Most AI founders target problems models already handle roughly a fifth of the time.
Google’s Chief Scientist Jeff Dean argues that’s exactly the bet that gets you killed: 20% success means the capability is already arriving and the general model is likely to catch up within six to twelve months.
Drawing on his full conversation at YC Startup School, this issue breaks down Dean’s:
actual selection filter
durability test
three moats general models structurally cannot copy
context-engineering plays any founder can run today with nothing but an API, so you can build for where capability is going, not where it already is.
Inside you will find:
Why Agents Will Soon Run for Days
The Napkin-Math Bet Behind the TPU
Context Engineering Without GPUs
Keeping Agents on the Rails
The 1% Rule for Picking Problems
Three Moats General Models Can’t Copy
Specs and Taste Are the Scarce Skill
The Evaluator That Ate Six Months
This issue is available exclusively to paid subscribers!
Paid subscribers unlock the full library: 200+ decks, 10,000+ investors, AI Agent playbooks, and fresh high-signal content every week.
Not subscribed yet? Try it free for 7 days and unlock this guide, plus our full premium library:
We’re already 80,000+ founders and operators inside!
Let’s begin!
