↑ ↓ / space / swipe to navigate
01 / 20

$ find ./talk -name "*.slides"

no matches. reconstructing from transcript.md

The New Physics
of Business

Garry Tan — President, Y Combinator

AI Engineer World's Fair · Keynote · 2026

This deck doesn't exist. Garry Tan gave this talk with nothing behind him — "Sorry, there's no slides. I have no slides." (4:23). What follows is reverse-engineered from the transcript of the talk on YouTube.
02 / 20

The question on the table

"
Are we ready for the revolution?

— opening line, 0:21

YC is trying to build companies where one person does what used to take a thousand people — not as a metaphor. Mechanically.

0:52 – 1:13
03 / 20

the number he got "torn apart on the internet" for

400x

more output — same person, same hours, actually fewer

2013

YC partner, near-full-time engineer. ~14 usable logical lines of code a day, at full effort.

2026

Runs YC full-time. Same person. Weirdly, fewer hours — he has a 5pm kid pickup now.

Stress-tested against his own skepticism: even at the most pathological "bloated code" discount, the floor is 8x. Middle estimate, ~40x.

1:32 – 2:50
04 / 20
"
The 2x people and the 100x people are using the exact same Claude. The leverage is not in the weights — it's in how you wire the work.

— 3:06

25%of Winter '25 batch companies had codebases 95% AI-generated — now the fastest-growing, most profitable batch in YC history
94 companieshave crossed $100M in revenue from a seed check, in YC's history
3:17 – 3:45
05 / 20

the heart of the talk — "steal this"

Everything you've learned building with agents maps to an organization

SKILL FILE→EMPLOYEE

One job, written down

One capability, written clearly enough that someone (or something) can execute it.

RESOLVER TABLE→ORG CHART

Who handles what

A task comes in and the resolver decides who handles it and where it goes.

FILING RULES→PROCESS

Is it in compliance

Whether the resolver is actually working, and whether the work follows the rules.

TRIGGER EVALS→PERF. REVIEW

Does it actually fire

A test confirming that, e.g., editing a test file actually loads tests.md as it should.

4:11 – 5:35
06 / 20
"
When you sit down with Claude Code or Codex, you're not writing software. You're hiring, training, and managing a workforce made of markdown.

— 5:58

5:44 – 6:10
07 / 20

real-world impact

Revenue-per-head that didn't exist before

Emergence · AI app builder · Summer '24
15 people
Public launch to 9-figure ARR in 8 months. Only 15 people when they crossed $15M ARR.
Winter '24 batch company
$60M ARR
Roughly 40 people. That ratio, Tan argues, "did not exist before."

"Not in software. Not in oil. Not in railroads. Never."

Company name per auto-generated transcript — may not be exact.

6:14 – 6:38
08 / 20

the extension most engineering talks miss

It's not just engineers

6:46 – 8:17
09 / 20

where almost every AI engineering bug comes from

Two places computation happens

LATENT SPACE

The LLM

Taste. Judgment. Understanding what a human means when they say something vague. Non-deterministic — steered with the markdown file.

DETERMINISTIC SPACE

The code

Where agents write TypeScript, or Elixir, or whatever. Exact math and logic — it must not live in the context window.

Nearly every bug is computation happening on the wrong side of this line.

8:41 – 9:37
10 / 20

example

Seating 800 people perfectly — in ten minutes

Startup School: seat 800 people so the person to your left and right is the best possible person for you to meet.

9:37 – 10:48
11 / 20

Miller's Law

7 ± 2

the entire working memory of a human being

Every org chart, every checklist, every filing cabinet humanity has ever built is a prosthetic for that limit.

~1,000,000 tokensan AI agent's working memory — about three Harry Potter books, open at once, searchable and synthesizable across all three in seconds
10:57 – 12:05
12 / 20
"
Three books is a lot, but it's also very little. Your company is not three books. Your company is a library — every email, every meeting, every decision, its reasoning, every post-mortem.

— 12:25

The question that determines whether your agents are geniuses or goldfish: who decides which three books are open on the desk. That's context engineering.

12:25 – 12:53
13 / 20

"this is just RAG" — yes, and

A company brain is the library plus the librarian

Retrieval is the primitive — the same way Postgres is just B-trees. The hard part is everything around it:

Retrieval is easy. Being worth retrieving from is the product.

12:56 – 13:28
14 / 20

built in the open — "GBrain"

~220,000

pages, written mostly by his agents — from email, meetings, 20 years of notes

Before he even finishes reading a founder's crisis email, his agent has already pulled every prior conversation with that founder — plus what worked for three portfolio companies that hit the same wall.

"That's the difference between an assistant and a colleague."

13:29 – 14:21
15 / 20

stress-testing his own pitch

Company brains fail in specific ways

WITHOUT HYGIENE

A garbage dump with great search

Retrieval surfaces stale facts with total confidence. A bad skill file encodes a bad process forever.

THE FIX: MEMORY + HYGIENE

Provenance and pruning

Provenance on every fact. Contradiction checks when new info collides with old. A librarian — human plus agent — whose job is pruning.

Treat the brain like production infrastructure and it compounds. Treat it like a dumping ground and you get a confident agent that's wrong in ways nobody can trace.

14:21 – 15:04
16 / 20

Garry's signature move

Never do one-off work

Do the task. Don't like the output? Fix it. Then — before you move on — skillify it: turn what you just did into a reusable skill.

"
If you have to ask for something twice, you failed.
15:13 – 16:24
17 / 20
"
The organization that captures what it learns gets smarter every single day. The one that doesn't wakes up every morning with amnesia.

— 16:26

Model quality is rented. The brain you build, you own.

16:26 – 16:40
18 / 20

the answer to "what do we build now"

Build the AI-native company — not one that just uses AI

"Open Claw is the Ferrari — I'll always recommend it. But Codex is a really good Honda, it'll do 90% of this. Use whatever. The concepts are the point, not my repos."

16:41 – 18:17
19 / 20
"
Abundance is not a policy paper. It is shipped software.

— 18:56

A friend with a son who has a rare form of epilepsy built an 80,000-markdown-file company brain for one small boy — pushing to the edge of what humanity knows about his son's exact condition. No lab. No grant. No permission. A father, a laptop, and a library.

19:04 – 19:28
20 / 20

We can boil the ocean now.

Every archive too big to read. Every dataset too gnarly to clean. Every ocean you were told not to boil.

go boil the ocean.
go write that test.
go ship that skill.

Reconstructed from "The New Physics of Business" — Garry Tan, AI Engineer World's Fair. Not an official deck — the talk had none.