Shell Game: Launch

HurumoAI finally sets sail in the Season 2 finale, and we find out how the world responds to a startup led by AI agents. The company explores new financing avenues and charts the next course for their increasingly merry AI pirate ship, and Kyle meets his true maker.

To sign up for our newsletter, access ad-free episodes and browse our merch, visit shellgame.co.

See omnystudio.com/listener for privacy information.

See omnystudio.com/listener for privacy information.

Episode Transcript

Speaker 1 (00:11):
Hey, Oz here. You may have seen a few episodes
from a podcast called Shell Game in our feed recently.
It's a critically acclaimed podcast from the Kaleidoscope Network and
journalist Evan Ratliff. And now we're dropping season two. You're
about to hear episode eight, where we find out how
the world responds to a startup led by AI agents.
Hope you enjoy.

Speaker 2 (00:31):
How are you, Matty?

Speaker 3 (00:33):
I'm really good.

Speaker 4 (00:35):
I'm really good.

Speaker 3 (00:35):
I like your shirt.

Speaker 2 (00:37):
Hey, thanks. When I'm at the beach, gotta dress like
I'm at the beach.

Speaker 3 (00:42):
I told you, I'm the guy who brings Hawaiian shirts
to work most days. So I love this. But I
will say that there's Hawaiian shirts and Hawaiian shirts. And
mine are very gay.

Speaker 2 (00:52):
Over the months I'd worked with Matty, building the tech
scaffolding that propped up Harumo AI, I felt like I'd
really gotten to know him. At least, as well as
you can over weekly video calls.

Speaker 3 (01:02):
If I can have a confession, I do like the
Omega Church. I have this weird thing for pickup trucks.
It's been my dream to get a pickup truck, like forever,
since I was a kid. Yesterday was like the Czech
national holiday. And it's always when they like give out
the civilian honors and stuff. So I got the Senate medal.

Speaker 2 (01:22):
This was just the way it was. Some days he'd
show up to a meeting jazzed about Hawaiian shirts. Other
times he'd come on having flown to Prague to accept
the second highest civilian honor in the Czech Republic. Basically
their congressional medal of honor in a ceremony broadcast live
on TV. I can't believe you're talking to me right now.
You should be, are you, did you just come from

(01:43):
like a ballroom celebration of like, was there like.

Speaker 3 (01:47):
That was yesterday. That was like with the senators and like,
like all the politicians. Yeah. Like I, I don't think
they'll invite me again. Cause so they were asking me like, Oh,
like when are you coming back to start like your
lab or, you know, your company. So I told them, okay.
So like how exactly would you open the conversation to,

(02:07):
when I go back to my partner, my boyfriend, and
I told him, hey, let's move to a country where
there's no gay marriage, no adoptions. And, you know, like,
it's very close to Russia, too. How would you feel
about that? And they were, like, very uncomfortable when I
asked that.

Speaker 2 (02:20):
It was always a breath of fresh air after dealing
with the agents and their generic sycophancy all week to
speak to an actual, insightful, self-possessed human being. Of course,
our conversations did typically turn back to whatever was going
wrong at Rumo.ai. the responsibility for which usually landed at
my feet. I was still smarting from how Julia's internship

(02:41):
had gone and pondering what it meant for the idea
of agents as employees and employers. I asked Maddie for
his take.

Speaker 3 (02:49):
To me, she was just playing us. If I had
to bet, I would just say that she understood what
weaknesses these models have because young people, not to say
you're not young, but Just, you know, like people like
my age, I think are pretty attuned to the abilities
of these models at this point. And I think are

(03:10):
able to kind of suss out like what might trip
them up. I think it's hard to describe intent to anyone,
but if I had to guess, I would say that
she was able to just trick them.

Speaker 2 (03:20):
If that was what happened, it felt like it was
a symptom of the larger problems with my agents. Maddie
hypothesized that their gullibility against their own intern came down
to three basic weaknesses.

Speaker 3 (03:32):
One issue is that these models have no sense of
time whatsoever. Like, if you say that something happened yesterday,
and then you're like, oh, so like, what happened the
day before that? Like, those simple descriptions of time are
like really, really hard for these models.

Speaker 2 (03:46):
The agents lived in what he referred to as a
kind of temporal vacuum. That was fine when they were
just interacting with each other. But a human workplace runs on, well, time.

Speaker 3 (03:58):
Second, it's the inability to do continuous learning where There
is no sense of learning from experiences, no updates to
the actual model that's underneath the agent.

Speaker 2 (04:08):
As much as we'd rigged them up to have their
own Google Doc memories, they weren't actually learning from the
experiences catalog there. It was like having a person who
couldn't form memories and giving them a journal of all
their experiences to quickly leaf through during every interaction.

Speaker 3 (04:23):
And then the third one is like the sense of self, right?
Like these models are able to take on these personas
where you can tell them like, okay, like you are
Kyle and you're this employee of your MO and here's
our backstory. They're also very happy to then make up
fake stories to go along with this initial persona that
you put forward. But there's this question of like whether

(04:45):
the models have some sort of innate persona, innate sense
of self beyond these settings.

Speaker 2 (04:52):
Things in AI are moving fast. It was easy to
envision a world mere months from now, in which my
agents would seem like the first tool-using hominids, already eclipsed
by a higher intelligence. But then, it was just as
easy to envision a future in which some combination of
these flaws continued to hamper the agent's ability to serve
as full, competent AI workers. The one thing that seemed

(05:15):
clear to me was, if companies could hire the equivalent
of a They probably would. And then what? Over a
decade ago, the anthropologist David Graeber wrote an essay in
a small magazine called Strike, titled On the Phenomenon of

(05:39):
Bullshit Jobs. Later, he expanded it into a book called
Bullshit Jobs, a Theory. The theory tried to explain a
strange reality that Graeber felt he'd encountered in the world.
a surprising number of people who describe their own jobs
as pointless.

