Plan AI Products With a Flexible 60-20-20 Rule
Ship something that works today, without closing off what could become magical as models improve.

Lenny's Podcast: Product | Career | Growth · What it actually takes to build frontier models | Tulsee Doshi (Head of Product for Gemini)Tulsee Doshi, who leads product for Google's frontier-model work including Gemini, builds around one reality: models keep changing, so today's limits may not hold.
Her workaround is a rough 60-20-20 split. It is built to ship something useful now while staying open to what models may soon do.
Start by making about 60 percent of what you build work now, so the prototype is at least functional. That departs from traditional startup advice to ship only a minimum product that already works.
Put about 20 percent into features that are imperfect but close, working well enough that near-term progress could plausibly finish the job.
Save the last 20 percent for where you want models to go. Look for glimmers of light, early signs that an experience could become magical if you cracked it.
The numbers are flexible, Doshi says. The point is balance: ship something great that works, without betting everything on the present or the future. Building that way leaves a working product today and room to grow as models improve.
She finds it hard to say exactly where models will be three months from now, even working on them. She prefers concrete proxies in your own product, places where you see the gap starting to close.
Questions people ask
How is this different from just building an MVP?
A minimum viable product usually means shipping only what reliably works. Doshi argues that is too narrow for AI, because models are consistently changing. Her mix keeps a working core while reserving space for near-term gains and bigger breakthroughs.
What does a glimmer of light look like in practice?
It is an early, inconsistent sign that a hard task might soon be possible. That could be a demo that succeeds once, a difficult prompt that partly works, or a workflow that fails but fails in a promising way. Builders use those clues to decide where to invest.
How often should I revisit the split?
Doshi does not give a timetable. She says it is hard to forecast exactly where models will be months out, even for someone building them. She suggests watching your own product for places where the gap is starting to close, rather than betting on a date.


