She shipped Terraform Cloud features in the morning and shipped cartoon explainers in the evening. Now she writes the checks. Yoko Li is the a16z Partner who still commits code - and she wants to know which LLM plays the best Tetris.
There is a specific kind of VC that makes founders relax. Not the kind who rattles off TAM math or drops brand names in the first three minutes. The kind who asks about the database schema. Yoko Li is that kind.
A Partner at Andreessen Horowitz focused on AI and infrastructure, she came up through the engineering trenches: AppDynamics writing frontend tools in React and Angular, Transposit as a founding engineer who transitioned into product, then HashiCorp where she ran product for Terraform Cloud - the infrastructure-as-code platform that DevOps teams relied on to not accidentally delete production. When she joined a16z in mid-2022, she brought that operational scar tissue with her.
What makes her unusual is what she does with the nights and weekends. Instead of networking dinners, she ships open source projects. AI Town - a virtual settlement where autonomous AI agents wander around, have conversations, form opinions, and occasionally hallucinate in ways that make them more believable, not less - came from her own need to understand what multi-agent systems actually felt like from the inside. It became one of the most-starred AI demos on GitHub. MIT licensed. Deployable in minutes. Tens of thousands of developers used it as a starting point.
The future of AI user interfaces demands new companies - the incumbent tools were not built for a world where the interface itself can reason.
- Yoko Li, a16z PartnerHer investment thesis is simple to state and hard to execute: she looks for the foundational layer. The infrastructure that other infrastructure runs on. The tools that developers reach for before they reach for anything else. Resend (email delivery for developers), Mintlify (documentation rebuilt for the age of AI agents), Clerk (authentication that works the way developers actually think), Stainless (the API SDK layer), Inngest (workflow orchestration for AI agents) - every bet is a layer in the stack. The stack she has publicly called a trillion-dollar opportunity.
She is also, genuinely, a cartoonist. The @stuffyokodraws persona - active on Twitter and Instagram since January 2019 - posts drawings about tech, cooking, culture, and the occasional absurd observation about the startup world. The AI Explained series on ai-explained.yoko.dev turns machine learning concepts into illustrated panels. It is not marketing. It is just how she thinks.
The TetrisBench project was the clearest window into how her mind works. After playing Tetris 99 on her Nintendo Switch, she wondered: what would it feel like to play against an LLM? Not as a party trick - as a genuine evaluation of strategic reasoning under time pressure. She built the benchmark. She ran the models. She published the results. Gemini 3 Pro won with a 62% win rate by playing with minimal interventions and surgical patience. Other models showed different optimization patterns entirely. It is the kind of thing you build when you genuinely want to know the answer, not when you want a press release.
The Rice CS degree is load-bearing. Yoko Li's technical depth is not decorative. When she evaluates an infrastructure startup, she reads the docs the way a developer reads the docs - looking for friction, missing primitives, and the gap between what the README promises and what the API delivers.
She hosts and appears on a16z podcasts covering AI agents, the Model Context Protocol, the future of software development, and the evolving shape of developer tooling. When MCP emerged as a potential standard for letting AI agents connect to external systems, she was among the first at a16z to dig into its implications - hosting the MCP co-creator for a detailed technical conversation that went deeper than most.
The Fly.io engineering blog, which is not known for puff pieces, profiled her under the headline "How Yoko Li makes towns, tamagoes, and tools for local AI." The three words that are doing the most work in that headline: "and tools for local AI." Her local AI starter kit lets developers run private AI on-device for document search. Her AI-Raspberry-Pi-Starter-Kit connects a cat detector to a narrating LLM that sends text alerts. These are not demos. They are experiments with opinions embedded in the architecture.
The portfolio she is building reflects a clear conviction: the next decade of software will be written differently, deployed differently, and maintained differently - and the developers building that future need better primitives. She is trying to fund those primitives before everyone else figures out they need them.
Email communication infrastructure for developers. The modern alternative to legacy transactional email services, built API-first with developer experience at the core.
Developer tooling focused on the next generation of infrastructure primitives.
Documentation rebuilt for the AI era. Moving from static docs toward interactive, agent-aware knowledge bases that can answer questions and onboard developers automatically.
Authentication and user management designed the way developers actually think. Drop-in auth that removes one of the most painful early decisions a startup makes.
Workflow orchestration for AI agents. Reliable event-driven execution that handles the messy reality of long-running AI jobs without rolling your own queue system.
Generative AI for photography. Raised a $5.6M seed led by a16z to build AI tools that work at the intersection of creativity and computational photography.
The SDK layer. Generates clean, idiomatic API client libraries across languages - solving a problem every API company faces and almost none solve elegantly.
Rate limiting and caching infrastructure for AI applications. Serverless Redis and Kafka with a pricing model that actually makes sense for variable AI workloads.
An open source virtual settlement where autonomous AI agents wander, converse, and form social relationships over time. MIT licensed, deployable in minutes. Used by tens of thousands of developers as an AI agent starter kit. The hallucinations were a feature.
Open Source ▮ AI Agents ▮ MIT LicenseA Tamagotchi powered by LLaMA 2 7B - written entirely in JavaScript. No Python. No GPU server. A digital pet with a personality shaped by a language model, proving that local AI can run in environments developers already know.
