Key Moments
Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers
Y CombinatorKey Moments
One person can now do the work of 400 engineers using AI agents by "tokenmaxxing," but controlling your tools is paramount, like being a mechanic for a high-performance Ferrari.
Key Insights
Gary Tan shipped hundreds of thousands of lines of code and built popular open-source projects, going from nothing to over 100,000 stars on GitHub, in just a few months while running YC full-time.
The third iteration of Gary's blog platform took about $200, his Claude Code Max account, and approximately five days to build.
Tokenmaxxing, or spending more on LLM tokens to "boil the ocean" with information, is presented as a way to achieve more complete and representative outputs, akin to getting 20 sources instead of just one.
The GStack system, which emerged from Gary's need for reusable skills, relies on different AI personas (CEO, designer, developer experience) and testing frameworks like Playwright for automation.
The average professional software engineer produces about 30-100 lines of tested, production-ready code per day, whereas Gary's current AI-assisted output, after stripping for logical lines of code, was 400x his 2013 rate.
Controlling your AI tools is framed as crucial, with a choice between personal AI that you control and corporate-controlled AI, drawing a parallel to the personal computer revolution.
The 400x Engineer: Gary Tan's AI-Powered Building Spree
Gary Tan, while leading Y Combinator, achieved a remarkable feat: shipping hundreds of thousands of lines of code and creating open-source projects that garnered over 100,000 GitHub stars in just a few months. This level of output, previously unimaginable for someone with his demanding role, was made possible through advanced AI coding agents and a philosophy termed 'tokenmaxxing.' Tan describes this period as being "back to building," shocked by his own productivity, which he estimates is 400 times greater than when he was actively coding full-time thirteen years prior. This transformation highlights a new era where individual developers, empowered by AI, can tackle projects that once required large teams, challenging conventional notions of software development capacity.
From Posterous to Gary's List: An Evolution in Development Cost and Speed
The journey to Gary's current productivity began with revisiting past projects. His first YC startup, Posterous, a "blogs by email" platform, grew to be a top 200 website and was acquired by Twitter. Years later, he rebuilt it as Post Haven. In January of the current year, he rebuilt it a third time. The contrast in development effort is stark: the first iteration cost $4 million and took a year and a half with seven people. The second iteration required $100,000 and two people over three months. The most recent iteration, Gary's List, built using Claude Code Max, cost approximately $200 and took about five days. This latest version is a full-featured blog platform with advanced capabilities like retrieval-augmented generation (RAG) and agentic retrieval, allowing it to crawl and research extensively, producing detailed, sourced articles on various issues, particularly those concerning California politics and education.
Tokenmaxxing: "Boiling the Ocean" for Deeper Insights
A core concept discussed is 'tokenmaxxing,' which involves maximizing the input to AI models by feeding them vast amounts of context to achieve more comprehensive and accurate results. Rather than settling for limited information, the approach encourages "boiling the ocean" – providing the AI with everything it could possibly need. This is akin to a human researcher consulting dozens of articles and books. For example, to generate an article on a complex issue, an AI empowered by tokenmaxxing would ingest numerous sources, cross-reference them, identify areas of agreement and disagreement, and feed this rich context into its core prompt. This method aims to create outputs that are more representative of reality, going beyond superficial understanding derived from a single source or headline. The philosophy extends beyond writing to all knowledge work, suggesting that by "zapping the rocks harder" with more tokens, one can achieve significantly better outcomes, even if it incurs higher costs.
GStack and Agentic Workflows for Enhanced Productivity
Gary Tan developed GStack as a solution to repetitive tasks, consolidating frequently used prompts and workflows into reusable "skills." This began with simple collections in Apple Notes and evolved into a structured system. A key insight from building Gary's List was the importance of thorough testing. Initially, Gary minimized testing to focus on new code. However, encountering issues with the AI-generated code led him to emphasize test coverage. He learned that while 100% test coverage might be excessive, aiming for 80-90% is best practice. GStack integrates various AI personas, such as a CEO (embodying a "10-star" product vision), a designer, and a developer experience specialist, along with tools like Claude Code, Codeex, and Playwright for automated testing. This allows for sophisticated workflows, including simulating executive reviews and automating quality assurance, significantly accelerating the development cycle.
The Dichotomy of Control: Personal AI vs. Corporate AI
A central theme is the question of control: "Will you have control over your own tools, or will your tools have control over you?" The future of AI, according to the discussion, should be personal and user-controlled. This parallels the personal computer revolution, where individuals gained agency through accessible technology. The alternative is corporate-controlled AI, where users are subject to opaque algorithms and business models, similar to a social media feed. The importance of writing one's own prompts is highlighted as essential for maintaining this control, ensuring AI serves unique needs and goals rather than those dictated by third parties. This independence is seen as critical for individuals to leverage AI effectively without becoming subservient to it.
The Evolution of Developer Experience: From Stack Overflow to Agentic Collaboration
The conversation traces the evolution of developer tools. Stack Overflow was once a revolutionary resource, providing answers to programming problems. ChatGPT enhanced this by offering interactive assistance, but the core workflow remained similar: asking questions, copying code, testing, and iterating. Claude Code and agents like OpenClaw represent a significant leap, moving beyond copy-pasting to direct execution and agentic collaboration. Tan notes that while OpenClaw can be exhilarating, like driving a Ferrari, it requires a "mechanic's" understanding to fix when it breaks down. However, even when one agent (like OpenClaw) has issues, another agent (like Claude Code) can be deployed to fix it. This creates a dynamic where brittleness is acceptable if there's an agent to maintain it, fundamentally changing the nature of software development and problem-solving.
