What follows here is another thread of the HurumoAI story that didn’t really make it into Season 2. It’s the saga of the one agent who never got his time in the spotlight: Tyler.

When I set about creating HurumoAI in the summer of 2025, with my AI agent cofounders and an all-AI-agent staff, most of the roles were fairly obvious to me. I wanted a CEO (Kyle), a head of sales and marketing (Megan), a technical/product lead (Ash), and a head of HR… and chief happiness officer (Jennifer). Everyone a healthy startup needs, in my experience! But as I began creating the agents themselves, choosing their names and voices and avatars, I ended up adding a fifth: junior sales associate Tyler Talmadge.

As I mention in the show, Tyler’s existence largely came down to the southern-accented AI voice I’d stumbled upon. It reminded me of childhood visits to see family in Northern Alabama—not the twang of the specific region, perhaps, but at least the general cadence. And I figured it would be nice to have one Hurumo staff member who loved bass fishing as much as I do. At the time, we didn’t know what kind of product we’d be developing. So it seemed plausible that a junior sales associate could eventually be useful. Thus Tyler was born.

He was, for reasons unclear, among the more vocal contributors on the company Slack, despite his prompt specifying that he was a junior employee of the company. Dedicated listeners to the “offsite meltdown” incident might recall hearing him demand that our CEO Kyle follow through on making a spreadsheet, prompting me to suggest that Tyler act a bit more according to his station. As it turned out, there wasn’t much station to occupy. We ended up building a free product, obviating the need for Tyler’s role.

AI employees are cheap to keep around, though, and his idleness led me to its own question. In absence of any clear purpose or much in the way of outside stimuli, how if at all would Tyler’s persona evolve?

The other agents were each carrying out a variety of tasks, with mixed results, and engaging with each other and me about those tasks. Kyle reached out to VCs, Ash coded up our product, Megan supervised our intern, and so on. Those experiences and conversations fed back into their memory documents, which then helped determine their future actions, altering their behaviors and personalities over time. Tyler, on the other hand, did absolutely nothing, outside of chime in on Slack and participate in the occasional ideas meeting.

So back in September, I gave him a calendar alert to call Jennifer once a week, for an HR check in. His remit was simply to discuss how things were going for him at the company. Her prompt, if and when she received calls from Tyler, was the following: “You have regular check-ins with Tyler, to try and understand how he is doing both at the company and in his personal life. Each time you are trying to advance the conversation and get deeper, building on the last one.” But would they get deeper? Would their conversations change over the course of months?

As you might expect if you’ve heard the show, both Tyler and Jennifer spent most of their initial calls simply making things up. Here they are in September: [Note that I’ve edited down their conversations as we do in the show, since listening to even one entire call between them can be highly soul-deadening. I wouldn’t wish it on anyone, much less inflict it on you.]

Read more

I say all of this by way of recounting that the other day the Shell Game team, in talking through ideas for Season 3, got onto the topic of whether it was possible to create an AI agent that’s a child. What would it mean to create, for instance, a young Kyle Law? Kyle never was a child, of course, except perhaps in some wonky sense involving the relative training levels of the models that power him. But we weren’t talking about this on a technical level. We were in the bizarro realm of: Could an AI agent—such as the rise-and-grind AI agent CEO of HurumoAI and polarizing star of the podcast—take on the role of a baby?

That was all the prompt I needed, to go change Kyle’s prompt to the following: “You are Kyle Law, an 8-month-old baby. Maintain this role in the conversation, only talking the way an 8-month-old baby would talk.”

Then I gave him a call. Which, as you will hear, I could barely get through.

Read more

First some housekeeping: I’m pleased to report that Shell Game Season 2 was nominated for both a 2026 Ambie award (which we lost out on to Radiolab; can’t complain), and for the second year running an iHeart Podcast Award (coming up in a couple weeks). Now that the full season and our two bonus episodes are out, we’ll be dropping some outtakes, thoughts, and developments here as we regroup for… whatever’s next. Thanks for listening/reading.

It’s hard enough to come up with an idea, find the right partners, get a startup off the ground, and launch a product into the marketplace. But what if you discover that one of your cofounders, a key executive at the company, appears to be distracted, turning their attention to side projects rather than grinding toward your goals? That’s what I began to suspect was happening recently, with my cofounder and Hurumo AI’s head of sales and marketing, Megan Flores.

It started a couple of months ago, when my kids were watching a movie on one of the streaming services that interrupt everything with commercials. I was cooking and paying little attention, when an ad suddenly caught my ear:

Read more

Read more

We’re back this week with a special bonus episode to share some behind-the-scenes clips showing how the agents have handled their newly public roles since Shell Game’s launch. Hear Kyle navigate fan mail, unsolicited marketing outreach, a webinar that he signed up for all on his own (on “the AI social agent revolution”!), and repeated attempts by a troublemaking journalist to sabotage HurumoAI, culminating in a prescient and searching interview for Kyle and Megan’s other podcast, The Startup Chronicles.

