DeepSeek was a side project at High-Flyer Quant. Qwen was a side project at Alibaba. Twitter was a side project at Odeo. Mac was a side project at Apple. Meanwhile: Windows Phone was a core project at Microsoft. Metaverse was a core project at Facebook. Google Glass was a core

December saw my highest Cursor usage so far, with ~530M tokens in the past two weeks. Yet it still feels like I'm barely scratching the surface. Tokens should flow 24/7. I am the only bottleneck, my standards, my taste, my agency.

Found the perfect spot to bleed out

I often think about this but for software. The meta problem is that traditional anti-entropy tools themselves decay. Documentation, tests, configs, runbooks go stale. AI can help with maintenance, but we need to invest in harnesses that let AI observe, interpret, and act on our

Thanks Cursor Ben Lang for such a cool gift! Time to put my cat to work :)

"...the reality is that in production, 89.6% of Tusk Drift requests never reach similarity scoring. They [only go through] input value hash matching in the priority cascade before the best mock is chosen for replay testing." jytan writes about why similarity scoring is

Happy to see this grow to ~100 stars! Added SSH support and key fixes the past week, supports Claude/OpenCode/Cursor Agent/Gemini. Give it a shot and let your agents go yolo (safely)

One of the greatest joys in life.

Just got burned by Cursor's unsandboxed agent mode. I instructed the agent to use the github MCP to read an issue and investigate it, but it went ahead and use the gh CLI to push a commit straight to main, created a response for the issue and closed it. But it's so annoying

This is my experience as well. Codex 5.3 is less trigger-happy and does a better job judging when to stop and ask and push back vs go ahead and implement - so it seems a lot more trustworthy. Definitely appreciate being able to prompt less defensively.