Local AI ▮ JavaScript ▮ LLaMA 2A benchmark that makes LLMs play Tetris and evaluates their strategic reasoning. Started after a Nintendo Switch session. Gemini 3 Pro won with 62% win rate. Other models revealed entirely different optimization strategies. Published as a full analysis on the a16z blog.
LLM Eval ▮ Benchmarking ▮ ResearchDocumentation ingestion software for private, on-device document search. Runs entirely locally. Built to demonstrate that useful AI applications do not require sending data to a cloud API.
Privacy ▮ Local AI ▮ Developer ToolsBite-sized illustrated explainers for technical AI concepts at ai-explained.yoko.dev. Turns transformer architecture and attention mechanisms into panels that non-ML developers can actually parse.
Education ▮ Illustration ▮ Machine LearningA computer vision setup that detects cats, narrates what it observes using an AI language model, and sends text notifications. The kind of project that tells you more about the builder than the technology.
Computer Vision ▮ Hardware ▮ LLMShe reads documentation the way a developer reads documentation - looking for what is missing, where the API is awkward, and what assumption the author made that you are expected to already share.
The @stuffyokodraws persona predates the VC job by years. She draws about tech, food, culture, and whatever seems funny from the inside of the startup world. The illustrations are how she processes complexity.
When she wants to understand something - how different LLMs reason, how autonomous agents behave, how local AI feels from a user perspective - she builds a thing and runs it. The TetrisBench came from curiosity, not content strategy.
Years managing Terraform Cloud gave her a specific lens: she asks what the foundational layer is, who controls it, and what becomes possible once it is standardized. Every portfolio company is a layer in a stack she is mentally assembling.
The Model Context Protocol is not a household name yet. It will be. The protocol - which standardizes how AI agents connect to external systems like databases, APIs, and tools - is the kind of infrastructure investment that looks obvious in retrospect and invisible in the moment. Yoko Li started paying attention to it early.
Her podcast with the MCP co-creator went deeper than most coverage. Not "what is MCP" for a general audience, but the specific questions that matter for developers building on top of it: How do you handle authentication at the protocol level? What does trust look like when an agent is calling tools on behalf of a user? How does the protocol evolve as agents become more capable and more autonomous?
She has named a specific thesis about where the opportunity lies: the AI user interface stack needs to be rebuilt from scratch. Incumbent tools were designed for humans interacting with static interfaces. AI agents need something different - interfaces that can reason about context, that understand intent rather than just mapping inputs to outputs, that can communicate back to both the agent and the human simultaneously.
This is not a theoretical position. It is embedded in the portfolio. Mintlify is reimagining documentation for a world where the primary reader might be an AI coding assistant. Inngest is solving the orchestration problem that appears the moment an AI agent needs to do more than one thing in sequence without losing state. Clerk is rethinking authentication for an era where the entity authenticating might itself be an AI agent acting on behalf of a human user.
The "nine emerging developer patterns" essay she co-authored is essentially a map of this thesis. Version control for AI-generated code. LLM-driven user interfaces that reconfigure themselves based on context. Documentation layers specifically designed for coding agents to ingest. Each pattern points toward a company that needs to exist and, in some cases, already does - in her portfolio.
The key question Yoko Li keeps asking: "What is the assembly language of this era?" Not what the finished applications will look like - but what are the lowest-level primitives that everything else is built on, and who owns them?
She is currently seeing the field from a position few people occupy: an active technical builder who also controls capital. The open source projects she ships are not just intellectual exercises. They are research. They tell her what is hard, what is missing, and where the next generation of developer tools needs to go. The portfolio and the GitHub commits are the same inquiry, run in parallel.
"The future of AI user interfaces demands new companies - the incumbent tools were not built for a world where the interface itself can reason."
On the AI UI opportunity"We're still figuring out what the right abstractions are for AI agents - it's like being in the assembly language era of this stack."
On the current state of AI agent infrastructure"The best developer tools feel like superpowers, not scaffolding. You want the developer to feel like the tool disappears."
On developer experienceThe TetrisBench project started on a Nintendo Switch. She was playing Tetris 99 and wondered what it would feel like to compete against a language model. One curiosity spiral later, she had a rigorous multi-model benchmark with published results.
Her @stuffyokodraws persona on Twitter and Instagram - a cartoonist posting about tech, cooking, and startup life - predates her VC career by several years. The illustration work is not a brand exercise. It is just how she communicates.
AI Tamago is a Tamagotchi with a LLaMA 2 7B brain, written entirely in JavaScript. No Python, no GPU server required. She built it to prove a specific point: local AI can run in environments developers already know, with tools they already have.
The Fly.io engineering blog - which focuses on technical depth over PR-friendly narratives - profiled her under "How Yoko Li makes towns, tamagoes, and tools for local AI." Engineers write about people who build things. She builds things.
Her Raspberry Pi cat detector uses computer vision to identify cats, generates narration using a language model, and sends text alerts. It is a small project. It also reveals the full mental model: physical sensors, local AI processing, real-world notification. Stacked, not separate.