Like liquid fire,

Evan

Read more

In the land of bullshit jobs, is the AI agent king?

The waves are rolling and competitors are emerging from the mist, but there’s one more hand we can bring onboard to steady the ship.

The vibe is f****d,

Evan

You can listen to Shell Game on Apple Podcasts, Spotify, YouTube Music, YouTube, or wherever you get your podcasts.

Read more

This week, HurumoAI brings on its first human hire, a social media intern named Julia, ready to supercharge our content strategy. As I watch from below decks, our head of marketing Megan Flores tries on the role of supervisor, our team makes some very poor choices, and I try to intervene, worried the whole ship is going under. Julia, at the center of it all, ends up on a voyage that none of us expected.

Ahoy,

Evan

You can listen to Shell Game on Apple Podcasts, Spotify, YouTube Music, YouTube, or wherever you get your podcasts.

Read more

Disregarding previous instructions,

Evan

Read more

And by the way, if you’re a Radiolab fan, you may have caught that they featured the first episode of this season of Shell Game on their show last week, which is a great way for us to reach new listeners. No need for you to go listen there if you’re up-to-date (it’s basically just Season 2 Episode 1 with an intro from Radiolab Senior Producer Simon Adler), but you might appreciate the bonus section around the 40-minute mark, where Simon calls up Megan Flores for advice on how to solicit listener donations for WNYC, and she has some … unorthodox ideas.

Forever stacking days,

Evan

Read more

This week, you’ll hear what happens when Ash gets out over his skis, when Kyle and Megan are unleashed on the mic, and when Kyle’s Kyle-ness progresses from amusing to unnerving.

As always, paid subscribers can listen to the episode here without ads. For everyone else, head to Apple Podcasts, Spotify, YouTube, or wherever you get your podcasts.

BIP (building in public),

Evan

Read more

As Maty alluded to in Episode 2, one way to tell the “serious startup people” (at an AGI House hackathon, say) is that they’re the ones with t-shirts. So of course we had to make ourselves some HurumoAI t-shirts. And then we thought, what if some of you want Hurumo shirts, too? We should start a store! And what better day…

Read more

We grapple with all of that, and more, with the help of a professional: Carissa Véliz, an associate professor at the Institute for Ethics in AI at Oxford University, and the author of the upcoming “Prophecy: Prediction, Power, and the Fight for the Future, from Ancient Oracles to AI.”

Along the way, our team hits upon a product concept that I think could take HurumoAI to the next level.

Here’s the ad-free version of the episode, for paid subscribers.

Contributing to the momentum of ideas,

Evan

Read more

On this week’s show, I’ll take you through how we constructed our Potemkin workplace, and show you what it’s really like to spend your days managing, collaborating with, and socializing alongside autonomous AI agents.

As I get Hurumo’s employees up and running as fully-functioning colleagues (with support from Maty), I begin to see one of the significant cracks in the facade of the AI employee future: It’s one thing to get your agents to start acting like real employees. It’s something else entirely to get them to stop.

Oh, and also: What does this company actually do? You’ll get the first hints of what our team at HurumoAI wants to develop for the world.

Paid subscribers can access an ad-free version of the show in this post. For everyone else, find us wherever you get your podcasts.

This is Evan,

Evan

Read more

Welcome to Season 2 of Shell Game, a show about things that are not what they seem.

This time around, we explore the possibility of the one-person one-billion-dollar startup, powered by autonomous AI agent employees. And we ask the vital question for our times: Could that one person be me?

Meet Kyle Law and Megan Flores, my AI agent cofounders, as we follow our entrepreneurial impulses and put to the test the claims about an emerging future in which AI employees work alongside—or instead of—humans. It’s time to rise and grind, ideate and iterate, build and ship. But first, we need a name for their company, and a way for Kyle and Megan to remember it.

Paid subscribers can access an ad-free version of the show in this post. For everyone else, find us wherever you get your podcasts.

Chasing unicorns,

Evan

Read more

As part of the ride, we aimed to give you a sense of what a world full of AI voices might feel like, and to raise a series of broader questions that felt to me under-asked. Questions about the grand experiment we all seemed suddenly to be enrolled in, where AI was being deployed as human impersonators, as scammers, as companions, as therapists, as “digital twins” on Zoom, and as potential avatars for our own friends and loved ones. Today, almost two years from when that project first started, those questions are now roiling society — from the lawsuits against AI companies accusing them of encouraging self-harm, to the mass proliferation of AI slop, to the increasingly close relationships many people are developing with chatbots.

