What we read, and why.
Short notes on published work in user simulation and self-improving agents. The research is other people's; the commentary is ours, and every note links to the original paper.
Park et al. · Stanford University and Google Research · arXiv:2304.03442 (2023)
Twenty-five language-model characters live in a small sandbox town with memory, reflection and planning wired together. The interesting result is that plausible social behaviour emerges from fairly simple machinery: remember what happened, reflect on it, plan accordingly.
Why we read it. It makes a clear published case that a simulated person is worth taking seriously as an evaluation instrument, and it is honest about the failure modes: memory retrieval that pulls the wrong thing, and behaviour that drifts toward the polite and agreeable. Both are exactly the failures we watch for in our own personas.
Reader's noteUser simulation
Wang et al. · NVIDIA, Caltech, UT Austin, Stanford and Arizona State University · arXiv:2305.16291 (2023)
An agent plays Minecraft with no human in the loop and no fine-tuning. It sets its own next goal, writes a program to reach it, checks the result, and keeps the programs that worked in a growing skill library it can call later. Capability accumulates in that library as reusable code, not as a longer prompt.
Why we read it. It is the same shape as the loop we care about: attempt, verify, keep what worked, reuse it next round. The paper is candid that progress depends almost entirely on the quality of the feedback, which is why we put evaluation before autonomy.
Reader's noteSelf-improving agents
