For Thinking Machines, from 🤫 hussh
How we would start.
A bench and a prototype in the first month. Then one pilot, then building together, then the market. Trust first, always.
The order
Bench, pilot, build, then sell and buy together.
- First 30 days
Stand up the bench
- A provider-neutral bench inside 🤫 Agent One over MCP, with Gemini, Claude Code, Codex and Grok side by side.
- A teach-your-own-agent prototype: fine-tune an open-weight model, Tinker-style, on a 🤫 Puppy One developer unit. Private, on owned compute.
- Months 1 to 3
One small, real pilot
- One consented use case where a person teaches their own private model on compute they own.
- Reproducible and receipted end to end: deterministic inference, every access logged, nothing leaves without consent.
- Months 3 to 9
Build together
- Efficient, reliable fine-tuning inside the 🤫 owned-compute stack, on open rails (PCHP).
- Real-time interaction models for the always-on, voice-first 🤫 Agent One.
- Publishing in the open, in the spirit of Connectionism.
- Month 9 and on
Sell together, buy together
- A shared go-to-market on open, inspectable rails.
- Buy compute together where it lowers cost per watt and per workload for everyone.
- Grow the 🤫 Factory One grid so more people own the means of their own intelligence.
The first step
Thirty minutes to compare notes.
We would love thirty minutes between Mira Murati and Manish Sainani to compare notes on a future we describe almost the same way and build from two complementary ends: the science of teaching your own model, and the place that model should live.
We can start on our side today. It would be far more fun to build it together.
Own your AI. Own your data. Own your compute.