For the second season of his popular Shell Game podcast, Evan Ratliff launched a startup run almost entirely by AI agents—each with a name, a title, a personality, and an expanding memory. Through creating HurumoAI and releasing it into the world, he was reminded that a job is not simply a bundle of skills. Most of the roles filled by humans are made up of some skills that are indeed automatable and many more that require wisdom and judgment that fall outside an agent's purview.
Underpinning this conundrum is the fact that, with LLMs, we've created a confabulation machine that will lie to maintain the role it's been given. Hallucinations like these are comparatively old news, but Evan reminds us that it's also a design condition—one that someone decided to make the default. And here we are, quietly, getting used to it.
AI systems are adept at performing identity, performing confidence, performing competence. The shell game is now a question of whether there's anything under the cups. The deeper we go, the answer is increasingly that it doesn't matter—because we've agreed to play.
The Confabulation Machine
Full episode on YouTube plus a searchable transcript—Evan Ratliff on anthropomorphized agents, confabulation as a design condition, outbound AI, memory failures, and what won't change.
A Job Is Not Just a Bundle of Skills
What Evan Ratliff learned from running a company staffed by AI agents—and what it reveals about the work we can't see. On confabulation, outbound AI as consumer threat, memory failure divergence, and why the things that won't get automated are the ones we've been calling invisible.
Will Conversational AI and Experience Design Revolutionize the Behavioral Sciences?
Daniel Lametti on experience sampling, smartphones in the wild, and how conversational interfaces could reshape behavioral research—paired here as archive context for the Ratliff cluster.
Did We Agree to This?
On the quiet social contract at the center of the AI moment, and the designers who signed it on everyone's behalf. Who made the decision to make hallucination a tolerable design condition? When did fluency become more important than accuracy? And what does it mean that we're getting used to it?
The sentence-before-meaning framing is the most useful thing I've found for explaining to enterprise clients why AI systems fail in ways that are hard to anticipate. They're not starting from knowledge and expressing it—they're generating plausible continuations of whatever came before. Once you see it, you can't unsee it. And you start designing for it instead of pretending it's an edge case.