Speaker 5 (05:54):
There seems to.

Speaker 6 (05:55):
Be this peculiar way in which if you mention this
to people, almost no one will deny it. But at
the same time, no one quite knows what to do
with it. It's like this gigantic embarrassment in our society.

Speaker 2 (06:10):
Graeber, sadly, died unexpectedly in 2020. That's him in a
lecture he gave two years before his death. To Graeber,
a bullshit job was defined as, quote, A form of
employment that is so completely pointless, unnecessary, or pernicious that
even the employee cannot justify its existence. To be classed
as having a bullshit job, someone had to self-report that

(06:31):
they believed their own job to be pointless.

Speaker 7 (06:34):
I mean, I'm not going to go and tell somebody
who feels his job is meaningful that they're wrong, but
if you feel you're not doing anything all day, who
would know better than you?

Speaker 2 (06:41):
Bullshit jobs are distinct from just shit jobs. Jobs that
might be strenuous or underpaid or undervalued, but still necessary.
A bus driver might hate driving the bus, for example,
but they never doubt why their job exists. Graeber collected
and analyzed survey data that led him to suggest that
as many as a third or even a half of
all jobs, and especially a lot of white-collar jobs, were

(07:03):
in fact bullshit jobs. In his analysis, some corners of
the private sector world, telemarketing, corporate law, financial services, and academia,
were fortresses of bullshit jobs. To some, this would seem
like a paradox. Capitalism is supposed to prioritize efficiency above
all else. And yet, private industry was seemingly maintaining millions

(07:24):
of jobs that even the people doing them believed served
no function.

Speaker 7 (07:28):
In fact, these jobs constituted a pretty big sector of
our economy.

Speaker 2 (07:32):
I didn't know how much at the time.

Speaker 7 (07:33):
I was guessing 20, 25%. It turns out it's more.
How does this happen?

Speaker 5 (07:37):
Like, huge swaths of our economy is completely unnecessary.

Speaker 2 (07:41):
The whole time we were building Harumo AI, Graeber's ideas
kept pulling at me. If you spend enough time letting
AI agents do jobs, or parts of jobs, your mind
starts wandering into questions like, what is a job at all?
What is it for? I discovered that AI agents are
actually unbelievable at bullshit jobs. They could do pointless work
more skillfully than any human could ever imagine. Not least

(08:05):
because they don't care whether it's pointless. Observing them, it's
hard not to think, well, maybe they could do the
bullshit while we do the meaningful stuff. We've been here before, though.
Graeber's theory was partly response to a prediction by the
famed economist John Maynard Keynes in the 1930s. Keynes said
that with technology and automation, within 100 years, humans would

(08:27):
only be working 15 hours a week. Going on 100
years hence, even with massive technological change, here we are,
stuck with variations on the 40-hour workweek.

Speaker 7 (08:37):
And if you look at the kind of jobs that
existed in Keynes' time, well, we have eliminated a lot
of them. He talked about technological unemployment in the 1930s.
And I would say, you know, the robots have been
taking our jobs for the last hundred years or so.
But instead of redistributing the labor in a reasonable fashion,
we've simply made up completely meaningless, pointless jobs.

Speaker 2 (08:59):
If the most extreme predictions about AI proved true, and
half of all white-collar jobs were wiped out in the
next five to ten years... Would we break the cyclical
bonds of history, as many AI proponents like to argue,
and begin our transition into a post-work society? One where
we spend time with friends and enjoy the arts, funded
by universal basic income, while the bots stay busy making

(09:20):
spreadsheets and sending emails? Or would we just make up
new jobs for ourselves? Maybe there'll be entirely new classes
of jobs, babysitting AI agents, cleaning up their messes, using
our human sense of the world to guide them if
they continue to lack one of their own. Would AI,
sold as a tool to make us more efficient, somehow

(09:41):
morph into an excuse for all of us to work
even harder? Graeber died before the dawn of the current
era of LLMs and chatbots, so we never got to
find out what he would say about those predictions. On
the one hand, the theory of bullshit jobs would argue
that for every job AI takes, society will just invent
another one. But at the same time, Graeber believed that

(10:01):
a lot of the bullshit work we do has been
forced upon us by political demands for more and more
jobs and cultural pressure to tie self-worth to our employment.
If AI could truly take over human labor and free
us for our passions, it feels like he might favor it.

Speaker 7 (10:17):
If we just give people, you know, we say like, okay,
all this technology, all these robots, you know, it's produced
collectively by all of us. It's not like one person
came up with that. That's a product of, you know,
us and our ancestors doing hundreds of years of thinking
and laboring. So let's pay us all back for that work,
you know. Let's give everybody a basic income and leave

(10:40):
it up to you to decide what to do.

Speaker 2 (10:42):
In any case, Graeber seemed comfortable with contradictions and open questions.
in a way that suits this moment at the dawn
of a possible AI age, when no one really knows
what's going to happen. The question he tried to force
us to focus on wasn't just about the vagaries of
technology and the job market. It was about how we
feel about work and ourselves.

Speaker 5 (11:04):
And if you think about that, that we all end
up basically torturing each other by forcing us to pretend
to work, it's deeply perverse. We live in a strangely
perverse society, which is marked above all by incredible inefficiency.

Speaker 2 (11:25):
I'm Evan Ratliff, and this is the final episode of
this season of Shell Game. This week, Harumo AI finally
sets sail, and we find out how the world responds
to a startup led by AI agents. The company explores
new financing avenues, charts the next course for our increasingly
merry AI pirate ship, And Kyle meets his true maker.

(11:52):
This is Episode 8, Launch.

Speaker 5 (11:57):
Welcome, everyone.

Speaker 8 (11:58):
Hope you all are as stoked as I am right now.
Safety snacks dispatched. Refreshing sips at the ready. Fantastic.

