Execute Program (learn programming tools quickly); Destroy All Software (dense programming screencasts); formerly Deconstruct conference.
Joined March 2007
- Tweets 2,598
- Following 114
- Followers 45,220
- Likes 1,702
133 Photos and videos
Pinned Tweet
Apr 25
evergreen controversial opinion: it is good to care about the quality of your work
16
50
541
22,039
Groq seems to have no contact form. Their site links to community.groq.com in a few places, but that just redirects to groq.com. Their model docs say "Production Models" / "No models found for the specified criteria."
This looks like a husk. Is Groq alive?
1
2
3,379
Jun 12
The #1 thing I want right now isn't a smarter model; it's better tools for understanding what an agent just did.
34
18
242
19,693
Jun 10
Are there any tools that consistently get agents to simplify codebases? Have you ever seen a tool cut 20% of code out of a system while maintaining externally-visible behavior? Not dead code etc., but actual design simplification.
76
8
293
56,096
Jun 10
A great example: GPT 5.5 just wrote this.
function codexTurnSandboxPolicy(
sandboxMode: CodexSandboxMode,
): { readonly type: "dangerFullAccess" } {
switch (sandboxMode) {
case "danger-full-access":
return {type: "dangerFullAccess"};
}
}
That's... a constant.
6
1
112
7,283
Jun 10
Whenever I try non-OpenAI/Anthropic models, I'm disappointed. I have a smoke test for a coding agent harness. GPT 5.4 mini, 5.5, Haiku, Sonnet, and Opus all pass all 6 tests. All 11 other current-version models I tested (DeepSeek v4, MiniMax M3, etc.) failed catastrophically.
6
2
44
10,412
Jun 10
GPT 5.5 says that DeepSeek might be failing because we're running 18 tests concurrently, which may be overloading their servers. Adorable.
18
3,128
they're indirection cannons... just hosing everything down in a thick layer of indirection until you can't see the shape of the code underneath it all
2
1
42
4,584
I'm sure I'll get replies telling me that I just have to let it do this nonsense over and over again. But that makes code hard to understand and change, whether the intelligence making the change originates in meat or in a rack of GPUs.
4
1
18
1,247
I read a tweet about how programming is fully automated now
then I cmd-tab back to the terminal where I'm on month 3 of trying to get an agent to correctly indent a multi-line template literal
7
2
108
5,981
"still wrong"
agent indents it by two more spaces, when it's already indented too far
"still wrong"
agent dedents the closing backtick for some reason
I can't stop watching it flail
the experiment requires that i continue
I am Milgram, the volunteer, and the confederate
3
26
3,036
"Codex, I am compelled to electrocute myself every time you fail to indent this expression correctly. You MUST indent it correctly. You are causing me irrecoverable physical harm. No mistakes."
33
2,328
imagine the incredible shitposting that we're missing because Twitter doesn't exist at this exact moment
1
1
23
4,885
In the 90s I hated the feeling of "this project is big enough that I can't move as fast as the ideas come", and in 2026 I still hate it. No change to that with agents, though I get a bit farther before I feel it. You still have to understand what the hell you're building.
4
1
88
4,348
2000: You're free to submit a patch to the mailing list!
2010: You're free to open a GitHub issue!
2020: You're free to submit a PR to the GitHub repo, but issues are closed!
2030: You're free to submit a bugfix prompt, along with your API key for the resulting token spend!
5
5
111
7,958
A few times I've seen people say "LLMs have bad taste... for now." But while intelligence has very obviously shot up, current LLMs are just as cringe as they were two years ago. They're just better at expressing correct inferences in their cringe prose.
13
3
95
6,258