Serial medical data startup founder. Posts on Rust, Svelte, GeoData, DB/Analytics, ML/AI
Joined April 2007
- Tweets 4,212
- Following 763
- Followers 799
- Likes 2,241
237 Photos and videos
Jun 4
Beehiiv really needs a setting to opt out from anyone else adding you to a newsletter. I’m getting subscribed to Beehiiv newsletters I don’t want like twice a week now
1
36
Jun 3
One cool thing about jj revset aliases is that you can use them in all sorts of places. In my diffview.nvim fork that supports jj now I can do `DiffviewOpen stack_base` and quickly get a diff of the current stacked PR
stack_base = "latest(heads(bookmarks() & ancestors(@--)), 1)"
39
Jun 2
So apparently if you update a kids iPad and it needs updated T&C acceptance to use iCloud then the screen time controls just turn off until you accept the new terms.
40
Jun 1
Added this note about AI-diagnosed issues to the issue template on the svelte-maplibre repo
35
May 28
Wow, running my PR review prompts with Opus 4.8 and... it's really bad.
- It is pedantic about unused imports
- It has a serious hallucination problem. Finding issues in zod schemas that don't event exist.
I never saw this with 4.6 and 4.7.
2
3
188
May 28
Looks like maybe a harness issue, I got an odd message about "file read tooling being blocked" mixed in with the session logs.
1
44
May 28
I don't know, very confusing. Everything seemed to work fine, and the intermediate output from the subagents/workflow looks right. It's just the final output that was messed up.
15
May 21
tried github.com/mattpocock/skills… on a state machine definition and I'm sold. A quick interactive walkthrough made it very very clear when I was pressing a button to advance through a state that just doesn't need to exist.
45
May 13
Really happy to see Svelte support in oxfmt. Just switched over our work monorepo with about 1.3 million lines of code, and it's between 10x and 25x faster per package to format.
4
90
May 13
Claude Code getting much less useful in June
May 13
Starting June 15, paid Claude plans can claim a dedicated monthly credit for programmatic usage.
The credit covers usage of:
- Claude Agent SDK
- claude -p
- Claude Code GitHub Actions
- Third-party apps built on the Agent SDK
1
1
76
May 6
AWS recommends enabling MFA delete on S3 buckets and also makes it a huge pain to actually turn it on...
1
1
171
May 2
Any illusions about Claude’s superior apprehension of time are broken when you come back to the computer after 3 hours and find Claude has spent the entire time waiting for a clearly stalled test process.
61
Apr 24
When @grok rolled out periodic tasks, one of the samples was a daily news task. It started out useful and then at some point it veered into fiction. Today was Cerebras's IPO popping with the price up 150% on first day of trading... But the IPO hasn't happened yet.
Also room temp superconductors, GPT-6 running on quantum computers, etc.
So far I haven't been successful in changing the prompt to avoid this behavior.
1
34
Apr 21
So when generating PR review guide style output, it works much better to generate the diffs yourself, store them in a file, and then tell the LLM to just insert something like `<diff ref=1>` instead of having it insert the whole diff and inevitably screw it up or omit lines.
43
Apr 18
When I see things like Anthropic changing the default effort to xhigh and thereby chewing through everyone’s usage quota I’m reminded of this tweet
6 Oct 2025
needless to say: dogfooding the hobby tier so that we never slip and make it a great experience. easy to let it erode over time
57
Apr 18
Ran into an infinite loop in @pierrecomputer Diffs (wonderful library btw you should check it out).
Gave Codex nothing but the stack trace and and the code and it figured out exactly what the problem was.
56
Apr 14
Tell Claude to generate a walkthrough of a PR and then followup prompt to find issues, and the issues are mostly nitpicks.
Fresh session that just finds issues, and it does a proper review.
It's really interesting how previous context can anchor the model to what's normal.
1
103