Founding AI Engineer at LayerLens.ai. I build evaluation systems for LLMs and agents. Before that, I studied math at USC and played Division I NCAA Men's Water Polo.
I'm interested in world models, machine agency, and agent-to-agent communication.
Garrett Allen
AI Engineer
San Francisco, CA
Recent Work
How do you pay people for information without making them gamble? A sketch.
Selected Projects
Evaluation platform for LLMs. Benchmarks and observability.
Infrastructure for agents to negotiate and sign contracts with each other.
Learning world models from Minecraft gameplay using JEPA.
Game environments for testing how well LLMs reason and plan.
Simulation environment for training agents to negotiate.
Favorites
Papers: Platonic Representation Hypothesis, A Path Towards Autonomous Machine Intelligence, IWM-JEPA, MindEye 2, o1 System Card, DeepSeekMath, SimpleQA, Voyager, Entropy SGD, Scaling and Evaluating SAEs.
Posts: Dealing with Daemons through Algorithmic Contracts, The Model Does the Eval, Why Tool AIs Want to Be Agent AIs, Jailbreaking Frontier Models (PRBO), Clever Hans, ATPAMI: Unsolved Technical Alignment Problems.
Books: UDL, Topology, Baby Rudin, Arbitrage Theory in Continuous Time, Programming Massively Parallel Processors, Neural Networks and Numerical Analysis.
Repositories: Mineflayer, V-JEPA, Verifiers, TAU-Bench.
