The
Napkin Math That Built Google — And Where Jeff Dean Says You Should Look Next
Jeff Dean spent 27 years at Google,
helping build the search engine, TensorFlow, and the TPU chips that now power
much of the world's AI. In mid-2026, weeks before leaving the company, he sat
down with Y Combinator for one of his most candid interviews yet.
Most people will read this for the AI
predictions. But the real lesson is smaller and more useful — a habit Dean used
twice to spot billion-dollar bets before anyone else could see them: doing the
math on a napkin.
Two Napkins, Twenty
Years Apart
2001. Dean and colleague Sanjay Ghemawat did
a rough calculation: could Google's entire search index fit inside a computer's
memory (RAM) instead of sitting on slow disks? The math said yes. They built it
in days. Search went from slow to instant — and became one of Google's biggest
advantages.
2013. Another napkin, another calculation.
Dean worked out that if every user used voice typing for just three minutes a
day, Google would need to double its entire server fleet. That one number
justified building custom AI chips from scratch — the project that became the
TPU, hardware now running much of the world's AI.
Neither started as grand strategy. Both
started as a calculation that made a hidden problem impossible to ignore.
Dean's pattern: before you build anything, find the number that tells you
whether it's worth building.
Are AI Models Already
"Junior Engineers"?
A year before this interview, Dean
predicted that by 2026, AI would work like a junior software engineer — coding
on its own for hours. Here, he admits something rare for a tech leader: he was
too cautious. AI improved on complex, multi-step coding tasks faster than even
he expected.
But he's careful about what that means.
Today's models excel at tasks close to what they were trained on. Move slightly
outside that zone, and performance drops fast. That gap — between "looks
like a junior engineer" and "reliably solves a new kind of
problem" — is where he sees the real opportunity for builders.
The Rule: Look for
0%, Not 20%
This is the most quotable, useful idea
in the whole interview. When picking a problem to build a company or product
around, Dean's advice is almost the opposite of common startup wisdom.
Most people look for a problem AI is
already partly good at — say, 20% successful — and try to push it to 90%. Dean
says: don't. If a general AI model is already at 20%, that's a sign the big AI
labs will solve it soon with more training data, because the capability is
clearly emerging.
Instead, look for problems where the
best AI models succeed 0% to 1% of the time. That sounds discouraging,
but it's the opposite. It usually means the problem needs something today's
general models don't have — specialized data, a unique workflow, or deep domain
knowledge that only a focused team can supply. That's where two or three
people, not a giant lab, can still win.
Why "Taste"
Is the New Skill
If coding becomes cheap because AI
writes most of it, what should humans focus on? Dean's answer is a word you
don't often hear from an engineer: taste. Knowing which problem is worth
solving, which experiment is worth running, and which shortcut will actually
work — that judgment doesn't come from a model. It comes from experience,
curiosity, and yes, more napkin math. As AI writes more of the code, picking
the right problem becomes the valuable skill, not writing the code
itself.
The Next "Fits
in Memory" Moment
Dean believes 2026 has its own version
of the 2001 RAM story. Back then, the breakthrough was data fitting in memory.
Today, he says, the breakthrough is inference — the process of an AI
model actually answering you — becoming fast and cheap enough to run on
specialized chips rather than general-purpose hardware. He believes systems 50
times faster than today's could unlock entirely new products, the same way
instant search once did.
He also predicts something bigger: AI
systems that can run their own experiments, break big problems into smaller
ones, and improve themselves automatically — a kind of automated scientific
method that could eventually speed up research in fields well beyond computer
science.
Key Takeaways
- Do
the math before you build. Google's biggest breakthroughs
started as simple back-of-the-envelope calculations, not grand strategy
documents.
- AI
can now code like a junior engineer on familiar tasks, but its
performance drops fast outside that comfort zone.
- Look
for 0-1% success rates, not 20%. Problems AI already partly solves
will likely be solved by big labs soon anyway.
- Judgment
("taste") is becoming the scarce skill. As code gets
cheaper to produce, knowing what's worth building matters more.
- Watch
inference speed, not just model size. Faster, cheaper AI responses —
not bigger models — may be the next big unlock.
FAQ
Who is Jeff Dean? Jeff Dean was
Google's Chief Scientist and one of its earliest, most influential engineers,
known for building core Google infrastructure and AI systems like TensorFlow
and the TPU chip.
What is the TPU, in simple terms? It's a computer chip
Google designed specifically to run AI calculations, instead of using
general-purpose chips. This makes AI faster and cheaper to run.
What does Dean mean by the "0-1%
Rule"?
When choosing a problem to build around, target things today's best AI models
almost never succeed at, rather than things they're already partly good at. It
signals real, defensible opportunity.
Are AI models really as good as junior
software engineers now? For coding tasks similar to their training, yes, largely.
For new, unfamiliar, complex problems, they still struggle — and that gap is
where Dean sees opportunity for builders.
What should a non-technical reader take
from this?
Big ideas often start with a simple, honest calculation, not fancy tools.
Asking "does this actually add up?" is a skill anyone can use.