May 8, 2026 · 41m · y-combinator
Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Lightcone Podcast, Y Combinator President and CEO Garry Tan discusses his return to active software engineering and how leveraging multi-agent AI tools, 'Tokenmaxxing,' and open-source frameworks like G-Stack enabled a 400-fold increase in developer leverage.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 4.6 Guest teaching 6.2 Guest disagreement 1.7 The partners pushing back 1.1
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 30:26 Defending lines of code metric against internet critics
Garry confronts critics of his viral productivity metric, arguing that measuring standardized logical lines of code objectively demonstrated a 400X output increase over baseline human engineering.
Hardest push from the partners ▶ 30:12 Host raises internet skepticism about lines of codeThe co-host raises the common engineering counterargument that raw lines of code fail to accurately measure developer productivity.
Biggest teaching moment ▶ 21:00 Reframing Markdown as compiled latent codeGarry dismantles the idea that Markdown is trivial formatting, explaining how plain English markdown functions as deterministic logic instructions in LLM latent space.
The partners hold their own ▶ 36:30 Host maps tokenmaxxing cost to SF founder rentThe co-host synthesizes the core investment thesis of token spend by drawing a direct parallel to the necessity of paying premium San Francisco rent for network density.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| The Origins of Garry's List and Political Reform | 4 | 5 | 1 | 1 | The host prompts Garry to discuss the origin story of Garry's List and political reform in SF. Garry explains the political and educational motivation behind creating the site, recounting the history of Posterous and building software to solve social problems. | |
| Building an Agentic Newsroom with Claude Code | 5 | 6 | 1 | 1 | The host points out that Garry's List functions as an automated investigative journalist rather than just a publishing tool. Garry agrees, detailing how agentic retrieval and deep research pipelines replicate a month of human research for a few dollars. | |
| "Boil the Ocean" and the Tokenmaxxing Mindset | 3 | 7 | 2 | 1 | Garry delivers a deep exposition on the 'boil the ocean' philosophy of tokenmaxxing, contrasting human incremental research with cross-referencing dozens of sources. The tone is highly educational and collaborative. | |
| The Birth of G-Stack and Automated QA Testing | 4 | 6 | 1 | 1 | The host bridges to the creation of G-Stack, and Garry shares his workflow insights on using ASCII diagrams and achieving high test coverage to avoid generating vibe-coding slop. | |
| Meta-Prompting and the 10X CEO Plan | 4 | 6 | 1 | 1 | Garry describes meta-prompting and the CEO Plan skill inspired by Brian Chesky's 10-star framework. The hosts listen as Garry shows how simple prompts unlock high-leverage model behavior. | |
| Multi-Agent Workflows: Conductor, Playwright, and Codex | 5 | 7 | 2 | 1 | The host asks about Garry's daily multi-agent workflow. Garry breaks down orchestrating Conductor, Playwright for automated QA, and delegating hard bugs to Codex while using Claude Code as the product planner. | |
| Maintaining Human Agency and Product Taste in AI | 3 | 6 | 2 | 1 | Garry argues that human product taste and agency remain indispensable in agentic software. He dismisses fully autonomous zero-human coding claims in favor of keeping humans firmly in the loop. | |
| Thin Harnesses vs. Fat Skills: Markdown as Code | 5 | 7 | 3 | 1 | The host references Garry's post on thin harnesses and fat skills. Garry pushes back against online critics mocking Markdown, arguing Markdown operates as executable declarative instructions for probabilistic models. | |
| The Homebrew Era of AI and Self-Healing Agents | 6 | 5 | 1 | 1 | The host adds technical color regarding how self-healing loops work when Claude Code repairs broken OpenClaw setups. Garry compares the current era to the Homebrew Computer Club and kit car Ferraris. | |
| G-Brain and Building a Personal Knowledge Base | 4 | 7 | 1 | 1 | Garry explains how building G-Brain required learning vector embeddings and hybrid RRF search to overcome grep's context window limitations, emphasizing how practical problem-solving accelerates learning. | |
| The 400X Developer Metric and Lines of Code Controversy | 5 | 7 | 4 | 2 | The host brings up the controversy around measuring productivity by lines of code. Garry directly addresses the debate, explaining how standardized logical line counts validated his 400X productivity leap over 2013 human benchmarks. | |
| The Fight for Personal AI vs. Corporate Feeds | 6 | 6 | 2 | 1 | The host observes differences in model performance across harnesses. Garry reframes the issue around personal computing versus corporate-controlled algorithmic feeds, urging developers to write their own prompts. | |
| Tokenmaxxing as Capital Expenditure: The SF Rent Analogy | 6 | 6 | 1 | 1 | The host introduces the SF rent analogy for token spending, explaining that token expenditure should be treated as essential high-utility capex. Garry reinforces this with YC core advice on living in the future. |
Statements from this episode (22)
Tan: Returning to Coding with AI Yields 400x Output
“It was 13 years of not coding and then suddenly, boom, I'm doing about 400 X the amount of work that I was that year. The last time I was even sort of like two thirds of the time writing code.”
