Daily Frontier Digest ·
A ten-year-old image network beat GPT-6 Astra and Claude Fable 5.1 at diagnosing skin lesions
Curated by Alfredo Borrás · Plymouth AI Lab
A fine-tuned 2015 image network out-diagnoses two frontier LLMs on 200 biopsied skin lesions, Nanbeige releases an Apache-licensed 33B agent model, llama.cpp gets its first CVE, a 125B MoE runs on a 16 GB card from an SSD, and the Elvis estate sues Anthropic.
The story today
A ten-year-old image network beat GPT-6 Astra and Claude Fable 5.1 at diagnosing skin lesions
Dr Magnus Lynch, a consultant dermatologist at Guy's Hospital in London, published a test on 25 October in which he gave the same 200 skin lesions to GPT-6 Astra, Claude Fable 5.1 and a fine-tuned Inception v3, an image network Google first published in 2015. Every lesion came from the public MILK10k dataset, had a clinical photo and a dermoscopic photo, and was confirmed by biopsy: 100 that should not be sent home (including 30 melanomas) and 100 harmless moles and seborrhoeic keratoses.
The old network named the right diagnosis out of six for 125 lesions, or 62%. GPT-6 Astra got 84 and Claude Fable 5.1 got 85, both 42%. On melanoma specifically the scores were 13, 8 and 11 out of 30. The network was fine-tuned on about 68,000 public dermatology images on one H100, with the test lesions removed from its training data.
The LLMs failed in a cautious way. They kept 97 and 96 of the 100 dangerous lesions but cleared only 32 and 22 of the 100 safe ones, so almost everyone would be sent for a check. Each also had a favourite wrong answer: GPT-6 Astra called 75 lesions basal cell carcinoma when only 25 were.
Lynch lists the limits himself: 200 cases and 30 melanomas are a small sample, each LLM was run only once, every lesion was hard enough to biopsy, and the LLMs may have seen this public dataset in training. Inception v3 is a deliberately weak baseline; regulated skin-cancer products do considerably better.
Why it matters: General-purpose chat models are not yet a substitute for small specialised models in narrow visual tasks like this one. If you build in health or any image-heavy niche, a cheap fine-tuned network can still beat a frontier LLM, and Lynch has published the cases, prompts and scoring code so you can check.
Sources: Lynch Lab
Highlights
Nanbeige4.3-33B: an Apache-licensed dense model aimed at agent work
BOSS Zhipin's Nanbeige team released a 33B model under Apache 2.0 with a context of up to 512K tokens. In the team's own table it scores 23.8 on Terminal-Bench 4.0 against 6.3 for Qwen3.8-27B and 65.5 against 42.5 on DeepSWE, while Qwen3.8-27B stays ahead on SWE-Bench Pro (61.7 to 60.3) and GDPval (86.4 to 84.8). All the numbers come from the makers and have not been reproduced independently yet.
open-weightslocalSource: Nanbeige4.3-33B on Hugging Face
llama.cpp gets a CVE for a heap out-of-bounds write triggered by a malicious mmproj file
CVE-2026-116096, published on 25 October, covers an out-of-bounds write in the Pocket-TTS code that a crafted multimodal projector (mmproj) file can trigger, and a public proof-of-concept exists. Version b11796 is listed as unaffected; on current builds the bad file now hits an assertion and aborts, though the fix that blocks it by design is still an open pull request. The score is low (1.9) because it needs local access, which in practice means loading a downloaded file.
securitylocalSource: NVD
Anthropic sued by the Elvis Presley estate over 274 songs
Authentic Brands Group, which owns Presley's music and likeness, filed suit on Friday alleging Anthropic used 274 Presley songs to train its models without permission. The Guardian calls it one of the first major cases brought directly by an artist's estate against an AI developer, after earlier claims from the RIAA and the major labels.
anthropiclegalSource: The Guardian
Claude Enterprise merges chat and Cowork into one conversation
Announced on 23 October: on Enterprise plans, every conversation now starts as a chat and moves into a Cowork cloud sandbox when a task needs to run code or work with files, then comes back. It is off by default in existing organisations and is due to switch on by default on 16 November, so admins have about three weeks to decide.
anthropicagentsSource: PYMNTS
Sweden's government offices block ChatGPT and Claude on work computers from today
Citing Sveriges Radio, ForkLog reports that from 26 October staff at the Swedish Government Offices can only use internally approved tools such as Microsoft Copilot. Some staff keep access to ChatGPT and Claude through an internal platform. The Swedish Tax Agency and other bodies already have similar limits.
policyeuropeSource: ForkLog
Concept of the day
Why a 125B model can run on a 16 GB card: mixture-of-experts offloading
A mixture-of-experts model splits most of its weights into many small expert blocks and uses only a few of them for each token. Qwen3.8-Flash-Next has 512 experts per layer but activates 10, about 2% of its weights. So the always-used parts can sit in GPU memory while the experts live on a fast SSD and are copied in only when the router picks them, with a cache keeping the recently used ones close. The cost is that speed now depends on how fast the SSD can feed the GPU.
Source: MaxDevv/lean-qwenflash on GitHub
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Nanbeige4.3-33B | Nanbeige (BOSS Zhipin) | 33B dense | Apache 2.0, up to 512K context, tuned for agentic coding and office tasks. Scores are the makers' own. |
Try this
Lean-QwenFlash Run the 125B Qwen3.8-Flash-Next on a 16 GB GPU
This open-source runtime streams the model's experts from an NVMe SSD into the GPU. The author measured about 20 tokens a second of generation and about 1,600 tokens a second of prompt processing at the full 262K context on an RTX 4090 capped at 16 GB of VRAM and 8 GB of system RAM. It needs Linux, an NVIDIA card with 16 GB or more, CUDA 12.4+ and a fast SSD; clone the repo and follow the README to build the expert store.
Source: MaxDevv/lean-qwenflash on GitHub
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #3 (26 Oct 2026) · Subscribe free