Why frontier labs are scaling-pilled - by Paras Chopra

Inverted Passion

SubscribeSign in

Why frontier labs are scaling-pilled<br>Investors underwrite scaling laws because the alternative (having an Einstein discover better algorithms) is high variance

Paras Chopra<br>Jul 14, 2026

Share

What would it take to make progress towards general intelligence, where general stands for any problem that might arise in our world?<br>Since our world is big and open-ended, the quest for general intelligence becomes a quest for solving more and more problems, including even the long tail and arcane ones. In fact, many people see AGI as the point when an AI model is able to solve any problem that any other human or a group of humans is able to solve.

Compute is scalable search for patterns

The famous bitter lesson essay starts with this assertion:<br>The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin

To solve any problem, you have to have insights into how the relevant domain works. For example, long before neural networks took over the world of chess, you’d have humans bake in their domain knowledge into chess engines about which positions are superior and which are inferior. Those human-generated insights didn’t come for free, they required human brain-compute. Somebody had to spend a significant amount of time playing chess intently in order to notice repeated patterns that led to winning or losing a game. Only after that expenditure of compute, could we have insights that power a chess engine.<br>But, here’s the key point… human-compute doesn’t scale.<br>Even if you have all the humans in the world search for patterns in a domain and even if the quality of their insights is superior, you’d cap the maximum compute available for a given problem to roughly 8 billion brains. Moreover, you’d have communication and co-ordination issues to integrate human-discovered insights into a cohesive whole. Contrast this with GPUs/compute where getting additional compute is a matter of manufacturing, which can be scaled on demand and discovered patterns can be designed to be better integrated right from the get go.<br>Today it is true that the brute-force nature of search via compute is an inferior proxy for the high quality and generalizable patterns that humans often notice. But the trivial scaling of the compute simply provides it an overwhelming leverage. This is exactly why frontier labs are scaling-pilled: instead of hiring ten thousand geniuses, owning a million GPUs that search for patterns non-stop is simply a more predictable path to general intelligence .<br>Frontier exploration is compute-hungry

The reason we have many branches of science is because the world is big, open and practically irreducible , which motivates discovery of patterns at all levels of abstraction. In theory, all you need is fundamental physics and all the other patterns of our world can be derived from it. We don’t do that, though.<br>We know the rules of quantum mechanics extremely well, but good luck applying it to finding how a Benzene ring forms. You need the rules of organic chemistry in order to do that. Similarly, it’s stupidly wasteful to use cellular biology to predict how consumers will react to a specific product.<br>So, in order to solve more and more problems, we need to keep pushing the frontier of knowledge to discover more and more regularities at multiple levels of abstraction. The entire industry of science is a testament to the fact that even if in theory the world is reducible, practically we need many different branches of science. In order words, to solve problems in the world, we don’t just need the quantum field theory. We need that + thermodynamics + biophysics + molecular biology + psychology + …<br>The beautiful thing about our big-and-open world is that we’d never run out of patterns to mine. We can always find more relationships between domains, a better solution, a faster or cheaper way or simply patterns that were not noticed before.<br>We can keep pushing the frontier of knowledge forward, and hence training of frontier model can always consume more compute .<br>Thanks for reading Inverted Passion! Subscribe for free to receive new posts and support my work.

Subscribe

To ask if we can train a frontier model cheaply is to ask if we can min a max. Whatever amount of compute you have, your competitor with more compute can solve a wider range of problems better by simply throwing more compute at it. This is why pushing the frontier of AI will always be compute-hungry.<br>This is not to say that you cannot aim to approximate frontier capability if you already have access to a frontier model. Even though discovery of frontier knowledge requires massive search, once knowledge is found, learning and using it is relatively efficient. Even high school students today can do basic quantum mechanics, but the original search for it required many...