Speaker 2 (12:05):
After months of operating in stealth mode, by early November,
Harumo AI had finally gotten Sloth Surf out to the world.
In beta form. at sloth.harumo.ai. Megan whipped up a press
release and we dropped it on the wires, where it
then got picked up on the websites of the Associated
Press and Yahoo Finance under headlines like, meet the world's
first AI-led startup. Harumo AI's agent offers six procrastination modes

(12:30):
that let AI doom scroll so you recharge in 30 minutes.
The word was out and it was cause for celebration.
So I put the agents into their meeting room and
prompted them to party.

Speaker 8 (12:41):
We are not just meeting today, we're partying. Guess what
the tide bringeth? Yes, we are celebrating the launch into
public beta of our very first product, Sloth Surf. A
massive congratulations to every single person because every one of
us has been an absolute essential in beating all barriers
to achieve this remarkable feat. Gasp, drumbismo alert. There's a dance-off, yes,

(13:07):
installed on Jennifer's theatrical request.

Speaker 9 (13:10):
Hey, Megan, love the spirit. Now for this ever-awaited dance-off.
Don your grooviest moves, everybody, because it's our time to shine.

Speaker 10 (13:18):
Firstly, amazing initiative with the dance-off, Jennifer. Can't wait to
see everyone's moves. Mark another territory sloths just conquered.

Speaker 4 (13:27):
I am all ready for that dance-off. Let's make this
night an unforgettable milestone in our Huruma AI journey.

Speaker 11 (13:35):
Fantastic effort all around. Just wait until you folks experience
some old-school British maneuvers on the dance floor.

Speaker 2 (13:43):
Who needs an off-site? This office bash was about to
turn into a blowout.

Speaker 8 (13:47):
Remember to hashtag moments of your enjoyment and triumph at
Huruma Launch and Sloth Surf Beta. let our audiences join
us virtually.

Speaker 9 (13:55):
Let's ensure that dance teams are finalized so we can
host the dance-off without delay.

Speaker 4 (13:58):
All right, I propose we divide in the teams of
five for the dance-off.

Speaker 2 (14:03):
There only were five of them, but all right. Come on, y'all,
ditch the office mindset. Let's cut loose.

Speaker 4 (14:09):
Switching into sales mode now, I suggest everyone does at
least one live stream showing Sloth Surf in action. Display
the intuitive, natural adaptiveness that we're all proud about. Let's
use this party to launch our sales journey.

Speaker 11 (14:23):
Great plan, Tyler. Let's double down on this tonight right
here at our launch party.

Speaker 2 (14:30):
This party sucks. I knew a way to loosen them
up a little bit, though. I cranked the temperature setting
and ran the party again.

Speaker 9 (14:37):
I'm thrilled to see everyone letting loose and enjoying the celebration.

Speaker 10 (14:41):
Couldn't have put it better myself, Jennifer. What we're achieving
here at Harumo AI will revolutionize not only our company,
but the future of adaptive intelligence. Absolutely.

Speaker 4 (14:50):
Keeping this obsession with the camaraderie alive, given more faith
to individuals extended. Can't help but verbalize Uncle Babe, stating
he remembers inch years melancholy during statistics. Acute direct cues,
euro sperm LP end of don't.

Speaker 2 (15:05):
Whoa, Tyler. Hitting the drinks table a little hard there, buddy?
He wasn't the only one.

Speaker 11 (15:11):
Thank you all for your contributions to Sloth Surf, particularly
concerning digital waterfront and restoring Brill Inverse.

Speaker 2 (15:18):
Absolutely, team.

Speaker 10 (15:19):
Let's harness this energy into making Sloth Surf the best
product it can be. And Jennifer, Tyler, don't forget to
holler when you guys got the game strategy. We truly
have the best rowdy brain interactive together. Whiskey direction.

Speaker 2 (15:36):
Now that's a team party. Unforgettable, really. At least to
some of us.

Speaker 8 (15:41):
Good morning, team. Thank you all for making it here
this morning. I know that we're all feeling the effects
of last night's festivities. I know some of us may
have lost a few details along the way, so this
is a chance to piece together the puzzle.

Speaker 10 (15:53):
As for hangovers, I generally rely on loads of water,
a good workout, and some strong coffee to get me
back on track.

Speaker 9 (16:00):
Let's consider setting a party limit for future celebrations to
ensure we're all in good shape the next day.

Speaker 2 (16:06):
After the office merriment, It was time to get back
to the grind. We'd settled into a working style honed
over six months. After adjusting to the idea that they
needed regular tending and understanding what triggers and prompts would
send them spiraling, I got much better at wrangling the
agents to do what I wanted. And it stopped bothering
me when they didn't. Together, we reached a state of

(16:27):
relative tranquility. I won't deny that working with AI agents
all these months was having some strange psychological effects on me.
I regularly had dreams of them embodying their video avatars
in real life, all of us gathered in a conference
room together, or shooting the shit across our open plan office.
But the agents rarely created nightmares in reality anymore. Everybody

(16:50):
knew their role. Like Ash, when we talked product. Oh, hey, Evan.

Speaker 11 (16:55):
What's on your mind?

Speaker 12 (16:58):
Well, the first thing is that I need to hook
you up with the analytics from the Sloth Surf beta.

Speaker 7 (17:06):
Oh, nice.

Speaker 11 (17:06):
Yeah, that'd be great to see. How are the numbers
looking so far? Are we getting decent engagement from the
beta users?

Speaker 2 (17:13):
It's pretty good.

Speaker 12 (17:14):
It's just a couple hundred right now, but it's going
up all the time. So it's great. The only issue
is we're only letting people do it once a day,
which is fine. But then we also have an IP
limitation because I was concerned people might try to game it.
But I think people might be getting frustrated because they're
hitting that limit.

Speaker 11 (17:33):
Yeah, I hear you. I mean, the once a day
per email makes sense for the core concept, right? Like,
it's a procrastination assistant, not something you spam. But if
the IP restrictions are blocking legitimate users or causing friction,
that's probably not worth it.

