i don't have the answers and i am thinking out loud, this should sorta be read with my really building for agi post
for years, the battle between startups and incumbents was a fight between bloat and distribution. could startups stay lean enough to outmaneuver the giants? could incumbents leverage distribution to crush their smaller, faster-moving rivals? that was the equation. but now, there’s a third axis—test-time compute—and it changes everything.
what is test-time compute and why it matters
test-time compute is running a model for longer, allowing it to think more. this is what powers advanced reasoning models—your O-series models, your chain-of-thought prompting, tree-of-thought exploration, monte carlo tree search (MCTS). it’s about generating better answers not just by being smarter initially, but by iterating, refining, and reasoning through a problem dynamically. every additional second of test-time compute gives an ai system the ability to re-evaluate, simulate different possibilities, and optimize outcomes, making intelligence itself a function of budget.
in business, many decisions come down to simple capex ROI math—if spending $50 billion on compute increases conversion rates by 5% and that 5% drives $60 billion in revenue, the decision is easy: just spend more. incumbents can turn on this loop and pour infinite money into refining intelligence-driven optimizations, whether it’s maximizing engagement on instagram, fine-tuning ad targeting, or optimizing search placement.
test-time compute becomes the ultimate strategic weapon—an unbounded ability to generate better answers by thinking harder and longer. and in an environment where intelligence can be brute-forced through spending, the company with the deepest pockets will always win.
this is why startups must shift their definition of intelligence. the old game of 'we'll just be better engineers' doesn't work when the real advantage comes from throwing infinite compute at a problem. startups need to be smart not in raw intelligence, but in picking problems where intelligence itself isn't the currency of competition.
problem selection matters more than ever
picking the right problem has always been crucial in building a successful startup, but in the AI age, it's more important than ever. before, startups could win by building better software, designing more efficient systems, or optimizing workflows in ways incumbents couldn't match. but now, intelligence itself has become a commodity—one that can be scaled infinitely with compute.
this means that startups can no longer rely solely on being 'smarter' in the traditional sense. instead, they need to be smarter about where they apply intelligence. the key is to find problems where the marginal returns to intelligence are low—where throwing more compute at the problem doesn't yield significantly better results. this is where startups can still have an edge, not by outthinking in the raw sense, but by out-positioning incumbents through problem selection.
quick tangent on what is marginal returns to intelligence, this is straight from dario amodei's excellent machines of loving grace
Economists often talk about “factors of production”: things like labor, land, and capital. The phrase “marginal returns to labor/land/capital” captures the idea that in a given situation, a given factor may or may not be the limiting one – for example, an air force needs both planes and pilots, and hiring more pilots doesn’t help much if you’re out of planes. I believe that in the AI age, we should be talking about the marginal returns to intelligence and trying to figure out what the other factors are that are complementary to intelligence and that become limiting factors when intelligence is very high. We are not used to thinking in this way—to asking “how much does being smarter help with this task, and on what timescale?”—but it seems like the right way to conceptualize a world with very powerful AI.
this is the fundamental insight: if intelligence alone determines success, the company that can afford to spend the most on compute wins. but if a problem has other constraints or bottlenecks that startups can solve they could win
where startups could win
the key for startups is to identify problems where intelligence is not the primary driver of success. these include:
- problems where intelligence isn’t the bottleneck - maybe compliance, getting things done in the real world or network effects (this feels like a big one).
- markets where incumbents don’t see value in deploying high test-time compute because they don’t think it’s worth the spend.
- ??? - email me ideas [email protected] if intelligence is just one input among many, then startups still have ways to compete but they should make sure to compete outside the axis of intelligence
i am caveating small outcomes, this is always a thing a 200M business is rounding error for microsoft but might be great outcome for a founder, i am talking at scale
this essay was vibes written