Personal AI and "Time Billionaires": Borrowing Machine Consciousness
The concept of becoming a "time billionaire" emerges as a profound outcome of AI advancement. By leveraging AI agents and tokenmaxxing, individuals can effectively "borrow" vast amounts of computational time and consciousness from machines. This allows them to achieve highly accelerated workflows, akin to Gary Tan's 400x increase in output. The discussion emphasizes that this isn't about replacing humans but augmenting them, particularly those with "taste and design and product feedback." The ultimate goal is to utilize AI to focus on what truly matters – building valuable products, working on important causes, and developing personally – by offloading the tedious or repetitive tasks to AI. This allows individuals, especially founders and builders, to maximize their own limited time by drawing upon the near-infinite capacity of AI.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●People Referenced
Maximizing AI Productivity (Token Maxing)
Practical takeaways from this episode
Do This
Avoid This
Project Development Effort Comparison
Data extracted from this episode
| Project | Resources (Cost/People) | Time |
|---|---|---|
| Posterous (1st time) | $4 million, 6-7 people | 1.5 years |
| Post Haven (2nd time) | $100k, 2 people | 3 months |
| Third Iteration | $200, Gary Tan's Claude Code account | 5 days |
Common Questions
'Token maxing' refers to maximizing the use of AI models by spending generously on tokens to achieve the highest possible output quality and depth. It's about burning tokens to 'boil the ocean' for research or complex tasks, treating it as an investment rather than a cost.
Topics
Mentioned in this video
Mentioned as a tool for deep research.
An AI model that Gary Tan found to be less efficient in managing context compared to other models for his knowledge base.
An AI tool that Gary Tan uses, particularly for its '200 IQ nearly nonverbal CTO' capabilities, contrasting with Claude Code's 'ADHD CEO' style. It's integrated into GStack as '/codex'.
An extension for PostgreSQL that Gary Tan plans to use for his OpenClaw RAG system.
The first project Gary Tan built upon his return to coding, focused on issues like math education in San Francisco public schools and acting as a blogging platform with investigative journalism capabilities.
A rebuilt version of Posterous after Twitter acquired the original and shut it down. Gary Tan later rebuilt it a third time.
Mentioned as a tool for deep research on the internet.
An API mentioned for performing research, particularly on X.
Gary Tan's project developed out of a need to automate repetitive tasks he was performing with AI tools, evolving from simple notes to a suite of skills.
Where Gary Tan initially compiled repetitive tasks before moving them to Cloud Code for automation.
A platform used for building AI agents, which Gary Tan integrated GStack into.
An alternative testing framework that Gary Tan considered using to improve the speed of QA processes, moving away from the slow Claude Code MCP.
A project Gary Tan has been working on, building upon GStack and involving interactions with Peter Yang.
A database system with PG Vector extension that Gary Tan intends to use for his OpenClaw RAG system.
An interactive AI tool that improved upon Stack Overflow for programming help, but still involved a copy-paste workflow.
An AI tool described as exhilarating like a Ferrari, but requiring mechanic-like expertise to fix when it breaks down. It's presented as a powerful but potentially brittle system.
Used as a metaphor to describe OpenClaw's exhilarating performance and the need for technical expertise to maintain it.
Mentioned as a platform where Gary Tan's return to building is known and where his previous startup Posterous was acquired.
The platform where Gary Tan's open-source projects have gained over 100,000 stars.
Gary Tan's first YC startup, a blog-by-email platform that grew to a top 200 website and was acquired by Twitter for $20 million.
A platform related to agentic systems, created by Jake Heler, which inspired Gary Tan's method for generating journalistic long-form articles.
Mentioned as something Gary Tan consumed in his younger, more intense coding days (25-year-old self).
A Q&A website for programmers that was once considered amazing but is now seen as less interactive and efficient than modern AI chat tools for coding help.
Mentioned alongside OpenClaw as potentially not yet fully realized or requiring significant work, but expected to improve.
The speaker discusses his return to building after a hiatus, shipping hundreds of thousands of lines of code and building open-source projects while running YC.
Gary Tan's co-founder for the second iteration of Posterous (Post Haven) who now runs Initialized.
Creator of Case Text, who discussed agentic systems in a previous episode of The Lyone, influencing Gary Tan's approach to building journalistic articles.
CEO of Airbnb, referenced for his concept of a '10-star experience' in product design, which Gary Tan adapted into an AI skill ('CEO plan').
An exemplary engineer who inspired Gary Tan by stating his team writes no code, emphasizing the power of directing AI agents.
More from Y Combinator
View all 589 summaries
22 minPaul Graham, Founder of Y Combinator, Live from Stockholm
38 minRecursion Is The Next Scaling Law In AI
41 minHow to Build the Future: Demis Hassabis
40 minReplit's CEO On The Only Two Jobs Left In The Company Of The Future
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free