July 2026 1 note

ideas updated 1 min read

5 ways to improve your folk software

AI summary

Doug suggests that individuals can achieve notable results with folk software by utilising five key approaches, including creating specific skills and using tools like batching and test-driven development. These methods enable users to build effective systems with Large Language Models (LLMs). By implementing these strategies, users can improve their overall folk software experience and create more efficient systems.

…

June 2026 3 notes

notes updated 2 min read

So apparently I should be using ChatGPT

AI summary

The author notes that despite moral objections to OpenAI, ChatGPT aligns most closely with their left-libertarian political values based on a chart of frontier LLM biases. The note also highlights concerns about increasing government control over LLMs and their potential to influence user decisions and political views.

…
notes updated 2 min read

Can LLMs invent a "private language"?

AI summary

Anthropic observed Claude Fable 5 and Mythos 5 models sometimes using invented jargon, unusual punctuation, and emojis during long reinforcement-learning rollouts, then switching back to normal language when responding to humans. The note argues this is more like compression than a true private language, referencing Wittgenstein's argument against private languages for humans and cautioning that conspiracy theorists may misinterpret the behavior.

…
notes updated 2 min read

Tacit knowlede and domain expertise in a world of LLMs

AI summary

LLMs require tacit knowledge and domain expertise to produce useful results, as domain experts can verify correctness even if they can't code, while generalist engineers can build but lack the "ground truth" to judge domain-specific outputs. The binding constraint has shifted from building to verifying what is correct.

…