8 articles
Local AI is real and I run it daily — but when a frontier model offered to optimise my llama-server script, invented its own benchmark, and found 15%, I was reminded why the big models are not redundant. Then it reminded me of something else.
My local AI setup AKA My journey of making a frankly unreasonable number of experiments in making 8GB of VRAM behave like more.
Local AI is not one magic download. It is a stack of models, inference engines, clients, prompts, hardware limits, and occasionally MCP plumbing.
When companies turned AI token spend into a performance metric, engineers started burning tokens to hit quotas
Dropbox's Nova, LinkedIn's MCP tooling, and GitHub's Copilot app show coding agents becoming a fleet that needs orchestration, sandboxing, and context plumbing.
Sparse models, MoE, quantisation, and better local runtimes have changed what is possible on modest hardware. Local AI is no longer only for people running dual RTX 3090 rigs — but it is still very much for developers who can supervise the machine when it starts confidently sawing through the floorboards.
Uber's $1,500-per-tool monthly cap on Claude Code isn't about frugality — it's the first public admission that no one can measure agentic coding's return.
A sabotaged jqwik release and a critical Starlette flaw expose one blind spot: coding agents run third-party code under a threat model nobody designed for.
