Good morning. A winning AI score can hide a losing decision. Today, a skin research comparison shows why choosing the right test matters, Microsoft makes hundreds of coding agents work together, and mathematicians ask who gets to judge AI’s claims. The common thread: impressive output still needs the right checks.

Monday, October 26 · Read Time: 5 minutes

The Bite: The newest model does not automatically get the job. In an October 25 comparison, dermatologist Magnus Lynch reports that a newly fine tuned network built on the 2015 Inception v3 architecture correctly named 125 of 200 skin lesions. GPT 6 Astra got 84 right. Claude Fable 5.1 got 85.

Then comes the uncomfortable part. At its preset discharge threshold, the specialist marked 18 of 30 melanoma cases as needing no further care. The two general models would have discharged three and four. Better at naming the condition did not mean safer at deciding who needed care.

Why it Bites: This adds a specialist comparison to LLM results reported October 11. The small, difficult sample and single runs limit the conclusions. It does not establish clinical readiness.

Our take: Pick the failure you cannot afford before picking the model. A support bot can classify a complaint correctly and still issue the wrong refund. Test the decision separately from the label. And keep this research demo out of medical decisions.

You can’t fix the job market. But you can fix your resume… And the easiest fix is to send a resume tailored to each job, instead of one generic one everywhere. 

Kickresume just gave their Resume Tailoring tool a major update. It still does the tailoring for you in about 2 minutes, but they rebuilt it with the help of a Certified Résumé Writer and a team of HR people. 

It’s not an AI gimmick. Actual humans put in the work to make it transparent, accurate, and built so you stay in control the whole time.

PS: Kickresume also banned the AI slop.

The Bite: One coding agent can write a patch. Getting hundreds to share findings and combine their changes is another job. Microsoft released Agensh’s code and reproduction guide on October 24, following a September 22 research preprint. Agents claim tasks, exchange messages and merge work through shared infrastructure, without a central coordinator.

Why it Bites: Researchers gave agents six hours to rebuild programs from observed behavior and documentation, without Internet access. Across five difficult tasks, the average share of tests passed rose from 19.31% with one agent to 28.78% with 128, using GPT 5.6 Sol. The 1,024 agent demonstration covered one program, Pandoc. It does not mean 1,024 times the productivity. That fixes elapsed time, not computing cost.

Our take: More workers help only if their contributions survive integration. For a team considering this approach, start with a small, testable project and count accepted changes, review time and total cost. This is research infrastructure requiring Linux, Docker and model access, not a ready made replacement for a software team.

Who Gets to Choose AI’s Referees?
The Advisory Group on Mathematics and Artificial Intelligence wants labs to stop appointing particular mathematicians to oversee their releases. Its guest statement on Terence Tao’s blog argues that this can restrict access and borrow reputations. It recommends independent review processes. A useful question for any AI claim: who chose the reviewers?

Rehearse the Attack Without Running It
Redwood Research tests agent monitors with simulated tool replies, letting researchers examine malicious behavior without executing the dangerous calls. Results in its LinuxArena evaluation were promising, but they do not establish protection in deployment. The practical idea: test whether your monitor notices trouble before giving an agent real permissions.

🛡️ Less Code for Attackers to Reach
Tinfoil says it replaced more than a million lines of virtual machine firmware with about 300 lines of assembly. Its October 25 report describes moving work into the build process or Linux, rather than shrinking the entire system. The tradeoff: this works only with supported machine configurations. For teams protecting sensitive workloads, reducing what runs during startup is a useful design question, not proof of invulnerability.

PaperPilot: This early document toolkit, updated October 24, offers local PDF merging, splitting and OCR, with an MCP option for AI assistants. Try it on copies first and inspect the output. Its community edition uses Apache 2.0. Local processing does not guarantee privacy when a connected AI provider receives the results.

Ollama Speech: Its October 24 update lets Apple Silicon Mac users turn text into speech with Phonon. Install Ollama 0.40.11 or later and try reading a short draft aloud. Windows and Linux support is still coming. Standard English Phonon uses Apache 2.0; the smaller 40M models allow only noncommercial use.

Source checked discoveries, not hands on reviews.

One image is a genuine archival photograph. The other is an independent AI recreation. Which one is real?

Original photograph: NASA/Frank Michaux, May 27, 2025. Source and answer reveal. No NASA endorsement.