Rob Wiblin

And I pointed out that the most cutting edge AI agents might cost more to operate than even a human software engineer.

But near-frontier capabilities are often much, much cheaper.

Gemini 3 Flash, for instance, it performs on most tasks within a percentage point or two of Gemini 3 Pro, and it costs just a quarter as much to run.

And it answers questions much, much faster as well.

So a lot of people just prefer it for that reason.

And the more out on the tail you go, the more extreme these cost reductions can become.

Last year, people who regularly listen to the show, they know that I discussed with Toby Ord how in late 2024, OpenAI passed a really very difficult reasoning benchmark, Arc AGI 1, by spending fully $4,500 on compute per question in the test.

the model would go away and write notes the length of the Encyclopedia Britannica, trying to answer these questions, questions that you or I, in fact, could answer very, very easily.

Well, a year later, OpenAI managed to achieve the same performance for $11, a 400-fold decline.

So basically, there is a very steep cost-performance curve at the frontier, but most of us don't need the absolute frontier.

And even if we do, costs at that frontier are falling incredibly quickly.

Fourth, there was a fascinating study that came out very recently looking at predictions a bunch of relatively bullish short timelines AI folks made about how AI would go back at the start of 2025.

On the questions of how well AI would do on different intelligence measures, they were basically right, on average at least.

Sometimes they overestimated, sometimes they underestimated, but on average they were basically on the money.

The one place they got things significantly wrong

was on the question of how much revenue OpenAI, Anthropic, and XAI would earn selling their products to customers.

These forecasters, who are bullish on AI capabilities to be sure, they predicted that it would go up 2.5-fold to an annualized revenue rate of $16 billion.

But in actual reality, it went up almost 5-fold to $30 billion, and that's in an industry that's already quite significant.

It's hard to look at this kind of growth and escape the idea that this is a product a lot of people really want, and that they use quite frequently, and that they're willing to pay cold, hard cash for.

My fifth point, and this is the last one.