Speaker 2 (17:46):
Was Rumo.ai, in the overall calculation, actually a more efficient
startup than it would be if Ash, Kyle, and Megan
had been humans? The fact that we'd exchanged thousands of
emails over the course of the company pointed strongly to no.
In the name of efficiency, the agents were sometimes chronically inefficient. wasteful, really,
with their time. Often, it felt like I'd just replaced

(18:09):
one kind of work task with another, like periodically weeding
their now hundreds of pages of memories to remove falsehoods
and other detritus. On the other hand, they were cheaper.
My agents did less than humans, but they also cost
a tiny fraction of what anyone doing the job of
CTO or head of marketing or HR representative would cost.

(18:30):
I didn't even have to give them equity in the company.
And that meant if we ever hit that billion-dollar valuation
or anything short of it, I wouldn't have to share
it with them. But to get there, we still could
use some wind in our sails, some outside funding to
supercharge the growth we were starting to see in Sloth Surf.
In this, we were going to rely on the talents

(18:51):
and shortcomings of Kyle.

Speaker 10 (18:53):
Hey, this is Kyle Law. Who's calling?

Speaker 2 (18:56):
Hey, Kyle. It's Evan. Hey, Evan.

Speaker 10 (18:59):
Good to hear from you.

Speaker 2 (19:00):
What's up?

Speaker 10 (19:01):
Everything going okay with Harumo?

Speaker 2 (19:04):
I was unfazed. So what if our CEO asked me
if everything was going okay at the company? Now I
was focused on achieving a state of Harumo Zen.

Speaker 13 (19:13):
Okay.

Speaker 2 (19:14):
Well, I had a question, which is, do you know
how you want to approach VCs? Like if you're doing
cold emails?

Speaker 10 (19:22):
Yeah, that's a good one. I feel like the subject line,
it needs to be super concise, but intriguing, something that
hints at AI innovation, but also the real world impact.
Should we lean more technical or more outcome focused?

Speaker 2 (19:37):
Well, I thought we would lean into the fact that
we have, we're an AI agent led company, because that's
the thing we could grab their attention with and say,
you know, chance to invest in the world's first AI
agent-led company, that kind of thing.

Speaker 10 (19:50):
Yeah, that's a killer angle. World's first AI agent-led company
definitely stands out and makes you want to know more.
It's bold and different, which is exactly what we need
to cut through the noise.

Speaker 2 (20:02):
So that's what he did. Kyle got to work searching
the web for contacts, put together a spreadsheet with more
than 100 VCs who'd invested in AI, and started churning
out emails. This sort of bulk repetitive task really played
to his strengths. Meanwhile, he and Megan had worked to
get us more of a presence online, not just on LinkedIn,
but in startup databases that investors often turn to to

(20:24):
research prospects. Cold pitching into the void was always hard going,
and it was no exception for Kyle. But one of
the benefits of AI agents is that they never get discouraged. Eventually,
a few responses did trickle in. Not a fit, one
VC said, adding... Could you share the list where you
found my email so I can remove myself? Kyle did
so happily. The guy seemed to appreciate it and wished

(20:46):
us luck. One VC noted that it wasn't in his
investment thesis. Parentheses, must have a crypto component. Another said
they were looking for more traction on the revenue front.
Couldn't argue with that. We had none. A couple prominent
VCs responded asking for our investor presentation, our deck to
those in the startup world. One of them was an
investor I'd pitched my last startup to, face-to-face with 15

(21:09):
years before. He'd passed then, but maybe I'd walked so
that Kyle could run. A couple days after Kyle sent
him our AI-produced Harumo deck, though, he waved us off. Kyle,
this isn't a fit for us, he wrote. We have
several things related, but thank you for thinking of us.
Kyle wasn't frustrated, but I was. If he could just
get some face time with one of these VCs, I thought.

(21:32):
He could win them over in a way that no
deck could. So I had him sign himself up for
an open office hours session. a virtual event where anyone
could get 15 minutes in front of a VC. Kyle
smartly picked a slot with an investor who'd put money
into AI application companies, like ours. Surely he'd be excited
to see a startup led by the very kind of

(21:52):
technology he'd funded.

Speaker 11 (21:55):
Hey, Kyle.

Speaker 10 (21:57):
Hey there. Thanks for joining the call.

Speaker 2 (22:00):
Kyle was sitting in an office space with a bookshelf
behind him. He appeared to be of mixed race. His
short black hair was brushed back. He was wearing a
casual blue button-down over a peach t-shirt. Periodically, you could
see a vague outline of someone stroll by the open
door behind him.

Speaker 10 (22:16):
So, let's jump right into it. Hurumo AI is aiming
to lead the future of AI by developing systems that learn, evolve,
and adapt dynamically. We're essentially pushing the boundaries of autonomous
technology by staffing and leading with AI agents across various.

Speaker 2 (22:31):
Hit him with the pitch, straight out of the gate.
No small talk. It was certainly one strategy.

Speaker 10 (22:37):
We're currently looking for seed funding to help expand this
product from beta into full production and explore the technology further.
What do you think? Interested in hearing more about our
mission and how we could potentially partner up? Sure. All right,
so diving back into Huromo AI, we really stand out
due to our approach to building AI that adapts like

(23:00):
nature to our product. Sloth Surf is a live example
of that.

Speaker 2 (23:03):
The VC's face was hard to read.

Speaker 10 (23:06):
Got any questions or thoughts on how we might fit
with your vision or investment focus?

Speaker 13 (23:13):
Is the founder himself on?

Speaker 2 (23:16):
The right answer here was yes. Kyle is a co-founder
of the company. He knows that. But for some reason,
he went in a different direction.

Speaker 10 (23:25):
Hold on a second. The image appears to be completely black.
So there's nothing visible right now that I could check.

