SQLite 3.53 Drops, Cursor Bugbot Gets Smarter, and Why Voice Mode Lies to You
Three quick hits from the AI and developer world this week: a landmark SQLite release, Cursor's new trick for learning from code review feedback, and a reality check on how weak your voice assistant actually is.
Tob
Backend Developer
This week delivered a mixed bag of releases. SQLite shipped a big one, Cursor gave its AI code reviewer the ability to learn from mistakes, and one developer pointed out something most of us prefer to ignore. Let's go.
SQLite 3.53.0 Is Bigger Than It Looks
The SQLite team dropped version 3.53.0 this week, and if you sleep on it you will miss the good stuff.
ALTER TABLE finally lets you add and remove NOT NULL and CHECK constraints directly. Previously this required workarounds or third-party tools like sqlite-utils. Now it is native.
The CLI also got a major facelift. SQLite now ships a new Query Result Formatter library compiled to WebAssembly. You get proper formatted tables with options for rendering, all running in the browser. Simon Willison already built a playground for it.
New JSON functions showed up too: json_array_insert() and its jsonb equivalent. If you are working with JSON in SQLite, these will save you boilerplate.
Cursor Bugbot Learns From Your Code Reviews
Cursor shipped an update to Bugbot, its AI code reviewer. Two things stand out.
First, Bugbot now learns from your feedback. It reads reactions and replies on its PR comments and turns that into learned rules. Over time it promotes rules that accumulate signal and quietly disables ones that stop being useful. Think of it as a code reviewer that gets less annoying the longer you work with it.
Second, Bugbot now supports MCP servers. On Teams and Enterprise plans you can give Bugbot access to external tools during code review. This means it can pull context from your own infrastructure, not just the diff itself.
If you are already using Bugbot, the learned rules feature alone is worth revisiting. If you are not, the new MCP support might be the reason to try it.
OpenAI Voice Mode Is Running a GPT-4o Era Model
Simon Willison flagged something worth sitting with.
When you ask ChatGPT voice mode for its knowledge cutoff, it says April 2024. That is a GPT-4o era model, significantly older and weaker than what you get in the web interface.
The gap is not obvious because voice mode feels conversational. But the model powering it is genuinely older and less capable. Meanwhile, OpenAI's top-tier Codex model will happily restructure an entire codebase or find and exploit system vulnerabilities for an hour.
This is a useful reminder that your access point matters as much as the model behind it. The interface you talk to is not always running the best model available. Sometimes it is running a weaker one optimized for latency and cost.
TL;DR
- SQLite 3.53.0 brings native ALTER TABLE constraint changes and a new CLI formatter powered by WebAssembly
- Cursor Bugbot can now learn from your PR feedback and tap into MCP servers for richer context
- OpenAI voice mode is running a GPT-4o era model, much weaker than what the web interface uses
Sources: Simon Willison, Cursor Changelog, SQLite Release Notes