Deep Dive · XiaoHu Explains

Former OpenAI CTO Mira Murati: AI Should Amplify Humans, Not Replace Them

From tacit knowledge to interaction bandwidth to model alignment — the case for why AI development can't do without human participation.

TL;DR
  • Thinking Machines Lab published a mission blog proposing an AI development path centered on "extending human will and judgment," positioned as a counterpoint to the industry's "autonomous intelligence" path.
  • The company lists three technical investments: training frontier models, building Tinker — a tool that lets users fine-tune model weights themselves — and researching "interactive models" that support real-time, multimodal interaction.
  • Using chess and mathematical proofs as examples, the piece argues that every domain where AI can autonomously surpass humans meets two conditions: the goal is static and clearly expressible, and there's no hidden tacit knowledge involved — conditions most real-world work fails to meet.
  • The article criticizes current evaluation systems — such as METR's "autonomous task duration" — for measuring only what AI can do on its own, not how well humans and AI perform together.
  • The article makes the case for "decentralized alignment," pointing out that labs commonly use their previous flagship model to generate training data and reward signals for the next one — a practice that homogenizes model "personality" across generations.
⚑ This piece is compiled from Thinking Machines Lab's official mission blog. All claims, arguments, and descriptions of technical direction are the company's own statements — not independent third-party evaluation or verification.
1Company & Mission

What Kind of AI Is This Company Trying to Build

Thinking Machines Lab (an AI lab founded by former OpenAI CTO Mira Murati) has published a mission blog stating the company's mission as "AI should extend human will and judgment," and laying out three technical paths to get there.

The blog sets a direction: use AI to amplify human will and judgment, then breaks that direction down into three concrete technical investments.

◆
What makes this piece notable is that, in the form of a company mission, it openly proposes a technical direction distinct from the industry's mainstream "autonomous intelligence" path, and publicly lists three concrete investments: training strong models, building the fine-tuning tool Tinker, and researching interactive models.
Train Strong Models

Keep pushing model capabilities — multimodal interaction, customizability — toward the frontier, since human judgment can only take effect through a strong model.

Build Tinker

A toolkit that lets users fine-tune model weights themselves, reshaping AI into a version built for their own needs.

Research Interactive Models

Give models native support for real-time, multimodal human-AI interaction, training this capability directly into the model itself.

The company also says it will make the related research and practical handbooks public to the scientific community, so more people can gain access to this fine-tuning capability.

2Core Argument

AI Can Win Solo at Chess and Math Proofs; Most Jobs, It Can't

AI serves the knowledge that people rely on to get things done: how to do something well, what's even worth doing. This kind of knowledge is generated continuously by the people doing the work. A chef refining a new dish, a shopkeeper adjusting how goods are arranged and priced on a shelf — both are chasing a complex, outsider-illegible set of goals, using their own tricks of the trade. Those tricks keep updating with real-world feedback; they can't be organized into a static list you file into a database, and they're local — change the restaurant, change the shop, and both the goal and the method change with it. The full knowledge of any chef or shopkeeper lives scattered across each individual chef's and shopkeeper's own hands.