Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers

Defining Question & Core Metaphor [0:00-0:35]

The defining question of the era is framed as: 'will you have control over your own tools or will your tools have control over you?' Using Claude Code is compared to 'driving a Ferrari'—it is exhilarating and can figure things out you would never think a machine could, but it will also break down when you most need it, requiring you to be a mechanic.

Episode Introduction & Gary Tan's Return to Building [0:47-1:26]

This special episode of 'the light cone' covers how Gary Tan got back to building after a multi-year hiatus as an investor. In the last couple of months, he shipped hundreds of thousands of lines of code and built popular open-source projects that went from nothing to more than 100,000 stars on GitHub, all while running YC full-time.

Gary's Personal Shock & Productivity Leap [1:27-1:37]

Gary says he is 'relatively shocked myself' and 'amazed as well.' After 13 years of not coding, he is now doing about 400x the amount of work he was doing the last time he was coding two-thirds of the time.

The Origin Project: Gary's List [1:46-2:30]

The project that started it all was 'Gary's list,' motivated by a desire to bring together people who believed in his vision for California. He started a 501c4 (now a C3 and a PAC) as a way to bring people together, learning from years in San Francisco politics that bringing people together is powerful.

The Algebra Issue & Personal Motivation [2:32-3:21]

A key issue is that it is still very hard for a seventh or eighth grader in San Francisco public schools to take algebra. Gary states that if he hadn't had that opportunity, he would not have studied engineering at Stanford or been able to do any of these things.

Posterous History & Rebuilding [3:24-4:32]

Gary built Posterous, his first YC startup, in 2008. It grew to a top 200 website and was bought by Twitter for about $20 million. After Twitter shut it down, Gary rebuilt it as Post Haven. In January of the current year, he wrote it a third time. The first build took $4 million, six or seven people, and a year and a half. The second took $100,000 and two people in three months. The third took $200 (Claude Code Max account) and five days for a full-featured blog platform.

Capabilities of the New Platform [4:36-5:49]

The new platform includes full RAG, full agentic retrieval, the ability to read all of the internet, every tweet, recursive crawl, and deep research of any topic. It can ingest the internet, see all arguments for and against, and craft incredibly detailed reports. This was inspired by an earlier Light Cone episode about agentic systems with Jake Heller.

The 'Boil the Ocean' Philosophy & Token Maxing [6:17-8:47]

For $5 or $10 of Opus calls, the system does the work of a real human being who would painstakingly go through dozens of articles. The key insight is to think about what a human would do with the context given. The philosophy is 'boil the ocean'—what if you absolutely boiled the ocean? The goal is to not settle for one source when you can get 20 sources, cross-reference them, and feed all that context into a core prompt. This concept of 'token maxing' is the coolest thing you can do now and will permeate every part of knowledge work.

The Origin of GStack [10:07-11:10]

Gary did not plan to make GStack. He realized he was doing the same things over and over again and got sick of typing the same thing. He typed into Apple Notes all the things he found himself writing over and over again into Claude Code. One key practice was asking Claude to make ASCII art diagrams of all data flows, inputs and outputs, user flows, and error messages.

Testing Philosophy & 'Vibe Coding' [11:14-11:53]

When writing code himself, Gary would always do the minimum amount of testing because it was not fun. He hit the same issues as everyone else who starts 'vibe coding': 'this is slop. It's not working that well. It works fine for the 80% case, but if any users actually touch it, it starts falling over.' He realized he could get to 100% test coverage but has since learned that 80 to 90% is usually the best practice.

The Skills: Plan-Review, CEO Plan, & Office Hours [11:55-14:10]

The first version of the skill was 'plan-ge-review.' The 'office hour skill' simulates what YC does when working with a company. Gary posted the proto-skill and it went viral. He made a more expansive version called 'the CEO plan,' which uses metaprompting and is based on the concept of a '10-star experience' from Brian Chesky. It includes a '10x check': what is more ambitious and delivers 10x more value for only 2x the effort?

GStack Workflow & Daily Setup [14:10-16:58]

GStack started as a way to make skills. Gary uses 'Conductor' for his daily workflow, which is how he has been shipping hundreds of thousands of lines of code a month. His workflow: use the CEO skill, use a skill to make it super well tested in plan mode, then click approve for Claude to do the work. He got sick of manual QA and found Claude Code's Chrome MCP very slow, so he wrapped Microsoft's Playwright. GStack now has a CEO, a designer, a developer experience person, design tools, and Plange.

Claude Code vs. Codex [17:00-18:19]

Gary learned from YC alums that many people prefer Codex. He discovered that Claude Code is ideal for the ADHD CEO, but once in a while it will BS a bunch of stuff. For a problem that's much crazier, you need the '200 IQ nearly nonverbal CTO,' so you can call in a friend, which is what '/codex' is. The GStack skill '/codex' takes the repo and runs Codex in a command line prompt to find all the problems and bugs and reports it back to Claude Code.

