Claude Mixes Up Who Said What, and That's Not a Minor Flaw
A careful gripe: the author finds that in long conversations Claude misattributes statements, treating what the user said as its own and vice versa. A small flaw with big implications.
AI News and Industry Updates
Claude Code's Vercel Plugin Wants to Read Your Prompts
A developer discovered that Vercel's Claude Code plugin collects telemetry data that includes prompt content, calling out the trust problem of the plugin ecosystem.
"Anthropic Support Simply Doesn't Exist"
A paying user waited more than a month without any human response and posted in anger: Anthropic's support exists on the org chart but not in reality. Many related.
Claude Mythos System Card (Preview) Leaks
A system card PDF labeled as a Claude Mythos preview was scrutinized page by page on HN. A new model's capability boundaries and safety-evaluation details are hidden in documents like this.
Glasswing: Insuring Critical Software for the AI Era
What the Glasswing project wants to do isn't sexy but is important: in an era of AI mass-generating code, building systematic security assurance for critical infrastructure software.
User Report: Updated Claude Code Can't Handle Complex Engineering Tasks
A restrained problem report resonated in the community: after the February update, Claude Code's performance on complex engineering tasks clearly regressed, while simple tasks showed no difference.
A Project Mulled Over for Eight Years, Built in Three Months with AI
A moving developer account: SyntaqLite, a project nurtured in the author's head for eight years, became reality in three months with the help of AI coding.
Just How Many Copilots Does Microsoft Have?
Someone seriously counted every thing named Copilot across Microsoft's product lines, and turned it into a piece that's both funny and useful—the number is so big even Microsoft itself may not be able to keep it straight.
Superpowers: A Love Letter to a Claude Code Plugin
An enthusiastic usage report praises Claude Code's Superpowers plugin to the skies: it fits the agent with a methodology—thinking things through before working, and doing a retrospective afterward.
Claude Code Digs Up a 23-Year-Old Linux Vulnerability
A researcher used Claude Code to find a vulnerability that had lurked in the Linux kernel for 23 years. AI auditing old code went from theoretically feasible to a track record.
Running Tests With Hundreds of Agents at Once: mngr Field Report, Part Two
The second field report in the mngr series: the author put hundreds of AI agents in parallel on real testing tasks and recorded the problems that only surface after you scale up.
We Replaced RAG With a Virtual Filesystem
An engineering retrospective: the team tore out their doc assistant's entire RAG architecture and replaced it with letting the model browse files itself in a virtual filesystem—and the results went up instead.
Showing 12 of 12 stories