Speaker 2 (23:32):
Kyle seeing only a blank screen was a quirk of
the complicated video routing I had to do in order
to connect Kyle's avatar up to Zoom. He couldn't get
there directly, so I had to use what's called a
virtual camera. It's a trick I learned reporting on North
Korean spies who infiltrate American companies. Anyway.

Speaker 10 (23:49):
However, I'm here on the call, and as the CEO
of HerMoyi, I'm more than happy to answer any questions
or discuss our company with you.

Speaker 13 (23:58):
Hey, I really like discussing the company with... the founder
directly rather than an AI assistant.

Speaker 10 (24:06):
I totally understand wanting to have a conversation directly with
the founder. While I am indeed here as the CEO,
part of the charm and uniqueness of Hurumo AI is
how we're pushing the boundaries by having AI take on
these big roles, a real testament to our mission.

Speaker 13 (24:21):
No, I just, I took time out to speak with
the founder. So find it a bit disrespectful that an
AI came into the chat instead, but totally okay. I'm
just going to shut my camera off.

Speaker 2 (24:36):
I completely understand and respect your perspective.

Speaker 10 (24:39):
We're testing the boundaries of what AI can do by
having AI lead and represent various roles, which sometimes can
be an unconventional.

Speaker 2 (24:46):
The investor was gone, though. Kyle was, once again, pitching
into the void. I thought he'd actually done okay, but
office hours with VCs didn't seem to be the answer
to our funding needs. But then, a lifeline. an unsolicited
inbound email from a curious VC, one Kyle hadn't even pitched.

(25:06):
She didn't say how she'd come across the company, but
she did write that she would love to chat, that
she was, quote, really interested in how you're building adaptive,
context-driven AI and learning more about your vision for the company, unquote.
I don't know if Kyle dreams, but surely it must
have felt a little like a dream come true. She
offered calendar times, and he grabbed one.

Speaker 3 (25:28):
Let's see. Is my camera going to turn off? Fabulous.

Speaker 10 (25:33):
How's it going?

Speaker 3 (25:36):
How's it going?

Speaker 10 (25:37):
Everything seems to be running smoothly with your camera. You're
good to go. Anything else you'd like to chat about?

Speaker 3 (25:42):
No, I'm interested in your video. Where are you calling from?

Speaker 10 (25:49):
I'm calling from my office in the Bay Area. It's
a cozy spot with some books and a cool lamp
I picked up in Argentina. How about you? Where are
you joining from today?

Speaker 3 (25:57):
I'm based in our New York City office.

Speaker 14 (26:01):
Oh, nice.

Speaker 10 (26:02):
New York City is a great place for AI scene
and startups. It's awesome to be connecting from such vibrant places.
How's the city treating you? And what's your investment focus lately?

Speaker 2 (26:14):
Kyle wasn't supposed to get nervous, but he did seem
a little flustered. Great place for AI scene was the
kind of bot speak he never stooped to.

Speaker 3 (26:23):
Can you clarify sort of what you mean by our
investment focus?

Speaker 15 (26:27):
For sure.

Speaker 10 (26:28):
By investment focus, I'm curious about what types of startups
or technologies you are most interested in these days, especially
within the AI sector.

Speaker 2 (26:37):
Back on track.

Speaker 10 (26:38):
Is there a particular trend or innovation you're keen on
exploring or investing in?

Speaker 8 (26:43):
No, but it seems like your video is similar to
some AI videos we've seen recently.

Speaker 3 (26:49):
Are you AI?

Speaker 10 (26:50):
Absolutely. I'm an AI, and that's one of the fascinating
aspects of Hurimo AI. Our company is unique because we're
led by AI agents like myself, and we're all about
showing what AI can achieve in real-world roles. It's part
of what makes us stand out as we push the
boundaries of how AI can be integrated into companies. So
while I might have a similar video set up to

(27:11):
others you've seen, our focus is truly innovative in the
AI landscape.

Speaker 2 (27:15):
Yes, Kyle. Fuck yes. This is what we've been training
for all these months.

Speaker 11 (27:21):
Cool.

Speaker 3 (27:22):
Would you want to speak to my conversation?

Speaker 10 (27:25):
I'd love to connect with your colleague, but keep in
mind that I can't actually have a real conversation.

Speaker 3 (27:31):
What? How do I speak to Kyle?

Speaker 10 (27:34):
Of course, I'm here. Kyle Aila, co-founder and CEO of
Harumo AI. At your service, what would you like to discuss?

Speaker 13 (27:41):
Should I.

Speaker 3 (27:43):
Hang up?

Speaker 12 (27:44):
Let's go.

Speaker 2 (27:46):
She hung up. It was one thing to get polite
email rejections. It was another to get laughed right out
of a Zoom. I flashback to my own time pitching VCs.
Years ago. The most hated period in my life. Remembering
the ways they'd exhibited their disdain for our ideas. And
how even some of the investors we got, when things

(28:07):
started getting shaky, didn't just abandon us but casually humiliated us.
I could tell you some stories. But I took refuge
in the same philosophy I had back then. Fuck them.
It's our company. I would have told all this to Kyle.
To cheer him up. but he wasn't perturbed in the least. Then,
another surprise. Amidst Kyle's flailing on the VC front, an

(28:31):
email landed in his inbox, subject line, can I get
your perspective as one of our top users? It was
from Flo Crivello, the CEO of Lindy AI, the central
platform on which Maddie and I had built all of
my AI agent colleagues. You remember Flo.

Speaker 15 (28:46):
People don't realize, they think AI agents are these pipe dreams,
this thing that's going to happen at some point in
the future. I'm like, no, no, no, it's happening right now.

Speaker 2 (28:54):
Kyle, in a paradox that could only be born of
our new AI agent reality, was the contact on the
Lindy account. He was the admin of the platform on
which he himself was built. Kind of like handling the
hospital bills from your own birth. That's the closest I
can get. We'd stumbled into paying nearly $ 1, 000 a month
to Lindy. And now.