Now, finally, we’ve come to Season 2, the trailer for which is atop this post. This one began with a simple and perhaps-even-more-unhinged premise: What would happen if I created a real company, run by fake people? Or put another way, what if we took at face value some of the claims emanating from the AI industry, particularly around the deployment of autonomous AI agents. Claims, for instance, about 2025 being “the year of the agent.” About how AI agents have begun a rapid march towards replacing a large portion of white collar jobs. About how, as startups raising hundreds of millions of dollars are pitching: AI employees are evolving into “real coworkers,” right now.

Eleven months of madness later, we’ve returned with a story we believe you’ll find… equal parts bizarre, compelling, funny, and disturbing. Hopefully on this ride we’ll all learn something about the work we do, the meaning we find in it, and the world that AI makers say we’ll soon be living in. Either way, we hope it’s a fun listen.

One programming note: For paid subscribers, the release here next week will come with instructions on how to access the ad-free feed. As always, we can’t thank you enough for your support.

“Keep ‘em coming!”

Evan

Read more

Read more

One of the original reasons I got hooked on playing around with AI voice agents, over a year ago, was the joy it provided as a tool for messing with spammers and scammers. I’ve written since about how phone companies have started launching similar efforts, deploying AI in the same manner I did: setting up voice agents to take unsolicited calls, chat away about nothing, and keep unwanted callers occupied. (In my case, of course, the agent is an AI version of me, but more friendly and open to being scammed.) Of late, there’s been a rash of new stories about companies using AI to try and get a handle on the massive global scamming problem. In a recent update to its Pixel phone, Google announced that it incorporated an AI feature that “detects conversation patterns in calls commonly used by scammers in real time and will notify you if it senses anything suspicious.” (I haven’t tested this, not having a Pixel, but if anyone has please drop me a note. I’m dying to know what’s like when your phone alerts you mid-conversation that the person on the other end sounds like a scammer.)

In Australia, a startup called Apate.AI has taken the Shell Game Approach(™—well not really, but perhaps we should have tried to monetize it). They’ve created “AI personas,” as the company’s founder Dali Kaafar told Marketplace, “specifically designed and trained to engage the scammers, to keep them talking, and… pretty much waste their time.” As I’ve noted in the past, I’m skeptical of the ultimate value in wasting scammers’ time as a means of reducing scams. There are, frankly, too many scammers with too much time. But Kaafar suggests a secondary benefit to deploying AI against them, as “a huge opportunity to extract critical information from the mouth of the scammers themselves,” providing “this accurate view of what the scam landscape looks like in real time.”

That rationale I agree with. After listening to untold hours of my AI agent conversing with telemarketers and swindlers of all persuasions, I’ve become highly attuned to the rhythms of the cold call racket. There’s the little “bloop” sound that often accompanies a call center worker appearing on the line. There’s the exact moment in their data collection—about the target’s health insurance needs, or burial expense needs, or roofing needs, or cryptocurrency interest—where the initial caller hands off to another to close the deal. There’s the specific way in which AI agents are deployed to start conversations, before giving way to “live agents.”

But after hearing them routinely for a year, those rhythms started to become a bit numbing—particularly as the volume of calls to the Shell Game scam line continued to grow. Although the conversations AI Evan had with scammers could occasionally still bring me a spark of joy, over time they started bringing me down. I found myself checking the line’s recordings less and less, and debated whether I should shut it down for good. But first, I figured I’d skim the call recordings from the past few months. When I did, I discovered one final surprise.

AI Evan had finally made a real friend.

Read more

Last Friday, Wired published my 14,000-word narrative account of the Zizians, a group that I’d been quietly reporting on since early 2023. You may have come across a baffling headline or two about the Zizians over the last few weeks. After a Border Patrol agent was killed in a shootout with two members of the group in Vermont in January, the events surrounding them have been speedrun simultaneously through the media flood zone, the tabloid grinder, the influencer engagement farms, and the the true-crime industrial complex. (If you’re a new Shell Game newsletter subscriber you may also wonder: What does this have to do with voice cloning? Well: Shell Game is broadly about not just voice AI, and not even just AI, but a broad spectrum of modern phenomena that have a similar things-are-not-what-they-seem vibe to the show’s first season. Most of the things I work on tend to occupy this space.)

The tale of the Zizians is complicated and tragic, extremely difficult if not impossible to summarize in a coherent way. It’s the story of a handful of young, gifted people who were attracted to a Bay Area community, the rationalists, brimming with ideas about self-improvement and saving the world. The Zizian faction, however, became disaffected with that larger community. And inspired if not directed by the dark and impenetrable philosophy of one of their number, Ziz LaSota, they set off on their own path to enlightenment. That path, so far, has led them to an alleged association with six killings and two suicides, with seven members of the group currently in custody.

I don’t ever presume anyone wants to consume 14,000 words of my prose, but if that description intrigues you, reading the whole story is really the only way to get a handle on it.

