dwmkerr

In this article I try to explain how System 1 models like Jev differ from regular LLMs, and some of the emerging use-cases.

Some practical techniques - 20 mins each morning and evening with Fable running through the day: Goal setting and SOWs, exploration loops, interactive specs, talk don't type, anchor to idioms, manage the blast radius with burner environments

Less of a shameless self plug than a collection of hopefully useful patterns.

I've been wanting to do terminal recordings via LLMs for testing my projects and dropping in videos etc to PRs - this is like Playwright but for the shell. Would love feedback. Recording the LLM closing Vim was satisfying.

LLMs have been helping me code more rapidly but are instucted at the system level to often be overly helpful, making changes without discussing, adding code withotut removing stale code, trying to anticipate future needs and so on.

You can prompt your LLM or use the MCP server to get it to read this guide that instructs it to follow a 'plan / implement / review' cycle, and has some common patterns and stanards that should be near universal.

I've been using this for a few months and it's greatly improved my productivity, but would love any suggestions.

Some ideas I'd been noodling with a while back and just updated. Would be very curious for any feedback.

love it, tracking now

Tracking now https://github.com/dwmkerr/hacker-laws/issues/448

thanks for the heads up, just fixed!

on this list to go in: https://github.com/dwmkerr/hacker-laws/issues/446

just saw this, fixing at the moment

"The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin." - Richard S. Sutton (2019)

For those closer to the edge of research in AI, especially over the last five years, I'd love to know whether you believe this to be valid and whether it is still the case.

...design your system assuming that your opponents know it in detail.

The more unusual or interesting the data, the more likely they are to have been the result of an error of one kind or another.

"The best way to get the right answer on the Internet is not to ask a question, it's to post the wrong answer." - Ward Cunningham (possibly)

Thanks to those who've commented with corrections/suggestions, I'll go through them try and get the content fixed up where possible. Will need to have a lie down and possibly an epidural before going through the comments again to pull out the changes needed.

Hello, this is dwmkerr the author - your comments are all spot on - the earlier in the book you go the more work it needs, I've learnt a lot along the way TBH probably the entire first part needs a rewrite, that's actually going on at the moment as I'm working with a publisher.

I've linked to this post/comments on the repo (issue #210), they should all get addressed but it might be a while before the online version gets sorted, I'm closing out the final chapters online, once I've done this I'll likely go back to the beginning and rewrite/edit/put in a more consistent style (more code snippets, fewer screenshots etc).