I've been building AI systems since 2007, back before the current wave had a name. And everywhere I go this year, I hear the same line: we're in the first inning. Conferences, board meetings, half the podcasts I listen to. So early. So much runway. The real gains haven't even started.

I think we might be in the fifth.

Not because AI is disappointing. The opposite. It's because the first inning framing rests on an assumption almost nobody examines: that intelligence scales forever. Pull on that assumption and the whole baseball analogy flips.

The loudest voices in AI disagree about everything except this. The accelerationists believe intelligence compounds without limit, so we should build as fast as possible. The doomers believe intelligence compounds without limit, so we should be terrified. Same premise, different conclusions. Both camps are arguing about the consequences of an unbounded quantity, and almost nobody is asking whether the quantity is actually unbounded. (Rob May's Investing in AI newsletter has a good essay on this hypothesis: https://investinginai.substack.com/p/the-speed-of-thought-what-if-intelligence)

History suggests it isn't. Every quantity we once assumed was open-ended turned out to have a wall. Temperature has a floor at absolute zero. Speed has a ceiling at the speed of light. Engines hit thermodynamic limits, and computation has hard physical bounds. Physics has a consistent track record here, and intelligence runs on physics.

Two pieces of evidence make the ceiling argument more than a thought experiment.

The first is the human brain. It runs on roughly 20 watts, about the same as a dim light bulb, and evolution spent a few hundred million years optimizing it under brutal energy constraints. That doesn't mean the brain is the best possible thinking machine. But it does suggest that biological intelligence already operates near some efficiency frontier, and that the room above us may be a multiple, not an exponent.

The second is the scaling data we can all see. Each new increment of model capability costs exponentially more compute than the last one did. The frontier models keep improving, but the curve of capability against investment looks logarithmic, not explosive. Chess engines tell the same story in miniature: decades of continued effort, genuine improvement, and an unmistakable flattening. Those are the economics of an asymptote. A system approaching a ceiling behaves exactly like this.

None of this proves there's a wall. But "intelligence scales forever" is a strong claim, and it's worth noticing that the empirical evidence we have looks more like saturation than takeoff.

Here's the part I find most interesting, and it's an analogy I've been using for years.

Once AI passes human intelligence, we lose the ability to measure the difference. Every benchmark we have is calibrated to us. Bar exams, math olympiads, coding challenges: these are human yardsticks, and the frontier models have already blown through most of them. Past that point, Human +10% and Human +100% might look identical from where we sit, the same way my dog can't tell a college freshman from Einstein. The dog isn't equipped to perceive the gap. Neither are we.

This has a strange implication for the inning question. If the ceiling exists and the iteration speed of AI development is as fast as it looks, we could get most of the way there quickly and never be sure we'd arrived. There would be no announcement. The models would just keep getting modestly better on benchmarks we can no longer meaningfully score, at exponentially increasing cost, while everyone kept saying "first inning" because it feels early.

Feeling early and being early are different things. People overbias toward the beginning of every trend. Nobody hosting a panel wants to say "you've missed most of this," and nobody raising money wants to say "the big capability gains are behind us." First inning is what people say when the incentive is to keep the story going.

Suppose the ceiling is real and closer than the consensus thinks. What actually changes?

First, the race to superintelligence is mispriced. Hundreds of billions of dollars are being deployed on the thesis that recursive self-improvement produces an unbounded payoff. If capability saturates, the winners aren't the labs that spend the most on the last few percent. They're whoever applies the existing capability best.

Second, a lot of hard problems stop waiting for smarter models, because they were never intelligence-limited to begin with. Climate is a coordination problem. Aging is an entropy problem. Most organizational dysfunction is an incentive problem. A model twice as smart doesn't fix a supply chain or get a committee to agree.

Third, and this is the part I see every week in my consulting work: nobody is anywhere near the ceiling of what current models can do. A CFO at a $150M company told me last week that an analysis that used to take him 40 hours now takes one, on tools that have existed for over a year. His company had done basically nothing with AI eighteen months ago. That's not a story about frontier capability. That's a story about deployment, and there are thousands of companies still standing where he stood. If capability froze today, mid-market companies would have years of value to extract from what already exists. The gap that matters isn't between this year's model and next year's model. It's between what the models can do and what companies have actually deployed.

I could be wrong about all of this. Maybe there's no ceiling, or it's a thousand times above us, and the accelerationists get their intelligence explosion. But notice which assumption your plans are pricing in. Almost every AI strategy I read implicitly bets on "smarter models will solve this eventually." The fifth inning bet, applying what exists now to problems that exist now, pays off in either world.

And here's the thing about the fifth inning: in baseball, that's when the game becomes official. If it rains after the fifth, the result counts. Whatever happens with the ceiling, this game already counts.

Homework: Run a freeze test with your leadership team. Assume model capability stopped improving today, permanently. List the workflows in your company that current AI could already handle but doesn't. If that list is long, and it will be, then the ceiling isn't your constraint, and waiting for better models is costing you money that better deployment would capture now.

Two things I'd genuinely like back from you. First, hit Reply and tell me which part of this matches what you're seeing. Second, if there's someone in your world who should be on this list, forward it to them.

- Robbie