justin․searls․co

I joined Twitter in 2007 and my brain slowly morphed over the next 15 years from hopelessly verbose to nihilistically pithy. I've kicked the Twitter habit, but the takes keep flowing. That's why I post them here and format them as a social network of one. You're welcome to bookmark any of these takes, though I'm not sure why you would.

By the way, the hearts and like counts are fake. They're just there to make you feel safe.


Pro-tip: the codex CLI can't search the web by default (even if you bypass all sandbox restrictions). You need to explicitly enable --search.

If you ask codex to search the web without that flag, it'll literally guess domain names and try curling their homepages.

I wish coding agents came with those Green/Red coasters they give you at Brazilian steakhouses:

🟢 Green: go ahead and pile stuff on my plate
🔴 Red: stop adding, we need to make room first

I don't know who needs to hear this, but despite being bare bones from a feature-set perspective, Codex CLI with GPT-5 is much, much better at some coding ecosystems than Claude Code with Opus 4.1/Sonnet.

Codex writes competent Swift that does what I ask, nothing more. Claude hallucinates code all day.

With Swift, I'm really speed-running the list of stupid things you do when learning a new language. 3 days ago I wrote a dependency injection framework, 2 days ago I convinced myself I'd found a compiler bug, yesterday I wrote my first macro, today I made a mocking library.

The nice thing about server-side LLMs hitting the point of diminishing returns is that it gives local LLMs a chance to catch up and for their utility to approach parity.

I would pay so much extra for a version of Claude or ChatGPT that paid the same toll I do whenever I fuck up. Make guilt a stateful property that decays over weeks or months. Trigger simulated self-doubt when similar topics arise. Grant my account bonus GPU-time so the chatbot works ridiculous overtime to make up for its mistakes, just like I would for my boss.

Despite not touching it for several years, I've noticed a marked uptick in KameSame adoption in recent months. I asked a few new users and, like a lot of my stuff, it turns out ChatGPT is driving far more people to it than Google ever did.

Keep hearing about Finntech and how much money people are making, but never hear anything about tech startups in the other Nordic countries. Does Norway not have as many programmers?

Interesting analysis of the distinctiveness of the Japanese Web. The biggest cause in my mind has always been bottleneck effect. Japan's Web developed and remains more isolated than any other "free" nation.

If every non-Japanese website disappeared tomorrow, many Japanese would go literal months without noticing. THAT's why its web is different. sabrinas.space

sabrinas.space -

I don't wish them ill, but the stock price of DuoLingo (and that entire class of language learning apps) hasn't made a lick of sense since ChatGPT released. It's just going to take a single LLM-based product to obviate the entire business model yro.slashdot.org/story/25/08/17/194212/duolingos-stock-down-38-plummets-after-openais-gpt-5-language-app-building-demo

Duolingo's Stock Down 38%, Plummets After OpenAI's GPT-5 Language App-Building Demo - Slashdot

A group of Italian-American feminists should buy an island off the Amalfi coast to establish a women-only community and call it Old Country for No Men.

You know that meme where the best developers actually wind up deleting more lines of code than they add?

The more time I spend wrangling agentic codegen tools, the more the task feels like chiseling than sculpting. I suspect the deleters are better poised for this moment.