At "Claude Code Anonymous"

Earlier in the month I was at the first Brighton Claude Code Anonymous meet-up. Lightning talks and chat, around agentic coding (not limited to Claude).

I’m unsure of the etiquette, so I’ll name the things I found interesting rather than the people presenting them.

On running models locally, the hardware to look out for is unified memory and LPDDR5 RAM. The Radxa ROCK 5B+ single-board computer gets you 7 tokens/second on a 7bn parameter model for ~£120.

On instructing LLMs, one speaker reported success putting in the time to write up detailed instructions for a task, with diagrams and structured folders, and only then letting the agent go do the implementation. There’s less back and forth with the agent to get it right, and you get a return in terms of reduced time and token spend. It did slightly remind me of the bad old days of waterfall.

A couple of nice demos:

• The Google Chrome team have a service which uses LLMs to tell you why a particular web feature works the way it does. It digests the HTML spec and trawls the W3C mailing lists to answer a question from the social history.
• Code optimisation via pi-autoresearch: define an objective function, and allow the agent to make hypotheses about code improvements, implement them, measure, and repeat. Reminded me of GAs, but with a wide set of operations for mutating code.

I’m looking forward to the next event, as it looks like they will be monthly.