David McCabe

.me

[David McCabe @da_be19]

I built a fully autonomous AI coding agent.

In Bash. All 63,000 lines of it.

It runs 100% offline on my phone using 3B parameter models. No cloud. No subscriptions. No Python. No Node.js.

I’m not a software engineer—I’m a process control guy. Here is the story. 🧵

[David McCabe @da_be19]

@BryceWeiner In a Gödel-lian fashion, perhaps someone will start some research using OpenAI that OpenAI will steal and bring to fruition, but which ultimately undoes the company.

[David McCabe @da_be19]

@BryceWeiner Yes, I agree with this. If nothing else, I guess the Western labs at the least aren't having to pay for the development and release of the open weight models, at the cost of continuing to lose out on the goodwill and soft influence that could be gained from it. It's frustrating.

[David McCabe @da_be19]

@BryceWeiner My point being, the model performance and data center policies seem to currently be the after thought in light of that. At least until one of the two economies blinks, that is.

[David McCabe @da_be19]

@BryceWeiner The technology is revolutionary, for sure, and likely deserves massive investment to develop it, but it's become the most palatable venue for the economic hot war between China and the US with the advantage of developing the generative tech alongside it.

[David McCabe @da_be19]

@BryceWeiner "Opinion" might be the wrong word, let's go with "analysis" instead.

[David McCabe @da_be19]

@BryceWeiner You've got some really nuanced perspective here, would be interesting to read your opinion in longer form.

[David McCabe @da_be19]

@tallyho_golf @RickyPeebs @tuckedpins Try looking into the model Udisc has developed. A successful mobile app that does what you are trying to do, but for disc golf. You might find some good ideas.

[David McCabe @da_be19]

@mitchellh Was the function signature identical between the LLMs initial implementation and its optimized version?

Was the naive function signature the same as your hand-written one?

[David McCabe @da_be19]

@ylecun @ronbrachman Perhaps the process control engineers could teach a thing or two to the ML/AI field, and vice-versa.

[David McCabe @da_be19]

@joaomdmoura I have found filesystem combined the FST5 in sqlite to be a good combination.

[David McCabe @da_be19]

@omarsar0 Here's mine, take a look, have fun. 😉

https://t.co/Q8IZhNCN0s

[David McCabe @da_be19]

It’s open source. No telemetry. Your code never leaves your device.

The "Lego blocks" are all here.

Check out the Blue Lodge: https://t.co/Q8IZhNCN0s

Thank you for the optimism, JJ! @INTELLOPY

[David McCabe @da_be19]

I built a fully autonomous AI coding agent.

In Bash. All 63,000 lines of it.

It runs 100% offline on my phone using 3B parameter models. No cloud. No subscriptions. No Python. No Node.js.

I’m not a software engineer—I’m a process control guy. Here is the story. 🧵

[David McCabe @da_be19]

The Constraints & Caveats:

• Just 4 dependencies: curl, jq, git, sqlite3.
• 3,578 test assertions across 43 modules.
• Built with heavy LLM assistance. I found the Lego blocks and mashed them together.

• Bash is a wild choice, and small models have real limits.