Agent + tool = work done?
Agent Harness Reliability
Give an agent a tool that knows how to find the right answer, then see what can still go wrong — and how much a better harness may help (or not).
Small personal lab exploring AI and tech through experiments.
Three fields of exploration
Work
Exploring how AI and software perform and augment knowledge work.
What happens after repeated translations?
Translation Loops
Send the same short French texts through repeated translation loops across languages and models, and watch what changes, what survives and where things go wrong.
Regulation
Exploring tools and methods for financial regulatory research.
How should AI-assisted regulatory research work?
regul.tech
Exploring search, retrieval, sources and AI assistance for financial regulation.
Early-stage prototype · limited access
Maths
Exploring how language models reason, calculate and use tools.
Can an LLM (reliably) find the next prime?
Next Prime
Test prime search in raw, prompted and tool-assisted modes.
Can ChatGPT count?
LLM Calculations
Compare raw, prompted and tool-assisted arithmetic across models.