Garry Tan: Twitter acquired Posterous for approximately $20 million
“Posturus was dead simple blogs by email. It grew to be a top 200 website on the internet, and then Twitter ended up buying it for about twenty million dollars, so that was sort of like my first bag, really.”
Tan: Building Garry's List with Claude Code took $200 and five days
“The first time it took about, you know, four million dollars and, you know, six or seven people and about a year and a half. And then the second time it, you know, took about, I don't know, a hundred grand and two people, me and my co-founder Brett Gibson, who…”
Tan: $5 to $10 of Opus calls replicates human research and annotation
“Basically, for the equivalent of, like, Five or 10 dollars of opus calls. I mean, I would estimate that it does the work of like, you know, a real human being that would have to like go painstaking through dozens of articles, read entire books about certain su…”
Garry Tan: Tokenmaxxing will permeate all knowledge work across society
“I think now it's going to permeate every part of society. Like, every thing that we would call knowledge work could be token maxed.”
Garry Tan: Tokenmaxxing will not replace humans because human agency remains essential
“I don't think that it means that we're going to get rid of people. I think it means that people need to still supply the agency.”
Garry Tan: Prompting Claude for ASCII diagrams dramatically improves code completeness
“One of the things I discovered is sometimes Claude would just get confused and like write bugs or not be complete. But once I started saying, actually, before you start your work, make an ASCII diagram of all the data flows, all the inputs and outputs. What ar…”
Garry Tan: 80% to 90% test coverage is optimal for AI coding
“I've since learned that a hundred percent is probably too much. Like hitting 80 to 90% is usually the best practice at this point.”
Tan: Asking for 10x value at 2x effort unlocks ambitious products
“One is what is the 10 X check? What is more ambitious and delivers 10 X more value for only two X the effort. Right. And so for whatever reason, coming out of latent space, this helps the model, like really visualize”
Tan: Submitted 13 pull requests in 48 hours via AI workflow
“So I dropped like 13 PRs in the last 48 hours.”
Tan: Claude models are very good but not the smartest
“Claude Code is ideal for the ADHD CEO, but once in a while, there's a, you know, Claude Code will just BS a bunch of stuff, like Claude models are very, very good, but like, they are not the smartest, it turns out.”
Garry Tan no longer uses Visual Studio, relying entirely on AI agents
“For like, I don't use Visual Studio at all. Like there's no reason to, like when I can talk to my agent and my agent can do this, right?”
Tan: Agentic engineering fails when developers put Markdown logic into brittle code
“All of the difficulty in energetic engineering today is when people try to do things that should be in markdown in code and it fails because code is brittle.”
Tan: Untested AI-generated code is 10x worse than human-written code
“Like if it's not tested and you're just throwing users in there, like it's slop. You know, 10 X worse than like human written code.”
Garry Tan: AI development is in its Homebrew Computer Club era
“This is homebrew computer club. You know, the moment when the Apple one came out, like The Apple one created by Steve Jobs and Steve Wozniak was a breadboard inside, like literally a wooden case hammered together with like nails and duct tape, you know? And if…”
Tan: We are entering the golden age of open source
“This is something that, you know, anyone can do actually, it's like this is why I think we're entering the golden age of open source.”
Tan: Professional engineers average only 30 to 50 lines of code daily
“If you look at the literature about software engineering, going back to like 2000. 1990. I mean, it's pretty clear that the average number of lines of code that a professional software engineer that's like tested and production ready, it's not like a hundred l…”
Tan: Technical people with taste benefit most from tokenmaxxing AI agents
“It's very significant for people who are technical because it actually raises the bar on like what you are capable of doing. Like all the people who are attacking me about lines of code, they particularly are, The people who are most likely to get wings if you…”
Tan predicts everyone on Earth will have a personal AI by 2027
“And then I guarantee you like this time next year, like everyone's going to be saying what you heard here first, which is like every single person on the planet will have their own personal AI.”
Tan: Users who do not write prompts fall below the API line
“Unless you have your own prompts and you can write it for yourself, like you are You know, below the API line for some PM or developer that is not you who like will not understand you will not understand your needs will not understand what you uniquely care ab…”
Tan: Founders should willingly spend $500 daily on AI tokens
“One of the key maxims for YC is You know, how do you find good startup ideas, live in the future and build what's missing? Right. And so this is a profound version of that where all you have to do is commit your brain to look at, you know, spending 500 dollars…”
Garry Tan: Tokenmaxxing turns builders into time billionaires via machine labor
“And then if you can token max, it's like, I mean, you can buy millions of years of consciousness of machine consciousness. Now I can be a time billionaire. It's not, you know, my own time. It's the time of a machine, like doing work for me and like the human e…”