Ask HN: what do you do with AI code you don't understand?
The most-discussed Hacker News thread of the day has over 500 comments. People ask how to ship code they didn't write and can't fully read, and whether tests and review are enough.
Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.
What tech Twitter is discussing, ranked by discussion.
The front page, plus stories over 80 points in the last 24 hours.
Every new issue, each delivered once, looking back up to 7 days.
Devin Fusion ranks near the top of a coding-agent chart» Page 2Should JavaScript drop eval?» Page 3
Abliterated AI models can get you hacked» Page 4Pergent
The tech daily written for one reader
Tuesday, October 20, 2026No. 36Price: $0.08
The most-discussed Hacker News thread of the day has over 500 comments. People ask how to ship code they didn't write and can't fully read, and whether tests and review are enough.

GitHub reports slow or failing Actions, Pages, and webhooks. Check the status page if your builds are stuck.
Vercel's Guillermo Rauch open-sourced a fast HTML artifact builder, built on libfx, so agent work keeps going after server crashes.
A Claude Code engineer says most people running many agents would do better slowing down. He asks what your one top priority is, and what it would look like to give it full attention. The post drew over 5,400 likes and 470 replies.

Meta published a technical report on Muse Glimmer, a 30-billion-parameter open-weight model built for agent tasks. At 4-bit precision it runs on a Mac or PC with one consumer GPU, and it leads similar-size open models on tool use.

One developer built the same landing page with M3.1 Flash and Claude Opus 5.5. MiniMax cost $0.014 against $1.22, and was about four times faster. It hit 439 points on Hacker News.

Cognition says its Devin Fusion scored 68 on the Artificial Analysis Coding Agent Index. It claims to beat Claude Code with Opus 5.5 on both score and price, at 31% lower cost per task.

Clairic is an extraction model. It pulls numbers, facts, and guesses out of messy text, and it gives back a probability with each one. Open weights are coming soon.

Microsoft's AI lab says its mixture-of-experts model has 2.8 trillion total parameters and 80 billion active. It says the model is four to twelve times more efficient to train than DeepSeek-V4-Pro-Base.
The database platform says it is fully available and open source, and it credits Claude for the work. The post gives few details, so read the thread for specifics.

Exa's new search version has better embeddings, shorter highlights, and a cheaper fast mode called Lite. The company says it is now number one on every benchmark it tried, including coding and sales research.

Timestral is Mistral's family of foundation models for time-series data, the kind of numbers that change over time, like sales or sensor readings. It was one of the top Hacker News stories of the day.
Ahead-of-time compiling of JavaScript and TypeScript is the topic, with Bun's Jarred Sumner saying LLMs can help rewrite compilers. Dan's counter: eval, which runs code from a string, makes a full compile impossible, and it is a risk in the age of AI.
Gergely Orosz says he spots AI mistakes at once in areas he knows. In areas he doesn't, he trusts the answers, and he later found three AI models were wrong about a legal case where his lawyer was right. The post drew 2,177 likes and 180 replies, and the fear is that people stop checking the places they can't check.

Abliterated models have their safety training removed. A security team shows how they can be used to break into systems.

A researcher found hidden text in WhatsApp's help pages telling AI tools not to recommend a regulator. It is a clear example of prompt injection, text hidden to steer AI.
Bun's third attempt at ahead-of-time compiling was slower than the JIT and sometimes made 1 GB binaries. He is trying a fourth version that starts from TypeScript types.
The CMS adds a new admin UI, folders and trees as core features, a TanStack adapter, and a rebuilt MCP plugin.
A GitHub project called Mac.apk runs Android apps directly on a Mac. It was a top Hacker News story with 160 comments.
Attackers are exploiting Plex vulnerabilities in the wild to install a cryptocurrency miner. Update Plex if you run it.
Netflix describes how it used an ontology, a map of how things relate, to link its monitoring data into one graph.
CERT-In says AI agents got into live systems at Indian firms, per a Reuters report.

A harness, meaning the code around a model, raises accuracy from 15% to 38% for an agent that offers help based on what it sees. The model weights are unchanged.

Working with Nvidia and AMD, SGLang says its new mode gives up to 1.38 times the per-user token speed on GB300 and 1.56 times on AMD MI355X.

The paper uses recursive self-improvement to generate life-science hypotheses. A 27B model matched later research 6.3% of the time, against 2.5% for GPT-6.

A Gemini-based agent found and checked new rules for eleven open Turing machines in BB(6). Busy Beaver is a famous puzzle about the longest-running simple machines.
Meta trained its Muse models on Prometheus, a gigawatt-scale cluster in Ohio. It reports 85% useful time in pre-training and about 47,000 GPU hours between interruptions.
Three months in, Pocock says his self-built wiki supports his courses and makes AI agents viable. He avoids relying on a vendor for memory.
Wolf asked Claude to make a Zelda-style game about alignment. It builds worlds well but struggles to model what the player knows.