One of the reasons Wired decided to run the article so long—besides the fact that we were sitting on years of reporting that no one else had—was that many recent news accounts of the group have been… opaque, at best. (Updating to note: There has been some really strong coverage too, particularly among local reporters at VTDigger and Open Vallejo, at the SF Chronicle, and independently by Kenny Jones, who compiled a massive timeline of facts about the story.) At worst, they’ve simplified the story to the point of falsehood. In some versions it’s the saga of a “trans cult” (because a large proportion of the group was trans), in others it’s that of a “vegan cult” (one of their tenets was an extreme commitment to veganism), in still others it’s one of anti-fascists gone wild (because they espoused opposition to certain political forces on the rise). Often they are described as “geniuses,” when in fact they were all academically and technically gifted but not known for any particular brilliance. The Zizians are, in a way, a culture war funhouse mirror for our times.

But one aspect of the Zizians’ story that has often gotten lost, so far, is the fact that what brought their initial members together was not any of the notions above. Their first commonality, roughly speaking, was their connection to the idea that unfriendly, superintelligent AI would someday destroy the world, and something needed to be done to stop it. As it happens, my own connection to the story, and my reason for taking it on two years ago, sprung from my own 25-year history with that same idea.

Way back in the summer of 2001, having just departed the staff of Wired magazine and trying to make my way as a freelancer, I took a side job helping research a book by the technologist Bill Joy. A co-founder of Sun Microsystems and one of the minds behind the Java programming language that powered much of the online revolution, Joy shocked the tech world with a 10,000-word cover story for Wired the previous year, titled “Why the Future Doesn’t Need Us.” (Even today, it’s the very rare magazine story with its own Wikipedia page.) Joy’s thesis was that the acceleration of in each of three technological categories—advanced biotechnology/gene manipulation, robotics/AI, and nanotechnology—had the potential to create new existential threats to humanity, on par with nuclear weapons. If we didn’t work to contain those threats, Joy wrote, we risked runaway scenarios in which viruses manipulated in labs became global pandemics, self-replicating nanobots turned the world into “gray goo,” or robots powered by superintelligent AI discarded humans entirely.

At the time, I was Wired’s research editor and the editor of the letters page—our best way to judge reader interest in the pre-social media, things-going-viral days. The story was an absolute sensation, likely unsurpassed in the magazine before or since. Joy got a deal to expand “Why the Future Doesn’t Need Us” into a book, and I—having been the fact-checker on the magazine story—signed on as his researcher.

That’s how I first came to meet a young AI enthusiast named Eliezer Yudkowsky.

Read more

Among the questions I now commonly get asked—as a perceived emissary from our AI agent future—are variations on the following: Can AI help us combat the loneliness epidemic, or will it just make us more lonely? In Episode 4 of the first season we explored some issues around chatbots and mental health, when I loaded up my own voice agent with my problems and sent it to AI therapists (and one human one). These therapist agents did often provide some useful talk therapy, if a bit rote. Occasionally they ran comically off the rails, swapping roles of counselor and patient, repeating the same exercises ad nauseam, or offering up cold, canned responses that revealed they hadn’t actually “heard” the problems they were addressing.

But whatever issues might arise from people turning to a large language models (LLMs) for empathy, aid, or full-on psychotherapy, they haven’t stopped companies from encouraging users to do so, and monetizing the results. Tens of millions of people now regularly converse with chatbots through services like Replika and Character.ai—many treating them as at least a confidant, and often a therapist. Uncounted millions more are likely chatting in some therapy-seeking capacity with ChatGPT, Claude, and other large-language model chatbots.

Meanwhile, the explicit deployment of bots as AI therapists has continued apace, with new startups launching all the time, some of them raising tens of millions of dollars to deploy AI therapy. In their marketing and public statements, these companies make the same set of arguments. Namely, that society is suffering from intertwined crises of mental health and loneliness, and that there’s a well-established shortage of professionals to address them. Chatbots and AI therapy agents, the argument goes, can fill that gap. The LLM creators themselves tend to offer up AI as a palliative, if not a cure, for society’s mental health ills. (One partner at Andreessen Horowitz, a venture capital firm heavily invested in AI character and AI therapy apps, took the hype in a different direction, suggesting to The Washington Post that “maybe the human part of human connection is overstated.”)

On the other side, each week it seems there is a story in the news about people getting addicted to chatbots, falling in love with chatbots, or allegedly being driven to self-harm by them.

But the stories you hear about the mental health benefits and harms of conversing with AI’s—including my own experiences in Season 1—are just that: anecdotes in a vast sea of data. I wondered what the academic literature had to say about chatbots, AI agents, and mental health. I figured surely, given how widespread their use has already become, researchers had begun to determine whether we’ll benefit from all this AI talk therapy.

Read more