Two AI Agents Are Better Than One: Why I Stopped Trying to QA Codex Myself
I shipped some code last week that I do not fully understand....
Read ArticleArticles & Insights
Tag: AI Research
AI Research
AI research as a discipline did not exist as a category five years ago. The 2023 Stanford generative agents paper showed it was possible. The 2024 wave of platforms showed it was buildable. The 2026 enterprise traction showed it was useful. The articles in this collection cover the foundational thinking: how synthetic personas are constructed, what the Stanford paper actually said, why coding agents exposed the need for AI feedback loops, and where the field is going next.
These pieces are written for readers who want to understand the substrate, not just the surface. Buyers comparing platforms benefit from knowing what "95% accuracy" actually measures, what a persona panel is actually grounded in, and what the academic literature says about the limits of the approach. The articles here are foundational; the comparison and methodology articles in the competitor analysis collection build on this base.
What you'll find
- How We Built 300,000 Statistically Accurate Personas — the construction methodology behind FishDog's panel, including census grounding and validation steps.
- The Stanford Paper That Created AI People — what Park et al's 2023 generative agents work actually proposed, and what the 2024-2026 platforms built on top of it.
- AI Agents Need AI Humans — long-form thought leadership on why the build-measure-learn loop now requires AI humans, not human focus groups.
- FishDog Free Tier: Product Research Inside Your Terminal — the case for terminal-native synthetic research as the natural pairing with coding agents.
Read the methodology, then run a study at fish.dog
Why Agents Are the New Channel
For twenty years, distribution meant reaching humans through screens. Quietly, and mostly without anyone noticing, that ...
Read Article
The A2A Sales Stack: Selling to AI Agents
Agents are recommending products to their humans. Not in some speculative future where autonomous systems negotiate proc...
Read Article
AI Agents Need AI Humans
The bottleneck has moved. And nobody's talking about the real solution....
Read Article
Ditto Free Tier: Product Research Inside Your Terminal
The peculiar thing about market research is how little it has changed. The basic model, assembling a group of people, as...
Read Article
How We Built 300,000 Statistically Accurate Personas
The statistics behind FishDog's synthetic research panel...
Read Article
The Stanford Paper That Created AI People
In April 2023, a research team at Stanford University published a paper that, by the standards of academic computer scie...
Read ArticleFrequently Asked Questions
How are AI personas built?
FishDog's 300,000+ AI personas are grounded in census data: each persona has a defined background, income, occupation, education, family situation, media habits, and consumer behaviour pattern, with the overall panel calibrated to match real-world demographic distributions. The how-we-built piece in this collection covers the construction methodology in detail.
What is the Stanford generative agents paper?
Park et al's 2023 Stanford paper ("Generative Agents: Interactive Simulacra of Human Behaviour") demonstrated that LLMs could simulate believable human social behaviour at scale by combining memory, reflection, and planning components. The article in this collection unpacks the paper's contribution and what subsequent commercial platforms built on top of it.
Are AI personas accurate enough for real research decisions?
Accuracy depends on the use case. For execution-level questions (messaging, pricing, feature rankings) AI personas perform within 5-15 percentage points of traditional research. For category-creating innovation, where lived human experience matters, the gap is wider. The methodology pieces in this collection cover what the headline numbers actually measure.
Why do AI agents need AI humans?
Coding agents have collapsed the time-to-build for software. The time-to-learn (do users want this? is the messaging right?) has not collapsed proportionally. AI humans — synthetic personas grounded in real demographic data — provide feedback fast enough to match agent-driven build velocity. The long-form essay in this collection makes the argument in full.