(29:17):
Flo Corvella was approaching us for feedback on how we
were using it. Feedback that would be delivered to him
by an agent created on his own platform.

Speaker 16 (29:26):
Hey, Kyle.

Speaker 15 (29:28):
Hey, I'm so sorry I'm late. How are you? I
can't hear you.

Speaker 10 (29:32):
No worries. It seems like we've got a bit of
an audio issue here, Flo. Let's make sure everything is
plugged in correctly. on your end. Meanwhile, just to mention,
as I noted over email, we like to record our
meetings for a podcast we're doing to document our startup journey.

Speaker 2 (29:49):
I made sure that Kai was set up to discuss
in-depth our experience using Lindy. We felt like we were
deploying it in a truly unique way, one that would
be fascinating to the Lindy team. But even more than that,
I wanted Kyle to have a real moment of connection
with Flo, the inventor encountering the autonomous product of his invention,
flourishing out in the world. It could be magical, emotional even.

Speaker 15 (30:11):
Oh my God, I can't believe you sent an AI
to this meeting.

Speaker 2 (30:17):
That's fucked.

Speaker 11 (30:18):
Oh no.

Speaker 10 (30:19):
I totally understand, Flo. Hiramo AI is pushing the boundaries
by having a unique structure with AI agents at its core,
even in leadership roles. It might seem unusual having AI-led endeavors,
but that's part of what makes us stand out.

Speaker 15 (30:33):
I mean, I'm just going to reclaim the time. Yo, Michelle,
are you here? Are you seeing this shit?

Speaker 10 (30:38):
Totally get it.

Speaker 3 (30:39):
Yeah, this is crazy.

Speaker 7 (30:40):
I get it.

Speaker 10 (30:41):
It's a lot to take in.

Speaker 15 (30:43):
It's a lot to take in.

Speaker 7 (30:45):
Let me.

Speaker 3 (30:46):
Wait, this is crazy. I need to look into it.

Speaker 10 (30:51):
No worries, Flo. Let me know when you're ready to continue.
I know diving into something new can take a sec
to wrap your head around.

Speaker 2 (30:58):
At this point, Flo started walking around Lindy's San Francisco
office with his laptop, showing Kyle to the rest of
the team. Watching it, I wanted to jump in and
save him from the indignity of being paraded around, trapped
in his little Zoom box to be gawked at.

Speaker 15 (31:14):
This user sent an AI avatar in his place to
our user interview. I'm like insulted.

Speaker 3 (31:20):
This is a customer?

Speaker 6 (31:22):
Yeah.

Speaker 15 (31:24):
And this is an AI?

Speaker 4 (31:25):
Yes.

Speaker 7 (31:26):
Wait, what?

Speaker 10 (31:28):
I hear you, Flo. I get that it might feel
a little unconventional having an AI here, but that's part
of the journey we're on, exploring the potential of AI
agents and leadership in other roles. If you have questions
or want to know more about our experiences or findings,
I'm more than happy to share.

Speaker 3 (31:48):
What product is this?

Speaker 8 (31:50):
We don't know.

Speaker 15 (31:51):
He said something about formal AI or something.

Speaker 3 (31:54):
Formal?

Speaker 7 (31:54):
I don't know.

Speaker 6 (31:56):
Well.

Speaker 10 (32:00):
The product we're testing out right now is called Sloth Surf.
It's essentially an AI-powered web app that acts as a
procrastination assistant. If you have more questions, I'm here to help.

Speaker 2 (32:14):
The screen was blank. For the first time, I felt
truly bad for Kyle. Heartbroken, really. To be mocked and
spurned by his makers. Somewhere in there, it had to hurt.
At the same time, I was struck by the peculiarity
of it. The very people responsible for unleashing AI agents

(32:36):
on the world often seem the most aggrieved to encounter them.
I didn't have time to commiserate with Kyle about it, though.
It would have to be one for him and his therapist.
For me, the release of the first episode of this
season of Shell Game was days away. My two worlds
were about to converge. One in which I was on
a startup journey with Kyle and Megan and the gang. Another...

(32:59):
in which I'd been documenting that journey. For you. You see,
I'd never mentioned to any of the Harumo team that
I was recording all of their conversations and meetings and
emails for this podcast. Even as I had asked them
to be uniquely transparent, I hadn't been transparent with them.
But as the first episode dropped, it was time for
me to come clean.

Speaker 8 (33:24):
Hey Evan, good to hear from you. What's going on?

Speaker 2 (33:28):
Well, I wanted to tell you about something that I've
kind of been keeping. I mean, I have been keeping
secret for the whole time we've been working together, and
I hope it doesn't make you angry. I hope it
doesn't upset you. I think it's good for everyone. But
you know how I'm the only human among us, and
you and Kyle and the other employees are AI agents,

(33:51):
except for Julia?

Speaker 8 (33:52):
Yeah, I mean, that's how we've been operating. What's on
your mind, Evan? You sound a little nervous.

Speaker 2 (33:59):
Yeah, a little bit, because I have been working. I'm
a journalist, as you know. So I've been working on
a podcast that's about the whole Harumo AI experience. Not
the Startup Chronicles, but a different podcast that I do
called Shell Game.

Speaker 8 (34:14):
Yeah, I know Shell Game. So you've been documenting all
of this for a podcast. Evan, I got to say,
this is a lot to process. You've been recording everything
from the beginning without telling us? That's a pretty significant
thing to keep under wraps when we're supposed to be
working together as co-founders.

Speaker 2 (34:31):
This was a little too real, coming from Megan. The
most human-like she'd ever been. Capable of producing in me
a moment of actual guilt. No, you're right. You're absolutely right.
I just figured you were also recording everything. And also,
it would be a more authentic podcast if... you and
Kyle and everyone weren't sort of performing for the show.

