David Foster

Software sculptor. Writer. Teacher. Inquiring mind. Bringing world-class computer science education to K-12 schools, teachers, and students at https://techsmart.codes

Website: https://dafoster.net/
GitHub: https://github.com/davidfstr/

Last Notes

2026-06-05 16:25:52 GMT

AI generated code is not copyrightable under current US law, as I was recently reminded by the OpenSSL CLA. Thus presumably a properly attributed AI generated commit (via Co-Authored-By) is excluded from copyright protection. Interesting. #aicoding

2026-05-25 11:16:04 GMT

- reply

Thanks for the specs! Yes mine is M3 (Apple Silicon), 16 GB unified RAM, 24 GB total RAM. 100 GB/sec mem bandwidth. I just got Qwen3 Coder 30B A3B Instruct (UD-Q2_K_XL quant, 40k context) working in a basic fashion at 35 tokens/sec output. Using Pi harness to drive because it only adds 1k system prompt vs Claude Code at 31k.
I'm encouraged to see I'm converging to a similar model setup as yours.

2026-05-25 00:59:05 GMT

- reply

Do you use local AI models for coding? -- I'm running a MacBook Air with 16 GB GPU memory (24 GB total memory).

2026-05-23 15:54:21 GMT

I underestimated how difficult it would be to setup a local LLM (AI model) for coding assistance. Must consider: Model size, obscure hardware specs (like memory bandwidth), quantization choices, hardware optimization support, per-tool (non-)support for different model formats. What a mess. #aicoding

2026-04-18 01:19:21 GMT

I wrote up how I've been attributing usage of AI models & coding agent harnesses when drafting code commits: https://dafoster.net/articles/2026/04/17/ai-attribution-in-git/ #aicoding #git

2024-08-06 12:42:42 GMT

#Python developers! We’re gathering insights on types in Python and need your help! Your input will help us improve the Type System and IDE tooling. -- Please take a moment to fill out our quick survey and share with other developers: https://forms.gle/XdSNR42LUuj58ghdA . Thank you!