I remember taking apart the first PC we had at home when I was seven or eight. This magical machine I played games on had suddenly stopped working. My dad and older brother were on their way home, and I thought I could fix it before they arrived.
I didn't fix it, but I wanted to understand what I had found inside.
Years later, learning that all of it starts with sand and energy felt like opening that computer again. Games, robots, and perhaps one day superintelligence.
A chip is sand that has been taught to count, and a model is a chip that has been taught to answer. We are dust too.
Now the sand is starting to think. As more of us can ask it for help, I wonder how much thinking I will keep doing for myself.
The price of an answer
On stage at Google I/O, Demis Hassabis marveled that "we turn sand into thinking machines," and almost in the same breath spoke of serving Gemini Flash to everyone who wants it at "incredible low cost." I find the order strange. We have yet to understand what we have built, and already we argue about how cheaply to give it away.
A reply at GPT-3.5 quality, among the best AI available a few years ago, has fallen from twenty dollars to seven cents per million tokens.
From a phone, intelligence is weightless. From a data center, it has substations, cooling loops, optical fiber, concrete, and a great deal of electricity, and between the sand and the answer stands the whole chain of fabrication, plants and lithography and packaging and memory, and behind all of it, billions of dollars. Silicon is abundant; frontier compute is not.
I expect access to spread quickly as we build more capacity. How much will the people with their own compute still be able to do that everyone else cannot?
Who decides?
Even judgment may become cheap. A model can propose several answers, choose one, act on it, and correct itself from 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, and no person reads the result. It can 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 put their name under what the model chose.
What optional costs
I want dangerous work to become optional. I worry about losing the habit of thinking through something difficult before asking for help.
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 AI to help draft and edit this essay. It made the argument clearer and the work easier. I honestly cannot tell how much of the thinking it extended and how much it replaced.
I can see the improvement in the essay. I have a harder time seeing what happened to my own thinking while I wrote it.
There may also be more room for questions with no useful answer. More conversation, more contemplation, more prayer. A strange result of the semiconductor industry may be more rabbis and monks.
We are dust that learned to think. We took sand, gave it energy, and taught it to think too.
Related: The Control Room
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," Anthropic, 2022.