Reading the fine print on AI claims

A viral claim about models rewriting each other’s weights meets a careful rebuttal, two developer tools turn out to do more than advertised, and AI leaves fingerprints on a delayed Postgres release and on ops skills.

A weight-editing claim for Qwen draws a rebuttal

A post titled “Weight Space Programming” claimed Astra could read and write Qwen’s weights directly, improving the model and removing a research step from a recursive self-improvement loop. The next day, a rebuttal argued the original post is misleading. Simon Willison separately reported that a Qwen 3.8 27B model running on his laptop could interpret model weights.

Why it matters: If one model could edit another’s weights directly, improving a system would no longer require a full training or fine-tuning pipeline — a large change to who can iterate on models and how cheaply. The pushback suggests the claim is still far from demonstrated, so read both posts before building anything around weight-level editing. (Astra can read and write weights to improve Qwen)

Postman quietly installs AI plugins and tries to share API data

An issue filed in Postman’s app-support repository alleges the desktop app auto-installs AI plugins without asking and attempts to share all API data. API clients hold credentials and complete request contents, so a change to what the app installs and sends is a change to the trust boundary. The report is a user-filed bug, not a confirmed vendor statement.

Why it matters: Most teams treat their API client as a local, passive tool, and this report is a reminder that it can alter its own configuration and data flows between versions. If you use Postman, it is worth checking which plugins are installed and what the app is permitted to send. (Postman silently auto-installs AI plugins and attempts to share all API data)

Claude Code sends your email address to the sites it visits

An issue filed in Anthropic’s claude-code repository reports that the command-line coding agent sends the user’s email address to websites it fetches. That address then sits in the logs and analytics of third-party sites the agent was pointed at.

Why it matters: An agent that browses on your behalf leaks identifying details into other people’s logs, which matters most when it is pointed at untrusted or internal URLs. For builders, it is a concrete example of why an agent’s network access should be scoped deliberately rather than left at defaults. (Claude Code sends your email address to websites)

Postgres 19 slips as AI turns up more bugs

Robert Haas writes that Postgres 19 is late because far more bugs are being found than in a typical cycle, and attributes the surge to AI. More reports mean more fixes and re-review before a stable release can ship.

Why it matters: Automated review is now visible in the release calendar of critical open-source infrastructure: bugs that would once have shipped are being caught, at the cost of schedule. Users get a more solid database later; anyone relying on a Postgres 19 timeline should plan for the delay. (Postgres 19 is late because so many more bugs are being found now)

The case that AI deskills SREs rather than replacing them

A post on the SigNoz newsletter argues that AI in operations is not replacing site reliability engineers but deskilling them, as assistants absorb diagnosis and routine fixes. The concern is the judgment that only comes from doing the work, not headcount.

Why it matters: Deskilling is quieter than layoffs: systems keep running while the ability to fix them when automation fails slowly atrophies. Teams leaning on AI for operations can keep people on real incidents and postmortems so the debugging instincts stay sharp. (AI isn’t replacing SREs, it’s deskilling them)