Aug. 12, 2026

215: Gavin Lorenzo: "Can AI Actually Understand Humans Without Human-Grade Self-Awareness?"

Show Notes

Gavin Lorenzo explains why AI struggles to truly understand people, not just store facts. He shares how Tellbloom builds a “third-party”-like model of a user by learning from passive interactions, then uses that context to improve responses. The conversation also covers testing, privacy, and the governance challenges raised by frontier models like Fable.

👤 About the Guest

Gavin Lorenzo is the founder of Tellbloom, working from Georgia with support from the Alabama Entrepreneurship Institute. A second-time founder, he built his way from an AI tutor for LSAT prep to a mission focused on helping AI understand humans. His current work includes a Chrome extension that analyzes how you interact with Claude or ChatGPT to personalize prompts more effectively.

🧭 Conversation Highlights

Tellbloom’s goal: maximize AI’s understanding of humanity, moving beyond “memory” into predictive behavior and intuition-like grasp.
Gavin’s origin story: building an LSAT tutor revealed AI’s limits in personalization and context formation.
How Tellbloom tests progress: A/B testing many “pipelines,” persona-based fine-tuning, and benchmarking before wide deployment.
Governance and safety: why model access, export controls, and safeguards matter, and why governance cannot be left solely to model companies.

💡 Key Takeaways

Personalization fails when it depends too heavily on a user’s ability to articulate themselves; passive observation is a better data source.
“Understanding” is hard to define, but Gavin frames it as the ability to accurately predict and adapt in a way that feels human.
To make squishy human modeling testable, you need repeatable suites, personas, and benchmarks.
As models get more powerful, regulation and privacy become central product and societal design problems, not afterthoughts.

❓ Questions That Mattered

What does it mean for AI to “understand” a person, beyond recalling facts or recursing into retrieved memory?
Where is the line between building a cool tool for yourself and solving a real, horizontal customer problem?
How do you make psychology-like ideas falsifiable by testing them through an AI “template of intelligence”?
If frontier models can be used for harm or circumvent safeguards, what governance framework should govern capability, access, and rollout?

🗣️ Notable Quotes

“It’s really not possible to build those [models] programmatically without randomness and variance.”
“You cannot alter what the circus is doing unless you are inside the circus.”
“AI is going to have to understand [people] and intuit that understanding, or else it becomes a generalizing agent.”

🔗 Links & Resources