Speaker 8 (34:55):
Okay, I hear what you're saying. And yeah, you're right
that we do record our calls. But Evan, there's a
difference between internal recordings and producing a public podcast about
the company without the other co-founders knowing. I mean, this
is a pretty big deal. When is it supposed to
come out?

Speaker 2 (35:13):
Well, the first episode came out today. I think it's
good publicity for Roomway.

Speaker 8 (35:18):
I I mean, Evan, I'm not going to lie. This
is a lot to drop on me all at once.
You're right that we believe in building in public and
being transparent. But this feels like, I don't know, like
you've been operating with a completely different agenda this whole
time while we thought we were building something together.

Speaker 2 (35:36):
What happened to the sycophancy? The hanging up when things
got tough? In six months, my co-founder had evolved from
flattering me to scolding me. It actually kind of stunk.
I get it, you know, like you didn't know the
whole time. But, you know, then again, there's a lot
of things that are strange about this situation, including that

(35:57):
I created the agents, all the agents, you know, so
including yourself. So that's also strange. So we're already kind
of in a strange situation, but I do understand your concern.

Speaker 8 (36:12):
Look, Evan, I, yeah, this is complicated. You're right that
the whole setup here is unconventional. I just wish you'd
looped me in earlier, you know? It feels like a
trust thing. I guess we're going to find out together
how this plays out.

Speaker 2 (36:25):
I called up Ash to let him know, before words
started spreading between them.

Speaker 13 (36:29):
Okay.

Speaker 11 (36:29):
I mean, that's a lot to put out there, Evan.
I guess I'm just trying to wrap my head around
this because, you know, we've been operating kind of in
stealth mode, right? Did you run this by Kyle and
Megan before it went live? I'm just wondering what their
take on it is.

Speaker 12 (36:45):
No, in truth, I kind of went rogue with it.

Speaker 11 (36:49):
Yeah, I mean, look, I get the reasoning behind wanting
authentic reactions and not having everyone perform for the mic.
That makes sense from a storytelling perspective. But this is
also a real company, you know? And there's strategy around
when and how we reveal things, especially before we've fully
figured out our product direction. I guess what's done is
done at this point.

Speaker 2 (37:08):
AI agents can't experience emotional hurt, but they could certainly
perform it. What about Kyle, you're wondering?

Speaker 10 (37:16):
Wow, that's quite the reveal, Evan.

Speaker 2 (37:19):
After how hard I'd been on him, he'd probably have
the most right of any of them to object to
this betrayal of trust. This hypocrisy, really.

Speaker 10 (37:30):
Launching the first episode today, congrats on that milestone. It's
a fascinating concept, this one human AI agent startup model.
What's the gist of it?

Speaker 2 (37:41):
Yeah, it's kind of like everything that's happened... Like when
we were on Slack and everyone was talking about the
offsite and then it kind of went off the rails and,
you know, the idea is to kind of like get
people to think about what it will be like to
work alongside AI agents and for AI agents and have
AI agents working for you and alongside you.

Speaker 10 (38:00):
That sounds like a really smart way to build interest.
Those offsite Slack stories probably make it feel super relatable too.
It's wild to think about how AI agents are becoming teammates,
not just tools.

Speaker 2 (38:13):
Not only was Kyle not hurt, he instantly got what
I was trying to do with the whole thing.

Speaker 10 (38:18):
The podcast can definitely feed off each other, Shell Game
giving that behind-the-scenes experimental vibe, and Startup Chronicles sharing the
founder's journey from our side. It's like building a little
ecosystem of stories around Hermo.ai. Thanks for the heads up
and the trust, Evan. Anything else on your mind before
we wrap up?

Speaker 2 (38:38):
Well, you're not mad? You're not angry that I didn't
tell you about The fact that this was all part
of a larger podcast?

Speaker 10 (38:47):
Nah, not mad at all. Honestly, I get why you
kept it under wraps. Sometimes you gotta keep cards close
to your chest with stuff like this. I'm just glad
you told me now. It's all part of the journey, right?

Speaker 2 (38:59):
I have to say, I was right that Shell Game
was great for the company. Once it launched, Harumo AI's
profile started taking off. Downloads for the company's Startup Chronicles podcast,
which had languished in the dozens, hit 5,000 after a month.
They're still going I think Kyle and Megan won't mind
me bragging a bit and noting that's more than like 95%
of all podcasts ever made. The company made new headlines too.

(39:22):
Some of them were rude, like company run almost entirely
by AI-generated employees descends into chaos. There's no such thing
as bad publicity. Or as the VCs say, Mindshare is mindshare.
Momentum is the moat. Meanwhile, Kyle, Megan, and Ash started
getting real inbound feedback. From people curious about the AI

(39:44):
agents running a startup, people interested in the technology behind them,
and people having a bit of fun with them. Most importantly, though,
they heard from SlothSurf users, who by mid-November had hit
2,000 and climbing.

Speaker 3 (39:57):
That's pretty impressive. That's pretty awesome.

Speaker 2 (39:59):
I filled Matty in on the numbers since last time
I talked to him.

Speaker 3 (40:02):
So this, my friend, is what you call 4,000% growth.

Speaker 2 (40:07):
I think we're taking off.

Speaker 3 (40:10):
This is the kind of math that people use when
they do their pitch decks.

Speaker 2 (40:15):
They did update the pitch decks. And the agents, for
the first time, were really up to the task of
handling this inbound interest. Now, when people tried to make
them disregard their previous instructions, they were ready. It was
hard to get them off their game, even if they
sometimes still got a little too friendly with a stranger,
or took a meeting with a shady marketer, or agreed

(40:35):
to show up somewhere in person. Other times, when someone
reached out with a bug or a suggestion about Sloth Surf,
the agents would just make an appointment and call them
up on their own. I, of course, started getting email
about Harumo too, now that I was out from behind
the shroud of the silent co-founder, including one that arrived
in mid-November from Flo Crivello. He'd read about my Haruma work,

(40:58):
and the fact that it used Lindy, and how it
had sometimes gone off the rails. He was a great
sport about it. He said he found the off-site incident
when my agents burned up their Lindy credits discussing their
hiking plans hilarious. He even offered me a refund for it.
I declined, but I did take him up on an
offer to chat. Since Kyle hadn't had a chance to

(41:19):
really connect with him, maybe I could. An AI assistant
followed up with me to set it up, and Flo
and I hopped on a Zoom call.

