Some “complete bullshit” about AI and energy

The post opens on what he admits may be a non-sequitur. Photovoltaic cells turn sunlight into electricity more than ten times as efficiently as plants turn it into biomass, which he offers as a reason he is extremely skeptical of biomass-based energy strategies. The comparison he actually wants is the opposite one: nature is terrible at capturing energy and extraordinary at converting it into intelligence. Training a large language model takes hundreds or thousands of megawatt-hours, while his own estimate for raising a human brain to the age of 21 is a few, or about twenty if you count the body as support equipment for the brain. He concludes, at minimum, that the human brain is at least a few orders of magnitude more efficient as a learning machine than the neural networks driving current AI.

From there he runs a deliberately crude extrapolation, and flags it before starting: from here on, he says, his own speculation may also be complete bullshit. Take the roughly 100 trillion synapses in a human brain, multiply by a secondhand estimate that one real neuron does the work of about a thousand artificial ones, and a human-level model would need something like half a million times the parameters of GPT-3. If compute efficiency flatlined, a single training run would consume a double-digit share of annual American electricity demand and tens of billions of dollars of power. Compute efficiency is probably not going to flatline, he says, so he assumes another tenfold gain over ten years, which brings the same run down to a couple of billion dollars and a price he calls very reasonable. The move that matters is that the frightening number becomes an affordable one on an assumption he footnotes as one he has no idea how to evaluate.

Then the turn, which is that training was never the interesting constraint. A trained model can be copied onto unlimited servers for negligible energy, but using it cannot. On an estimate he borrows, answering a year of Google’s query volume with ChatGPT would take a couple of thousand times the energy GPT-3 consumed in training. He is careful to say he has not found a source confirming whether compute for using a model scales with parameter count the way training does; if it does, actually deploying a brain-scale model could absorb much of civilization’s energy. That conditional carries the whole weight of his second conclusion, which is that absent a paradigm shift in how AI is built, approaching human-level intelligence would probably break the energy system.

His third conclusion is the one he is best placed to make and the one he states most plainly: most valuable tasks do not need frontier models. The energy industry was putting AI to work years before the large-scale era began, and for most applications there is not even a training dataset large enough to justify billions of parameters. He offers two companies from his firm’s portfolio as evidence and then generalizes from his own experience: in all his years observing and investing in startups putting AI into energy operations, he has never seen a company whose growth was constrained by the AI inside not being good enough. The stumbles were prosaic, namely missing data, human workflows that resist automation, and the difficulty of improving on a status quo already honed by decades of human intelligence. So he admits to being a skeptic that bigger, better AI changes much for energy or other large industrial operations, and puts the value of frontier models in fundamental scientific work such as materials discovery, where he hopes it accelerates the search for next generation batteries and solar cells. A cheap fine-tuning technique reinforces the point: one big foundational model can spawn thousands of specialized descendants without new training runs.