Originally published on LinkedIn. View the original post ↗
I've never played Go. I've played enough chess to appreciate its complexity, but Go is orders of magnitude more complex by possible positions. More than the number of atoms in the universe.
AI makes quick work of world champions.
At The AI Conference I attended a talk from Ramez Naam called "Jagged Minds". Mapping AI's success to a game - even an extraordinarily complex one like Go - to other things we might want AI to succeed at is folly, he suggested.
Take curing aging, for instance. Researchers using AI, or even AI itself, might come up with a promising idea, but finding out whether it works could take years, decades, lifetimes.
You can speed up go to run at computer speed, not so easy with life.
For those of us building software, I see a practical implication to that.
If we want to speed up delivery of software, a focus on only the coding will just yield pressure on the roadblocks that come next. We have to give the entire loop conditions upon which to evaluate success. Doing otherwise just means the process will pile up unverified work on the next task.
Determining success criteria is now the critical path towards rapid iteration. True agentic development needs to start there.
The more confidence you have in the tests, the faster your agents will iterate.
Your test strategy determines your speed limit.