Coding Model Switching Is Cheap
2026-07-04
AI coding
coding models
developer tools
cost
Loading blurb...
2026
2026-07-04
Loading blurb...
2026-07-04
I just posted that the cost of changing coding model providers is next to nothing. That is only true if you're small and nimble. Navigating IT bureaucracy makes switching hugely costly.
But, if you're noticing that one intrapreneurial team outperforming everyone else on speed and costs, I bet you anything that their shadow IT/AI game is very strong.
2026-07-03
Loading blurb...
2026-06-27
I've been spending some evenings playing with Apple's CoreAI. It is yet another way to export and run models in a format that takes advantage of Apple silicon.
Apple is making it easy for any developer to integrate high performance inference on the edge. I mean if I can make a multi-modal AI-powered native MacOS app, anyone can.
2026-06-24
Loading blurb...
2026-05-27
If you want a compute and token efficient 'agentic systems' your goal should be to use your non deterministic coding tools to replace as many parts of your non deterministic system as possible with deterministic code.
Just not all of them.
Start off with something wildly inefficient then invest in evolution towards efficiency.
2026-05-22
oMLX as my local LLM server. It's faster. Mainly due to SSD KV caching. They've also got turboquant support for KV cache, but I haven't tested that deeply.
Qwen 3.6 35B A3B 8bit is my new local non coding agent brain. It's fast and smart and works well with opencode as the harness. It's also good at image understanding.
2026-05-17
“On one occasion I even saw an aerodynamic hearse, which is about as far as the aerodynamic style can go”
Bruno Munari, on forcing things into a given style.
For months the most cost effective coding set up was Codex via a ChatGPT subscription, falling back to GitHub Copilot when daily limits were reached. API calls were the last resort towards the end of the month.
The party ended June 1 for both of these options.
Now all dollars are flowing to Deepseek, Kimi, Qwen, and GLM. The cost of switching is nothing more than some testing and some getting used to.
I've been inspired to dive deeper into Apple's CoreAI APIs in MacOS 27 Beta thanks to the work of john-rocky on GitHub. Check out this article and his Core AI Model Zoo in his GitHub repositories.
The biggest threat to Crypto Currency is not Quantum Computing, it's large scale edge-device hacking powered by AI. Crypto requires safe 'rails' for transaction. Wallets need a safe place outside of cold storage to exist without being drained. I don't know where that's going to be when virtually every system is compromised.
I've got no skin in this game. It's neither here nor there for me, but a crypto-techno-utopia seems further away than ever.