Leaving Tesla, returning to first principles, and learning how to recalibrate in public.
Originally published on Lei Hua's Substack.
Epigraph
"I think a few years ago, I sort of felt like AGI was — it wasn't clear how it was going to happen. ... And now I think it's very clear, and there's like a lot of space, and everyone is trying to fill it."
— Andrej Karpathy, Sequoia AI Ascent · 2024-03
I. The Photo
In March 2024, in a room at Sequoia's AI Ascent conference, Andrej Karpathy walked up to the stage. The host, Stephanie Zhan, had opened a background slide ahead of time — an early photo of him. Karpathy glanced at it and laughed: "It's a very intimidating photo."
The man in that photo belonged to another era. The line between that era and this one runs right through the summer of 2022.
II. The Departure
In July 2022, Karpathy left Tesla. This is public knowledge. He had been Tesla's Director of AI for five years, leading the vision-only Autopilot architecture. Stephanie Zhan, two years later in 2024, summarized the outsider's view:
"He was poached by Elon in 2017. ... For folks who don't remember the context then, Elon had just transitioned through six different autopilot leaders, each of whom lasted six months each. And I remember when Andrej took this job, I thought, Congratulations, and good luck."
Those five years, he lived inside an engineering culture heavily shaped by Elon Musk. He described that work style, looking back in 2024:
"He likes very small, strong, highly technical teams. ... Companies, by default, the teams grow and they get large. Elon was always like a force against growth."
"He doesn't like large meetings. ... if you're not contributing, and you're not learning, just walk out. And this is fully encouraged."
"Usually, a CEO of a company is like a remote person five layers up... It's not how he runs companies. ... If the team is small and strong, then engineers and the code are the source of truth."
"I like to say that he runs the biggest startups."
These are his 2024 recollections. He didn't say any of this publicly the moment he left, in the summer of 2022. He took about three months before reappearing on camera at all.
III. The Tone of Leaving
October 29, 2022 — Lex Fridman Podcast Episode #333. This was his first long-form public conversation after Tesla.
Three and a half hours. The conversation drifted from the mathematical elegance of neural networks to aliens, synthetic biology, the shape of AGI, the simulation hypothesis. In that version of Karpathy, the metaphors for AI were romantic — neural networks as "another kind of intelligence in nature," sitting alongside animal brains and human minds as a peer form of cognition. His tone toward AGI was curious and open: he neither rushed it nor denied it. AGI, in that voice, had no timeline. Only possibilities.
If this whole biography traces a curve, that moment is its starting point. No founder identity yet; no clear version of the educator mission; no LLM OS metaphor; no Software 3.0 framework; no "decade of agents"; no "summoning ghosts"; no "slop"; no "AGI is still a decade away."
Just an engineer who had just let go of the wheel and hadn't yet decided which direction to drive next.
IV. A Few Things That Hadn't Happened Yet
A few things hadn't happened to him yet at that October 2022 moment — and each one, in turn, would change him. Listed in chronological order, they almost form the chapter outline of this book:
On November 30, 2022, ChatGPT launched. This day would re-define the context of every public appearance he would make afterward.
In early 2023, he re-joined OpenAI — research scientist again, but this time also a public educator.
In May 2023, his "State of GPT" talk at Microsoft Build laid out the entire LLM training stack in a public, anatomical way that the industry would quote for years.
In November 2023, his hour-long *Intro to Large Language Models gave the public the "LLM OS" metaphor for the first time.*
In February 2024, he left OpenAI — for the second time. Half a year later, he founded Eureka Labs.
In June 2025, at YC AI Startup School, he laid out the full Software 3.0 framework.
In October 2025, on Dwarkesh Patel's podcast, he uttered the sharpest line he had ever said about this AI wave — "It's slop. AGI is still a decade away."
In December 2025 — what he himself would later call his "personal inflection point" — his own coding flipped from "I write 80%" to "agents write 80%."
In April 2026, he returned to the same Sequoia AI Ascent stage, the same host, and said: "I have never felt more behind as a programmer."
Each of these events did not turn him into a different person. His code aesthetic didn't change. His allergy to hype didn't change. His preference for "small teams, deep technical work, remove bottlenecks" didn't change. His reverence for education didn't change. What changed was the world. The world got faster every year, and his judgment had to be calibrated more sharply, every time.
And that, precisely, is what this book hopes to refract back to the reader: in a world moving faster than you, a thinking person is not someone who refuses to change. A thinking person is someone who changes gracefully.
V. One Line for This Chapter
The Karpathy of the fall of 2022 is this book's baseline. His romantic imagination is the road from which every later coolness, every later soberness, every later "slop" would emerge.
If you want to hear how a person changes, you have to first hear the version that hasn't changed yet.
Sources
- Lex Fridman Podcast #333 (2022-10-29) — https://www.youtube.com/watch?v=cdiD-9MMpb0
- Sequoia AI Ascent 2024 (2024-03-20) — Stephanie Zhan × Karpathy — https://www.youtube.com/watch?v=c3b-JASoPi0 ; transcript at https://aletteraday.substack.com/p/letter-228-andrej-karpathy-and-stephanie
- Sequoia AI Ascent 2026 (2026-04-30) — https://www.youtube.com/watch?v=96jN2OCOfLs ; cleaned transcript at https://karpathy.bearblog.dev/sequoia-ascent-2026/
- Dwarkesh Podcast — AGI is still a decade away (2025-10-17) — transcript at https://www.dwarkesh.com/p/andrej-karpathy