Let the Sand Think — Shai Magzimof

Shai Magzimof

EN<br>עב

Let the Sand Think

July 14, 2026

The best intelligence money can buy will soon reach almost everyone. What does that do to ours?

A billionaire can buy a better house, a faster plane, and more of a doctor's time. He cannot buy a phone from 2035. The best consumer technology reaches millions of people too fast for anyone to hold a private frontier, and intelligence is starting to behave the same way.

What happens when everyone can reach the best intelligence money can buy? Two things at once, I think: we get more curious, and we get less practiced at thinking.

The marvel, priced for everyone

Demis Hassabis put the economic story in two sentences without meaning to. On stage at Google I/O he marveled that "we turn sand into thinking machines," and almost in the same breath talked about serving Gemini Flash to everyone who wants it at "incredible low cost." The wonder and the price cut are the same event.

Running a model at roughly GPT-3.5-level quality fell from about $20.00 to $0.07 per million tokens between November 2022 and October 2024, more than a 280-fold drop in under two years. A strange scientific achievement is already cheap enough to hand to everyone.

The bottleneck moves

From a phone, intelligence looks weightless. From a data center it looks like substations, cooling loops, optical fiber, concrete, and a great deal of electricity. Silicon is abundant; frontier compute is not. Between sand and an answer sit the whole chip-fabrication chain (plants, lithography, packaging, memory), the power and cooling infrastructure it needs, and billions of dollars of capital.

We can build more of them, and as we do the cost of a useful answer keeps falling. Access never becomes perfectly equal, but the capability gap compresses toward the phone curve: a lead measured in months rather than lifetimes.

What stays scarce

Even judgment might not stay scarce. A model can propose several answers, pick one, act on it, and adjust based on what happens. Anthropic's Constitutional AI already runs a version of this loop: one model grades another's answers against a written set of principles, with no person reviewing the result. It can already pose a sharper question than I can and reach a better answer than I would.

Whose preferences is it serving, and who gave it the authority to act? A model can infer and balance preferences well, but it cannot answer either question about itself.

Not everyone who can reach the same capability will want the same outcome, carry the same responsibility, or be willing to stand behind a decision, even one a model made.

What optional costs

Making dangerous, exhausting work optional is mostly progress, and I am not defending hardship. Some human capacities stay alive only through use, and when producing an answer is always slower than asking for one, we can quietly stop building the ability to produce it ourselves. If choosing which answer to trust gets outsourced too, the risk just moves: we stop practicing how to set the goal, question the objective, and decide what is worth doing at all.

Take the things we say we want most: back to the moon, on to Mars, a real reading of the universe. The old framing for all of them was sacrifice, a career spent on one equation, a decade of training to sit on top of a controlled explosion. Losing that sacrifice is mostly fine, but losing the practice underneath it worries me, because practice is how the ability to think stays in shape.

I used artificial intelligence to help draft and edit this essay. It made the argument clearer and the work easier, and I honestly cannot tell how much of the thinking it extended and how much it replaced.

Practice moves rather than disappears. If producing answers gets cheap, the practice that matters shifts to noticing what deserves our attention, deciding what we actually want, and sitting with questions that have no useful answer. That could mean more conversation, contemplation, and prayer. A strange outcome of the semiconductor industry might be more rabbis and monks.

The authority question above gets an actual mechanism in The Control Room.

Shai Magzimof

Sources

Demis Hassabis, on-stage interview with Alex Kantrowitz at Google I/O (May 2025), with Sergey Brin joining: "we turn sand into thinking machines," and on serving Gemini Flash at "incredible low cost." Big Technology.

Inference cost for a model at roughly GPT-3.5 quality falling from $20.00 to $0.07 per million tokens (Nov 2022 to Oct 2024), a decline of more than 280 times. Stanford HAI, AI Index 2025.

Silicon as the second most abundant element in the Earth's crust, after oxygen. USGS, Mineral Commodity Summaries: Silicon.

Anthropic's Constitutional AI trains one model to critique and revise another's outputs against a written set of principles, using AI-generated rather than human feedback to judge harmlessness. Bai et al., "Constitutional AI: Harmlessness from AI Feedback,"...