Thinking Machines Lab: Mira Murati wants to make AI consistent. But what remains of the human in a world ruled by language and data?

Two billion dollars. A llm project, No product. One radical ambition: tame the chaos of artificial intelligence and turn it into computational order.

Mira Murati, former CTO of OpenAI, returns with Thinking Machines Lab, a startup that isn’t promising yet another consumer chatbot but something deeper — and far more dangerous: to make AI deterministic, that is always consistent, predictable, reproducible.

But while technique accelerates, philosophy stumbles. Behind this algorithmic challenge lies an existential question:
Can reality be reduced to language? And above all… should it?

If the machine learns to always give the same answer, do we stop asking questions?

LLM road to AGI: an impossible dream. Symbolic AI: an even worse nightmare

The industry is obsessed with AGI: a machine able to think and decide like — or better than — a human. Two main paths, two opposite delusions.

1) LLMs and generative AI — the probabilistic path

Statistical models trained on massive text corpora, linguistic imitation, stochastic “creativity.” Dazzling, but they don’t understand.

The impossible road to AGI: brilliant storefront, hollow foundations.

2) Symbolic/Computational AI — the logical path
Formal rules, mathematical deduction, ontologies. Seemingly more rational, but more dangerous: it doesn’t simulate intelligence, it claims the truth. It doesn’t converse, it imposes.

Two extremes, two risks:

• LLMs risk machines that appear intelligent without being so.
• Symbolic AI risks systems that impose their logic on reality.

Murati’s project moves between these poles: a generative AI with computational ambitions. A probabilistic machine… that must no longer be probabilistic. A new hybrid. Powerful. Potentially devastating.

The technical core: making AI deterministic

The problem Thinking Machines Lab aims to solve looks merely technical: the inconsistency of Large Language Models. Today, with the same prompt, models like ChatGPT, Claude, Gemini produce different answers. This blocks trust, reliability, adoption.

The invisible culprit of LLM: GPU kernels and inference

According to researcher Horace He, the chaos stems from the orchestration of GPU kernels during inference: micro-programs that execute computations not always in the same order. Critical factors:

• Batching of requests (how many queries run simultaneously)
• Floating-point micro-variations
• Execution context/environment of the runtime

Result? Even at temperature = 0 the model changes behavior. For finance, legal, medicine, research, this unpredictability is unacceptable.

LLM SOLUTIONS: a deterministic runtime (batch-invariant AI)

Thinking Machines wants to rewrite at a low level: not a prompt tweak, not a model patch. A new computational runtime that:

• always produces the same output for the same input;
• is independent of parallel load (batch size);
• makes inference 100% repeatable;
• orchestrates kernels in a controlled, stable, orderly way.

In short: mathematically reliable AI. It doesn’t change, doesn’t drift, doesn’t hallucinate.

Where deterministic AI actually matters

• Finance & Legal — Zero ambiguity: you need safety, not originality.
• Scientific research — Reproducibility is law.
• Reinforcement learning — Removing noise makes training more efficient.
• Hallucination reduction — Less pipeline variability = fewer fabrications.

The linguistic codification of reality: the real project of modern AI

Modern AI bets that language is enough to describe the world. LLMs turn reality into tokens, encode it in sentences, and entrap it in predictive structures. Thinking Machines goes further: it wants this codification to be deterministic. Perfect. Irreproducible. “True.”

But reality isn’t language. Life isn’t a dataset. The human being isn’t a function. And yet, if the world becomes fully legible to the algorithm, the algorithm can govern the world.

Represents for whom?
Optimizes for what?

The technocratic dream: smart cities, automation, surveillance, control

• Smart Cities — Every movement, consumption, decision tracked and optimized.
• Total automation — Intelligence isn’t creative: it’s operative, replacing every repeatable function.
• AI as authority — A deterministic model is the base for algorithmic governance: no ambiguity, no contradiction.
• Surveillance & management — If behavior is codifiable, it’s also predictable — and therefore controllable.

A perfectly predictable world. But in such a world… where is the unpredictability that makes us alive?

Others try “soft” fixes. Thinking Machines rewrites the code

• OpenAI — Trains model humility: legitimizing “I don’t know”. Noble intent, but it doesn’t touch the structural problem.
• Prompt Engineering & RAG — Improve accuracy and grounding but don’t guarantee consistency and repeatability.
• Murati — Not a new LLM, but a new computational logic: changing AI’s foundations.

LLM + Symbolic AI = the end of ambiguity?

The near future may merge the linguistic fluidity of LLMs with the logical coldness of Symbolic AI: algorithms that understand language, derive rules, and impose them as truth. A reality governed by coherent outputs, yet deaf to doubt. To contradiction. To grace.

LLMs vs Symbolic AI, AGI, and the algorithmic cage

LLMs chase intelligence; Symbolic AI formalizes it. Both miss the human. But one is more dangerous. Thinking Machines Lab presses on the most concrete, least narrative point: consistency, predictability, repeatability. If it succeeds, we’ll see industrial/scientific/institutional LLMs — reliable, unerring, unmovable machines — perfect for enterprises, governments, complex systems.

The catch? Life isn’t a prompt.
Reducing reality to a coherent output means sacrificing complexity. Contradiction becomes error. Doubt an anomaly. Chaos a bug.

Can reality be reduced to language? Perhaps.
Do we really want it to be?
The real danger of AI isn’t that it fails.
It’s that it succeeds. Too well.