Snowglobe tests AI applications by generating simulated users with different goals and communication styles, then having them carry on conversations with the application. This explores behavior that a collection of isolated questions might miss, including how the system responds as an interaction develops.
The company's representative recounts a customer worried about offensive responses whose testing instead exposed excessive refusals of benign requests. The value of that example is diagnostic: simulation can help a team discover which behavior needs attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → before choosing safeguards. Its findings describe generated interactions, so they do not establish how frequently the same problem occurs among real users.
Terms in this piece · Glossary
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
videoWorld Models & General Intuition: Khosla's largest bet since LLMs & OpenAI
videoExtreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code or review — Ryan Lopopolo, OpenAI
videoInside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa
videoExo: Harnesses should see their own code and logs — Alex Krentsel, UC Berekeley / Google Research
