The framing assumes cloud-first AI agents as the default caller. But there's another path: local-first AI where the human remains the orchestrator and the model never phones home.
The "humans as tools" model only works if the AI layer is centralized and owned by platforms. If inference runs on hardware you control, you're not callable - you're the one calling.
The "local models got good, but cloud models got even better" section nails the current paradox. Simon's observation that coding agents need reliable tool calling that local models can't yet deliver is accurate - but it frames the problem purely as a capability gap.
There's a philosophical angle being missed: do we actually want our coding agents making hundreds of tool calls through someone else's infrastructure? The more capable these systems become, the more intimate access they have to our codebases, credentials, and workflows. Every token of context we send to a frontier model is data we've permanently given up control of.
I've been working on something addressing this directly - LocalGhost.ai (https://www.localghost.ai/manifesto) - hardware designed around the premise that "sovereign AI" isn't just about capability parity but about the principle that your AI should be yours. The manifesto articulates why I think this matters beyond the technical arguments.
Simon mentions his next laptop will have 128GB RAM hoping 2026 models close the gap. I'm betting we'll need purpose-built local inference hardware that treats privacy as a first-class constraint, not an afterthought. The YOLO mode section and "normalization of deviance" concerns only strengthen this case - running agents in insecure ways becomes less terrifying when "insecure" means "my local machine" rather than "the cloud plus whoever's listening."
The capability gap will close. The trust gap won't unless we build for it.
Interesting that the industry's answer to "we've made you addicted to screens" is "let us listen to you 24/7 instead." The form factor changes but the extraction model stays the same. Local inference is almost good enough that this tradeoff isn't necessary anymore - been working on that opposite thesis at localghost.ai
The article mentions IndieWeb/POSSE but discoverability remains unsolved. I'm working on a pledge system for local-first projects - a /.well-known/freehold.json that crawlers can verify. Projects that break the pledge get delisted publicly. More at localghost.ai/manifesto
I wrote this over the holidays because I couldn't shake the feeling that we have a narrow window before Apple/Google ship "local" AI that's local in marketing only.
The manifesto covers the philosophical case (cypherpunk roots, enshittification, architectural immunity), but the real urgency is in the companion piece: https://www.localghost.ai/inflection
The TL;DR: personal AI is the final extraction layer — not just what you search for, but how you think. If that data flows to central servers by default, the capture is complete.
Hardware is commoditizing (sub-$200 NPU boards by mid-2026), but the software defaults are being set now.
Full honesty: LocalGhost is a vision, not a product. No working software yet, just architecture and this site. The Freehold Directory is real though — a /.well-known/freehold.json standard for local-first projects to declare their principles and get discovered.
If someone builds this faster and better, great. The goal is that the alternative exists.
Built this over Christmas. LocalGhost is a vision, not a product —
there's a repo and this website, nothing else yet.
Thesis: hardware commoditization is nearly done, but Apple/Google
will entrench "local" AI with cloud-mandatory features. Once
convenient defaults exist, alternatives become irrelevant for
most users.
The piece covers extraction economics and why the cognitive layer
(the questions you ask an AI) is the most intimate data surface yet.
Also buried some games in the homepage terminal if you're bored.
Your thesis is likely correct on consolidation and cloud tie-in. Your pipeline model is also intriguing. I like the direction and would like to chat more about the registry or how the ghost gets interconnected.
