Henry Zhang

Zhang describes working on inference and reinforcement learning at SpaceXAI, after earlier engineering work at Cohere and Amazon. His website also describes studies in computer science and biochemistry at the University of Michigan. His public projects span agent workflows and small language-model implementations.[1][3][2]

Compose agents with ordinary Python

SwarmCore lets developers join agents in sequence or run them in parallel, then pass their results to another agent. Each step receives the previous step’s full output; earlier outputs are summarized to control context size. An agent can ask for an earlier full result when needed. Python type hints and docstrings supply tool schemas. The approach makes a workflow visible in Python code while leaving model choice to a compatible provider.[5]

Make a language-model program small enough to inspect

For femtoGPT, Zhang compressed a small Python GPT implementation and represented the bytes as an integer. His write-up explains how to recover and run the code by reversing that process. It is an encoding and implementation experiment, not a claim that a competitive trained model or its knowledge fits inside a short number. The article describes the result as a probable prime because its test is probabilistic.[6]