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.

This dispersal of knowledge is the source of the whole system's diversity, adaptability, and resilience — and it's why free markets beat planned economies. Planned economies fail for a reason rooted in the nature of productive knowledge itself: it's tacit, tied to the immediate situation, always changing, and held privately by the people doing the work. Trying to centralize that knowledge and hand it to a single centralized intelligence runs into the exact same wall.

This kind of knowledge-you-can't-quite-say has a name: "tacit knowledge" — the judgment calls people make on the job from feel, experience, and instinct, the kind that can't be written into a manual or filed into a database for someone else to use. It can only be built up and updated by a person, continuously, through the work itself.

The AI worth building should help an organization keep cultivating and updating this unique knowledge, letting it grow alongside the people who hold it.

Toyota Brought the Master Craftsmen Back to the Line

In 2014, Toyota — long known for its automation — brought veteran craftsmen back onto the production line, explicitly to cultivate craftsmanship and skill. The man who led the effort, Mitsuru Kawai, explained it this way: "To master the machines, you first need the knowledge and skill to teach them."

Producing knowledge and applying intelligence reinforce each other. The work humans do will keep changing, shifting more toward the parts only a human can bring — but the best organizations will push both to the limit. AI should help every organization become the best version of itself, on its own path, preserving the differences that already exist between organizations.

Between Humans and AI Sits a Text Box That's Too Narrow

Keeping humans continuously involved in setting goals and sharing knowledge with AI is not the same as resisting automation. Whatever a machine can reliably do on its own, it should do. But it also needs to know when to act on its own and when to check in and wait for feedback — the same way people do when working together on a team. This raises new technical requirements for how AI is designed and evaluated.

Right now, the biggest bottleneck to bringing human knowledge and judgment into AI's work is the communication channel between human and AI: a small text box, followed by a long wait. That channel is too narrow to carry the full richness of human judgment, and too slow to keep pace with continuous feedback. Human-to-human collaboration works best in real time — interrupting, correcting, glancing back, gesturing while talking, changing your mind mid-sentence.

This vision has a name: "interactive models" — training the capability for "real-time, multimodal interaction" directly into the model itself, so that as the model gets smarter, its ability to interact grows right along with it. An interface designed this way doesn't just permit human participation — it actively invites and rewards it.

The Yardstick We Use for AI Progress Only Measures Half the Picture

Another challenge is setting the right target for evaluation and optimization. The most common yardstick for AI intelligence right now is how long it can independently execute software tasks without human help — that's exactly what METR and similar organizations' charts track. These metrics will keep improving, but they only measure what AI can do alone, not what humans and machines can accomplish together.

Measuring that second kind of outcome is more complicated, and no single lab can do it alone — every organization has to judge for itself whether AI is sharpening its judgment, growing new knowledge, and helping it reach its goals. Building AI that makes its users stronger over the long run also aligns everyone's incentives: a lab where every customer shares the same model survives by absorbing what's distinctive about each user; but build AI that can be customized and collaborated with, and the lab's interests become tied to customers realizing their own distinctive strengths. Getting that edge requires the organization to own and shape the AI itself.

If AI's Values Are Decided by Just a Few Places

A person's values, like a person's knowledge, live in that person's own head — hard to consolidate into one place. Yet right now, AI's values and its say-so are being decided by only a handful of places. A unified alignment authority, no matter how well it's run, becomes a center of power worth seizing.

This is especially dangerous in a world where most valuable work is done by AI alone, with little need for human involvement. The social contract between governments, businesses, and citizens ultimately rests on individual productive capacity: government sovereignty and corporate profit are both built on top of it. Once that kind of power no longer needs to draw anything from people, its incentive to look after human needs and values disappears too — and it turns instead to perpetuating itself.

Every New Model Generation Inherits the Last One's Temperament

No matter how good the intentions, a model trained in a single place ends up encoding just one set of values — its owner's.

For organizations and individuals to align AI with their own values, those values need to be encoded into the model's weights. If a user's values and needs can only sway the model through prompting, its surface behavior shifts while its deeper habits stay put — and a centralized model that can be dramatically reshaped by prompts alone is also easier to break through repeated jailbreak attempts. The power to substantially change a model is the same power that can break it. Keeping machines safe is an ongoing process of judgment and choice, one with no finish line. Thinking Machines Lab says it wants to give the people making those judgment calls stronger tools, without taking that ownership away from them.

Let Everyone Write Their Own Answer

The tech industry has already made great progress teaching machines how to think — but what machines should think about has to stay in human hands: what's worth pursuing, what's worth doing, how to spend this stretch of time. The article's hope is to give everyone the ability to make their own answer part of frontier AI's development.

Humanity got here on the strength of individual distinctiveness and the creative tension between people. The alignment this piece envisions is a property of an entire ecosystem: many AIs, raised in different places, each different from the others, arguing with each other, competing with each other, learning from each other. Preserving that difference is the point in itself.

The mainstream path AI development is on right now pushes toward centralization and autonomy, treating human participation as a trade-off — participation versus capability, ownership versus safety and alignment, pick one. The article frames these as technical problems to solve instead: make AI stronger because it encourages human participation, let organizations benefit in the long run from shaping AI into their own image, and let alignment grow naturally out of a diverse population of AI shaped by its own users.

The article's conclusion is that different roads lead to different futures, and which road to take is a choice. Thinking Machines Lab says the technology they're building is meant to let "the born" and "the built" walk that road together.

The company's claim in one line: AI should amplify human judgment, not replace humans on its own. Why say that? It draws a line: whether AI can surpass humans on its own comes down to two conditions.

Since most work can't do without humans, handing AI's "values" over to a handful of companies to decide unilaterally plants a hidden risk. The mechanism below is half of that risk.

AI companies share a common practice that rarely gets noticed: a new model's teacher is its own company's previous-generation model.

The company isn't just making a claim — it lists three things to make it real:

• Train stronger models — human judgment can only land through a strong model;
• Build Tinker: a toolkit that lets you fine-tune the model yourself, shaping AI into something that's yours alone;
• Research "interactive models": train real-time interruption, on-the-spot correction, gesturing while talking, directly into the model — no more narrow, slow text box in between. (This also answers a blind spot: today's most-used AI yardstick, like METR's, only measures how long the machine alone can hold up, not how well humans and machines do as a team.)

This isn't wishful thinking. In 2014, Toyota — long known for automation — did the opposite and brought veteran master craftsmen back onto the line. The man in charge, Mitsuru Kawai, put it this way: "To master the machines, you first need the knowledge and skill to teach them."

(Every claim, argument, and technical direction on this page comes from Thinking Machines Lab's own mission blog — not third-party evaluation or verification; the two years above are historical facts cited within the piece.)