Key points
- Founders should build for where AI technology is going, not where it is today, according to the clip.
- The clip warns that building where technology is already as good as it will be makes long-term differentiation very hard.
- It says short-term heuristics for current models have massive diminishing returns and founders should focus on user-learned experiences.
Founders Should Build For The Next AI Model
The episode is a short argument about timing in AI products. A founder should build for where the technology is going, because anything built on today's ceiling has no room left to improve and will be hard to differentiate. That bet has a price, since it means working on something that does not feel like it is working for an unknown stretch of time. The advantage worth holding onto is the user base, because usage shows you which experiences to build next.
Takeaways
- Build for where technology is heading. The speaker's advice is to build for where you think the technology is going, not for where it is today. Anything built on today's version of a model is betting on a ceiling that will move.
- Finished capability leaves no differentiation. Building in places where the technology is already as good as it will ever be makes long-term differentiation very hard, the speaker argues, because every competitor is working with the same completed capability.
- The forward bet means a fallow stretch. Building for the future carries a cost. It means being willing to work on something that does not feel like it is working for an indeterminate period of time, which the speaker calls a genuinely hard thing to do.
- Heuristics for today's model fade. Companies invest deeply in techniques that make the current model perform, but the speaker says the returns diminish massively, and the effort is sunk once the model changes underneath it.
- Users teach you which experiences to build. The advantage the speaker points to is having a lot of users. Because people use the product, you can learn from how they use it and build better and better experiences, and that is where founders should focus.
What the speaker expects next
- The next AI model will be much better than the current one, so the heuristics companies build to compensate for today's limits will become unnecessary.
- Capabilities that take heavy engineering now will probably arrive on their own within about six months, which is why the speaker treats that work as cosmetic.