AI Is Just a Hack. So Was Newton.

George Hotz — the guy who jailbroke the original iPhone and built tinygrad — set AI Twitter on fire by saying the whole AI direction is wrong: large models are just a hack, cobbled together, not a rigorous technical system.

He's not wrong about the diagnosis. Anyone who's looked under the hood knows today's LLMs are a giant probability game. They don't understand logic or the rules of the real world; they fit a distribution over text from massive data and statistically guess output that looks fluent and correct. Architecturally it really is an improvised hack — there's no complete, self-consistent theory of intelligence underneath.

Where I disagree is the conclusion: that we should therefore throw the whole route out.

A hack that still moves the world

Even if AI is only a stage hack, it has genuinely raised productivity across industries and solved real problems that were stuck for years. That counts.

Look at the history of science: there has never been a perfect, one-shot ultimate technology. The flawed, approximate, even "wrong" temporary solutions are exactly the stepping stones to the next, more rigorous thing. Without those imperfect intermediate steps to accumulate practice, we never find the direction closer to the truth.

Newton was a hack too

Take the physics everyone learned. Newtonian mechanics, viewed from relativity and quantum mechanics, is a bounded, incomplete approximation — in a sense, its own way of simplifying the world. At high speed, strong gravity, or the quantum scale, Newton's laws break down completely and can't describe reality.

And yet — even though Einstein overturned its underlying logic — engineering, aerospace, mechanical, and civil work still run on Newtonian mechanics today. Why? Because in the overwhelming majority of everyday situations the approximation is precise enough, cheaper, and simpler to compute. Its flaws only show up at extreme, special conditions that don't touch the value it creates every day.

AI is exactly the same.

Manage the hack — don't worship or trash it

We fully admit current LLMs and agents are probability-fitting stage hacks with inherent flaws and blind spots you can't fully remove. But you don't discard something just because it isn't perfect.

This is the whole logic behind what I've been calling loop engineering — and the three production AI loops we run on Molecule AI. We know AI has limits, so we wrap it in a closed loop of safeguards: multi-layer verification, human oversight on top, heterogeneous model cross-review, and objective tests as the hard backstop. Engineering patches the hack's weaknesses and plays to its strengths.

If we insist on chasing a flawless ultimate technology and reject every imperfect transition, progress simply stops. Every efficiency gain, every round of trial and error — flaws and all — paves the road to the more rigorous tech that comes next.

Hotz pointing at AI's foundational flaws is a valuable, clear-eyed reminder: don't blindly hand everything to AI, and don't let autonomous loops run unattended. But we shouldn't swing to the other extreme and write off the productivity revolution this stage of AI is delivering.

See clearly that it's a hack, hold the engineering line, and use the leverage it gives you. That's the more honest — and more useful — position.