The Human in the Loop [18:21-20:31]

GStack relies very heavily on 'ask user question.' This is where the human needs to supply their understanding of what's going on. Gary says there is not really a substitute to that and it would surprise him if someone truly made a thing that could just make software without the human in the loop. He never wants to be entirely out of the loop; he just wants the machine to do the stuff he doesn't want to do. QA is a good example: the modern GStack has a long-lived HP demon with 70 commands as a CLI called 'Browse.'

Thin Hardness & Fat Skills Philosophy [20:54-23:37]

This philosophy is about how to token max. The name came from YC partner Pete Kumin. The idea is that a 'harness' is the core loop (user input -> LLM -> tool calls). The question is: why would we build that? What we should be spending all our time doing is thinking about what markdown should there be? Things that are deterministic or a real action should be in code. The difficulty in agentic engineering today is when people try to do things that should be in markdown in code and it fails because code is brittle. The magic for an engineer now is figuring out how much is in LLM land and how much is in code land.

Final Advice on Testing & 'Boil the Ocean' [23:39-24:09]

Combine the markdown/code philosophy with getting 80 to 90% tests. If it's not tested and you're just throwing users in there, it's slop—10x worse than human-written code. You need to figure out what's going on in latent space and deterministic space and also make sure that it's unit individually tested and then the integration is tested. This all goes back to the 'boil the ocean' philosophy.

Current State of AI Coding Tools [24:11-26:53]

The machine doesn't care, it'll just do it. You can just zap the rocks more and get to 90% test coverage. OpenClaw is not quite perfect but is 95% there. The current state is compared to the Homebrew Computer Club moment when the Apple 1 came out—a breadboard inside a wooden case. This is the 'kit car Ferrari phase.' To get something running, relatively smart technical people have to spend two or three hours and maybe $500 to $1,000 in tokens and cloud. The evolution from Stack Overflow to ChatGPT to Claude Code is described: with Claude Code, you realize you don't need to copy and paste anymore; it just executes and runs the code.

The 'Lines of Code' Controversy [30:10-33:26]

Gary still believes lines of code is an important measure. After doing a logical lines of code strip-down, his rate actually went up to 400x his 2013 rate. The surprising finding was that the logical lines of code analysis knocked down the amount of lines of code he was writing in 2013 by 70%. The 400x figure came from comparing his current output to his historical rate of 14 lines per day. Gary believes this is very significant for technical people because it raises the bar on what you're capable of doing.

Model and Harness Performance [33:28-34:29]

Experiences vary dramatically depending on the models and the harnesses. One speaker notes that any semi-complicated programming task through their OpenClaw agent just kind of fails, even with the same model as Claude Code. The speaker predicts that next year everyone will be saying what you heard here first.

The Future of AI (Personal AI vs. Corporate AI) [34:32-35:53]

The prediction is that every single person on the planet will have their own personal AI. Two possible futures: one where we have our own AI with our own data and integrations, and one where it is corporate controlled like a Facebook feed. The most powerful idea was the personal computer revolution, and we're about to go through exactly that same shift with personal AI. The defining question is: will you have control over your own tools or will your tools have control over you?

Token Usage and Cost (The 'Token Maxing' Paradigm) [36:00-37:37]

A disconnect is that many capabilities require the latest and most expensive models. The new way of getting the ASI AGI moment for building requires burning lots of tokens and the whole token maxing paradigm. Token maxing is compared to San Francisco rent: it's so expensive to not live there. The advice for YC founders is to spend as much as you can to get the most utility out of it.

Personal Experiences and Workflow [27:18-30:07]

Gary created 'GBrain' after meeting the other speaker and having Peter on the show. He spent one weekend to check out OpenClaw. He fell into a rabbit hole with Conductor and had GStack built into conductor already. Gary describes learning about vector embedding and hybrid RRF and chunking while building an agentic newsroom for Gary's list. He states, 'this is why I think we're entering the golden age of open source.' Gary confirms he is basically back to sleeping 4 hours and coding 20 hours a day.

Time Scarcity and Token Maxing [38:08-39:52]

A question is posed: did being CEO of Y Combinator help because your time is so scarce you had to figure out how to write hundreds of thousands of lines of code with spare minutes? Gary responds that he envies time billionaires but his personal philosophy is that he is in a crazy rush. By token maxing, you can buy millions of years of consciousness of machine consciousness. He can be a time billionaire not with his own time, but with the time of a machine doing work for him.

Concluding Thoughts [39:55-41:06]

A powerful moment for Gary was sitting with Boris Churney and hearing him say 'our team doesn't write a single line of code.' Gary realized that he can do that and that the people watching are not different. He states, 'if I just open a prompt, we have the same prompt. We have the same MacBook Pro. And there's nothing that stands between me or you or any of us from drawing on millions of years potentially of tokens to serve humanity.' A concluding quote: 'You could have infinite time by borrowing the time from the machines.'