The Key To Succeeding With AI in 2026

3 Mar 2026 · 1 min

Watch on YouTube

0:50

In this segment, the speaker outlines a methodology for achieving model accuracy comparable to directly querying real people. The approach hinges on three core components: (1) acquiring very high‑quality seed data—either a business’s own audience data or closely matching external data; (2) building a synthetic population of agents that represent the target audience; and (3) placing these agents into detailed decision‑making contexts, presenting them with additional information, and interpreting their responses. The process is continuously validated through extensive evaluations—over 40,000 runs comparing synthetic outputs to known real‑world responses. The pipeline relies heavily on commercial large language models rather than an in‑house LLM, using a combination of these models to generate the necessary outputs at each stage.

Chapters

How do you end up with a model that is as accurate as asking the people? Well, we basically combine 3 things. So the first is like very high-quality seed data. So we either use like businesses' own data on their target audience, or we'll find data that can match that, build out a whole load of agents. We'll build out basically a synthetic population. We can then build quite a complex process of how you place them into a decision-making context, how you present additional information against that audience, and then how we interpret the responses to get results that match what we'd expect to see from the real world. And we're always running evaluations against that. So we've run 40,000 to date on that, measuring our outputs against where we know real-world responses have been. A lot of those stages are powered by commercial models. So we use, you know, we don't have our own large language model, but we use a whole different combination of large language models in that process to generate that output.