Update — July 16, 2026: Thinking Machines shipped Inkling — 975B open-weights MoE with Tinker fine-tuning and a self-finetuning OpenCode demo — the first model release behind this manifesto.
Update — August 22, 2026: A garbled "Thinky Machines" headline trending on aggregators is this same lab — see Inkling's free OpenRouter endpoint explained for the identity confirmation and what "free" actually means in the terms.
On July 10, 2026, Thinking Machines Lab published The Future Worth Building Is Human — a full manifesto from the lab Mira Murati and John Schulman built after leaving OpenAI, backed by a reported $2 billion seed round.
The thesis in one line: AI should extend human will and judgment — not freeze a snapshot of organizational knowledge and replace the people who generate it.
Same week as GPT-5.6 Sol rollout and OpenAI's $50K bio jailbreak bounty, Thinking Machines is arguing the industry's autonomy race measures the wrong thing.
TL;DR — manifesto at a glance
| Theme | Thinking Machines position |
|---|---|
| Mission | Extend human will and judgment |
| Anti-pattern | Frozen models trained in few places, not shaped by users |
| Pro-pattern | Distributed AI as diverse as the people it serves |
| Tools | Tinker — train open-weights with your knowledge |
| Interfaces | Interaction models — live multimodal collaboration in the model |
| Research | Connectionism publications — open science |
| Alignment | Decentralized — many AIs, not one spec |
| Benchmarks | Skeptical of METR-only autonomy horizons |
Bringing intelligence to knowledge
The essay opens with tacit, local knowledge — Polanyi and Hayek cited explicitly:
- A chef's recipe sense or shopkeeper's display judgment is not a static database row
- Central planning fails not from weak intelligence but from knowledge dispersion
- Chess and math are exceptions — static goals, no hidden local knowledge
- Real work (restaurants, shops, organizations) needs AI that cultivates knowledge through ongoing collaboration
Toyota callback: Mitsuru Kawai brought master craftsmen back to automated lines — "To be the master of the machine, you have to have the knowledge and the skills to teach the machine."
explainx.ai read: This is the philosophical counterweight to agents' last exam and METR time-horizon charts — those measure solo agent hours; Thinking Machines wants joint human-machine outcomes as the optimization target.
Four technical directions
From the manifesto:
- Train strong models — multimodal interaction, customizability; frontier competition matters so human judgment can shape sharp instruments
- Build tools — including training model weights (Tinker)
- Develop interfaces — broaden the human–machine channel beyond a text box and long wait
- Publish research — open science (Connectionism)
Murati's X thread (July 10) tied the year-one recap: interaction model previews, Tinker for anyone training open weights, Connectionism research drops.
Human participation is a technical challenge
Key claims:
| Problem | Thinking Machines answer |
|---|---|
| Narrow channel | Text box + latency can't carry rich intent → bet on native multimodal interaction in the model |
| Wrong metric | METR task-completion horizons track solo autonomy, not collaborative yield |
| Incentive mismatch | Single rented model benefits from absorbing user distinctiveness; customized AI benefits when customers own specialized advantages |
Parallel on explainx.ai: Fable advisor + Sonnet executor and Matt Shumer's goal/bar loops are practitioner versions of "don't optimize only for solo agent hours."
Decentralized alignment
The manifesto's sharpest policy section:
- Single alignment spec → single power locus → intelligence curse dynamics if AI needs little from people
- Parent-model-on-parent-model training loops → everyone gets the same character, generation after generation
- Prompt-only values change surface behavior while deep habits remain — Gwern on personalization vs safety cited
- Von Neumann 1955 — useful and harmful lie close; safety is ongoing judgment, not one-shot spec
Vision: An ecosystem of AIs — raised in different places, disagreeing like humans — vs one averaged morality from a handful of labs.
Contrast July 2026:
| Lab | Alignment shape |
|---|---|
| OpenAI | Central GPT-5.6 + bio bounty + government preview |
| Anthropic | Classifier-gated Fable 5 |
| Thinking Machines | Own weights + local shaping + diverse ecosystem |
What shipped before the manifesto
Murati's July 10 thread referenced work already public:
- Tinker — fine-tune open-weights with organizational knowledge
- Interaction models — multimodal collaboration previews
- Connectionism — research line on how models learn
The manifesto is the why behind those products — not a sudden pivot.
Who should care
Enterprise AI leads — manifesto articulates why renting one frontier model may erode proprietary know-how.
Open-weight builders — Tinker narrative aligns with Ollama funding and sovereign AI threads.
Policy readers — decentralized alignment as alternative to export-control theater and single-lab value lock-in.
Agent engineers — critique of METR-only success metrics; design for human-in-the-loop yield, not autonomous hours alone (human-in-the-loop guide).
What to watch next
- Tinker API — pricing, model bases, enterprise adoption
- Interaction model product — ship date beyond preview
- Schulman/Murati hiring — lab headcount vs $2B burn
- Benchmark proposals — do they publish collaborative evals to rival METR?
- Open weights release — manifesto promises strong models + customization; weights timeline?
Related on explainx.ai
- Inkling open weights — July 15, 2026 release
- Inkling-Small — 12B-active open weights (Jul 30)
- GPT-5.6 vs Fable 5 — centralized frontier week
- OpenAI Bio Bug Bounty $50K
- Agents' Last Exam — professional work benchmark
- Ollama $88M — open models momentum
- Human-in-the-loop — when to let agents run
- Yann LeCun — LLMs vs physical agents
- India sovereign AI status
Official: The Future Worth Building Is Human · @thinkymachines · @miramurati
Manifesto themes and product references reflect Thinking Machines Lab's July 10, 2026 publication. Funding figures and product timelines cited from public reporting and lab posts — verify on thinkingmachines.ai.