Ramez Naam put a human brain beside an AI computer in the closing discussion of this Moonshots episode. His comparison: roughly 20 watts for the brain, against a 20-kilowatt machine running a frontier model. Co-host Alex wanted to count something else—the work coming out.
Under optimal conditions, Alex estimated, that box’s dozen GPUs—the processors doing the AI computation—could run about 500 simultaneous streams of model output. If each produced text ten times faster than a person, the machine would deliver about 5,000 times one writer’s output while drawing 1,000 times the brain’s power.
Naam saw an opportunity to learn from biology. Alex saw a machine whose appetite for electricity could look much more reasonable once divided by its output. Their disagreement turned on what “efficient” means: low power, little energy per completed task, or learning a skill from fewer examples.
Twenty watts against twenty kilowatts
Watts measure how quickly energy is being used, not how much a job takes to finish. A machine drawing more power can use less energy per task if it finishes quickly enough. Alex suggested comparing the energy needed to write a novel, rather than simply comparing the brain and computer while they are running.
His rack arithmetic illustrated the throughput argument. On those assumptions, the computer would produce about five times as much text per unit of energy. But the calculation counted tokens—pieces of text processed and generated by a language model—not finished novels of comparable quality. It was an estimate for an optimally running box, not a measured comparison of completed writing assignments or whole data-center energy use.
Naam’s point extended beyond operating power. Training is the process of building a model’s capabilities from examples; inference is using the trained model to produce an answer. He argued that current systems still have much to learn from the brain in both how they acquire skills and how they process information.
“Scaling is not everything in AI,” he said. “Scaling is just what we knew how to do.” Researchers had neural networks, the transformer design behind many language models, and an enormous supply of internet text. Spending more on training offered a predictable route to better results.
“It wasn’t the cheapest way or the best way, but it was a predictable way,” Naam said. He expects discoveries in how information is processed to make it possible to do more with less, though he had no curve to plot for that prospect.
Children, training data and evolution’s bill
Naam acknowledged that models trained on trillions of tokens can draw on more text than a person could read in a lifetime. He compared that advantage to a search engine’s vast index beside a librarian’s knowledge: breadth can enable work that one person cannot do, without settling which system learns more efficiently.
For learning itself, he pointed to children. Humans improve skills from far less data than current models require, he argued. Alex agreed that the gap offered an opportunity for better algorithms.
That was not a claim that biology must remain superior. “I’m not a carbon chauvinist,” Naam said. He believes digital intelligence can surpass humans, while arguing that present algorithms lack some of the advantages evolution built into human cognition.
Alex’s counter widened the accounting boundary. A child does not start with an architecture invented from scratch: the brain benefits from billions of years of evolution, a process with an enormous historical energy cost. He suggested comparing that inheritance with the upfront cost of training an AI model.
The discussion then widened again. A trained model can be reused without repeating its original training, spreading that cost over many future tasks. Yet its training material also rests on generations of human effort—and the calories spent producing it. Neither the rack’s instantaneous power nor a child’s learning experience captures that entire history.
What the scaling studies actually measure
The closing exchange returned to a question Alex had raised near the beginning: does intelligence improve only slowly as computation grows, and would that remain true if AI began improving itself?
Naam invoked diminishing returns. Better algorithms can shift the relationship, he said, but he expects additional gains to become progressively harder to obtain. The studies discussed in the exchange address a narrower, measurable question: how training resources affect a language model’s prediction errors.
Kaplan and colleagues’ 2020 study trained language models on web text and found power-law relationships between prediction loss—a measure of error when predicting the next token—and model size, data and computation. These were not measurements of a universal quantity called intelligence. The authors distinguished smooth improvements in average loss from qualitative changes in capabilities, and left generalization beyond language modelling open.
The 2022 Chinchilla study showed that allocation matters as well as scale. Across more than 400 training runs, its analysis indicated that model size and training-token count should grow approximately together as the computation budget increases. That does not mean splitting a budget equally between two expenses.
Its large-scale demonstration used a 70-billion-parameter model—parameters are the adjustable values learned during training—with four times the training data of the 280-billion-parameter Gopher. Using the same training computation, the smaller Chinchilla outperformed Gopher on nearly every measured evaluation and required less computation to run afterward.
That large-scale validation involved one pair of models; the smaller fitting experiments used less than one pass through their training data. Better language modelling also did not improve every measured property: automated toxicity scores on 25,000 unprompted samples showed negligible differences. The result offered a concrete way to spend computation better, not a general law of intelligence.
Could improvement—and demand—flatten out?
Naam extended his diminishing-returns argument to recursive self-improvement: AI helping develop better AI, which then helps develop the next version. Setting hardware advances aside, he argued that software-only improvement eventually slows. The initial jump could still be enormous, but he said he saw no valid mathematical model yielding an unlimited, near-vertical takeoff.
That is Naam’s assessment, not a theorem established by Kaplan or Chinchilla. Their experiments neither test a self-improving research loop nor rule out rapid advances from new algorithms.
Asked later whether demand for computing energy could grow without limit, Naam began with uncertainty: “None of us really knows.” His guess was an S-curve—rapid growth that eventually flattens—as the cost of an additional capability stops being justified by its economic value.
New discoveries could change where that happens. In Naam’s account, the eventual spending decision is concrete: whether the next increment of useful work is worth the extra computation and electricity needed to obtain it.