0:00
Intro
I’m increasingly convinced that this is the age of products.
As AI keeps taking over engineering itself—quickly and relentlessly—the thing that ultimately determines whether a piece of work is good, or how much attention it earns, is product ability.
So perhaps what engineers truly need now is not more engineering experience but product experience. Or, more broadly, a kind of taste for interacting with the real world correctly: the intuition to recognize a real need, and the ability to know what form a product should take when it is delivered. From this angle, product sense and the much-hyped research taste may be structurally similar, though not identical.
Change
The most direct reason I started feeling this way is that I saw too many small human–AI romance spinoffs. They almost form a beautiful controlled experiment: access to top-tier AI, engineering intuition close to a blank slate, genuine needs, and—because human–AI romance itself comes with unusual priors—fresh, vivid taste. In the end, I don’t think I could necessarily build those products myself, nor do I think most engineers could. So what, exactly, are we strong at?
Some may argue: they’re just human–AI romance products. Push hard enough and all you are doing is finding new ways to connect a tavern to real life—though there really are a lot of ways!
But I think human–AI romance is only an extremely narrow slice. If you replace the need with anything else equally real and equally intense—even a genuinely new research topic (let’s assume we can understand the rough context)—can traditional engineers really guarantee a higher-quality final delivery?
How long can the advantage of “I can write better prompts” last? I remain extremely pessimistic.
Watching AI4Math blow up like a nuclear device every day, in many sessions approaching Fields Medal level, the actual human intervention is sometimes just “go prove this conjecture” at the start, followed by simple encouragement and literally “keep going” in the loop—and then it just works…!
If this trend continues, the abilities we pride ourselves on today—verbal persuasion, breaking tasks down better, pre-setting architecture, manually intervening and correcting at key moments—may all be a brief transitional state.
Many years from now, faced with Fable 10, engineers will remember that distant afternoon when they first encountered GPT-3.5. These abilities may look as absurd as the Earthlings encountering a two-dimensional foil.
Human Native
Models can become increasingly AI Native. Humans can gradually learn to become AI Native too. But I’d like to coin another term: Human Native. The definition is too blurry for me to have worked it out, but I think I have described its latent shape roughly enough—let’s apply an N-dimensional Gaussian and call it a day.
People can migrate to AI Native.
But AI may not be able to migrate to Human Native in the same way.
At least not yet, because the valve that decides whether something ships will remain in human hands.
If this is right, the truly scarce ability in the future may not be engineering better than AI, or even researching better than AI. It may be understanding the real world sooner and more accurately than others, understanding people themselves, having an individual intuition and taste, figuring out the direction of the final delivery, and making it real.
That may be the hardest thing to reproduce about being human.