I'm a standup comedian in New York, and I do open-source research in machine learning: program synthesis, LLM-guided coding and agentic software engineering.
The Library of Comedy Show, where we give away books
The live side of the archive. The October 2025 show, photographed by Hakim Dhouib.
Comedy T-Rex
Mary’s, 146 Orchard St. 9 photos by Yad Konrad.
What My $200 Codex Plan Was Actually Worth
What My $200 Codex Plan Was Actually Worth
Seven weeks of Codex logs lined up against my weekly limit, the new Pro tiers, and the 62,500 credits that showed up the night after DevDay: what t...
I perform standup around New York, at clubs like New York Comedy Club. I created and produce Baklava Comedy, live stand-up with fresh baklava at the door, and I built Library of Comedy, the archive, and run its show, where we give away books.
Open-source research in program synthesis, LLM-guided coding and agentic software engineering, since Description2Code with OpenAI in 2016 and Hopper in 2017. This year, with the Sutro Group, coding agents I ran reimplemented 53 of Hinton's experiments and 58 of Schmidhuber's papers in pure NumPy.
Systems where coding agents do most of the building: Chimera, a coding-agent framework built by coding agents; staxtrace, which watches them run; Novalis, a terminal built for them. Lessons on working this way at AgenticEngineer.ING.
Program synthesis took me through verification, type theory, homotopy type theory and algebraic topology, and back out with tools: Erdős Navigator, Claude Code on 1,217 of Erdős's problems, and Bourbaki, an agent for mathematical reasoning and proof with SymPy and Lean 4.
Twenty years of building software, since GTA mods in 2006. Two companies started, CTO twice, Head of ML twice. CongruentAI, Roby, On Deck, NEAR Atlas at the NEAR Foundation, and KayOS, started while I was building My Pixel Pilot, with 13 production AI platforms. At 20, Microsoft's Imagine Cup in Iraq and the worldwide finals in Sydney.
I photograph comedy rooms around New York and the people who play them, and I document other comics' work: 100 photos from 11 shows on the site so far, every one shown whole.
Penetration testing, LLM red teaming and application security. 158 attack prompts against gpt-oss-20b in Kaggle's 2025 red-teaming competition, and a pentester's guide to the OWASP Top 10 for LLMs.
Stories from the shows, the rooms and the people, on the blog here. The technical work at yad.codes, with the numbers pulled from the session logs, and program synthesis at programsynthesis.pub.
Baklava Comedy
The standup show I created and produce. Live stand-up with fresh baklava at the door.
Library of Comedy
A comedy archive tucked into a corner of the internet, and its live show, where we give away books.
Shows I've shot
I photograph comedy rooms around New York and document other comics' work.
Hinton Problems and Schmidhuber Problems
53 of Hinton's experiments and 58 of Schmidhuber's papers reimplemented in pure NumPy by coding agents I ran, with the Sutro Group.
Description2Code and Neural Engineers
OpenAI's Description2Code request for research, 2016: a dataset of programming problems and a first neural model that writes code from a description.
Hopper
An AI coding assistant I built in 2017, before GPT and Copilot.
staxtrace
Local-first observability and memory for coding agents: sessions from 20 providers in one local store, cost analytics, time-travel playback, a memory your agents query mid-task.
Erdős Navigator and Bourbaki
Program synthesis took me through verification, type theory, homotopy type theory and algebraic topology, and back out with two math agents.
My Pixel Pilot
Task and project ops run by LLMs: self-hosted, bring your own AI, import everything from Jira, Trello or Monday. I was building it when I started KayOS, the company: 13 production AI platforms and an AI coding IDE.
Coding agents on real codebases, machine learning and data, security, a set or a shoot: tell me what you have in mind.