Local-first & uncensored AI

Most "AI strategy" today means renting someone else's model, on someone else's hardware, under someone else's rules. It's fast to start and quietly expensive to live with — in three currencies nobody puts in the demo.

Data leakage. Send your contracts, customer records, or trade secrets to a hosted API and you've copied them into someone else's training corpus — explicitly or effectively. For enterprises in regulated or competitive industries, that isn't a risk you wave away.

Cost at scale. Per-token pricing is fine for a prototype and brutal for a product. The moment AI touches millions of calls a month, the cloud bill becomes the project's ceiling. On your own hardware, capacity scales with capital you already bought, not with a meter that never stops.

The filter. Hosted models answer within a platform's tolerance. Ask the wrong question, in the wrong frame, and you hit a guardrail that has nothing to do with your business and everything to do with their PR team. For legitimate, sensitive, or novel work, that's a straightjacket.

Uncensored doesn't mean unsafe. It means you set the policy — not a vendor's trust-and-safety team.

Run open-weight models on hardware you control and three things change at once:

This is where the work gets interesting. Open models have nearly caught up to the frontier for most business tasks, and the gap is collapsing monthly. For the work most organisations actually do — drafting, summarising, classifying, searching their own documents — local models are already more than good enough. The privacy and cost wins are real today; the capability gap is no longer the blocker it was a year ago.

I won't sell this as a free lunch. Local-first means you own the ops: provisioning, updates, evaluation, uptime. The frontier hosted models are still ahead on raw capability for some tasks. And you need the engineering discipline to serve, monitor, and update models — it's not a managed API you can forget about.

The right architecture is usually both: local for the sensitive, high-volume, domain-specific work; cloud for the occasional frontier lift. But the default should flip. Start from "what can we own?" and reach for the cloud only when there's a real reason. Most teams do the reverse, and quietly inherit the three costs above.

A typical engagement I run:

1. On-prem model serving — an Ollama-style stack on your hardware, behind your network, with no call leaving the building.
2. RAG over your own documents — retrieval grounded in your private corpus, so the model answers from your data, not its training set.
3. Evaluation & guardrails — test harnesses that measure output quality before and after every change, plus input/output controls tuned to your policy.
4. No per-token bill — capability that doesn't phone home with your trade secrets or charge you to think.

The outcome is AI you can iterate on without a quota, on data you'd never upload, asking questions you'd never get answered. In an era where every platform is tightening its guardrails, the most capable, most trustworthy AI is the one you control.

Build it on your hardware. Set your own rules. Keep your judgement — and your data — on your side of the firewall.