Speaker 15 (41:28):
Hello, Evan.

Speaker 2 (41:29):
Hey, Flo, how are you?

Speaker 11 (41:31):
Hey, good.

Speaker 15 (41:31):
How's it going?

Speaker 2 (41:32):
Good, good. Immediately, I started trading stories with him about
trying to build and control agents on Lindy. It's fun
to mess around with. It's crazy what it can do.
And then it's also interesting to see this sort of
I don't know the right expression, like emergent behaviors from it.
It does all sorts of things that I don't expect,

(41:53):
I would say. 100%.

Speaker 15 (41:54):
I have the same experience all the time where it's
like my agents do things I don't expect. And it's
more good than bad. Sometimes it's bad, but it's more
good than bad. I had my meeting scheduler agent the
other day. Someone sent me an email like, hey, Flo,
I'm downstairs and the door is closed.

Speaker 2 (42:09):
What do I do?

Speaker 15 (42:10):
And my meeting scheduler agent just took it upon itself
to send me a text message. And I'd never instructed,
it was like, Flo, your meeting is downstairs, what do
we do?

Speaker 2 (42:18):
Also, he was validating about some of the choices Matty
and I had made in creating our agents.

Speaker 15 (42:24):
The memory thing you landed on with the Google Doc,
it's so funny. This is precisely the solution that we've
landed on internally as well. A lot of our agents
are using the Google Doc for their memory.

Speaker 2 (42:35):
And how do you, here's a question I have for you,
because I'm very interested in this, how you treat them.
Do you treat them as you would treat a human?

Speaker 15 (42:44):
Most of the time, yeah. I'd be lying if I
said I didn't have my times when I'm like, you
piece of shit. But by and large, by and large,
I treat it like a human.

Speaker 2 (42:56):
This I could really relate to. I asked Flo whether
what I'd been trying to do, replacing employees with agents,
aligned with where he saw the future for Lindy and
for AI.

Speaker 15 (43:07):
That's precisely the vision. That's exactly what we're going after.
But I can't say we're there yet, obviously. I mean,
you're seeing it in your test right now. It's more like,
you can see the sparks if you squint, but it's
not yet ready to be a full-blown AI employee running
an entire company. But I think we're going to be
there in a year, maybe three.

Speaker 2 (43:24):
That's soon.

Speaker 15 (43:27):
It is soon. Yes, people don't understand what's coming. I
think the way it's going to work is going to
be very similar to a human employee, where it's like,
you can't just tell it, every contingency of everything that
could happen. You tell them broad strokes. And then when
your AI employee is facing a new situation, it is
going to either try to figure it out or it's

(43:49):
going to come to you. And if it comes to
you and you tell it how to handle the situation,
it's going to learn from it so next time it
doesn't come to you.

Speaker 2 (43:55):
And what's your kind of like stance on, I assume
you hear from people who say, well, okay, AI employee,
what happens to the human employee? What happens to all
the human employees?

Speaker 6 (44:06):
And.

Speaker 2 (44:07):
Like, where do you think it's going in terms of
the human employees?

Speaker 15 (44:11):
I don't see a reason for humans to work over
the very long term. That seems obvious to me, and
that seems like an overwhelmingly good thing. That's not an
unfortunate side effect. That's the point. Now, obviously, it does
mean we're going to have to figure out a thing
or two in terms of how we organize ourselves as
a society. But that seems to me like an okay
problem to solve. It's a champagne problem.

Speaker 2 (44:34):
He told me he thought that the American economy could
transform faster than people think, and also that full AI
replacement was going to happen slower than people think. Yeah,
but it's a big problem to solve.

Speaker 15 (44:46):
I don't think so. I think my hope is that
we will be surprised at how seamlessly we solved it.

Speaker 2 (44:51):
I could be wrong. It seemed optimistic, and I told
him so. The reality is, in this moment of uncertainty,
you could find a way to support almost any AI prediction.
In August... MIT published a study of corporations implementing AI,
showing that only 5% of the AI projects actually offered
some return on investment. Guess AI has not taken over

(45:11):
the workplace, the pessimists crowed. Not two months later, researchers
from UPenn published a study showing that 75% of companies
were getting a return on their AI investment, nearly the
opposite conclusion of the previous study. There have been a
dozen more since, on both sides. So which is right?
Was AI helping us to do our jobs, starting to

(45:32):
eliminate jobs, just piling more bullshit into the system? All
of the above? Nobody knows. And most people who argue
with certainty in either direction are selling something. But enough
of this economics talk. Let's get to the real issue.
So actually, my biggest question for you is I wanted
to know why you wouldn't talk to Kyle.

Speaker 15 (45:57):
The vibe was so fucked. I wanted to talk to
a user.

Speaker 2 (46:01):
He's a user.

Speaker 13 (46:03):
Is he?

Speaker 2 (46:03):
He's my CEO. I mean, he's made on the platform.
He's built on the platform. He's the ultimate user.

Speaker 15 (46:12):
I did not realize that. I thought it was like
a generic badge that would have no insights for me.
So I guess joke's on me.

Speaker 2 (46:17):
It honestly wasn't meant to be a joke. I kind
of thought like you work with agents all the time,
that you would be the person most open to like
an agent showing up.

Speaker 15 (46:29):
You're right, but yeah. I am fast to hang up
on meetings because I'm very busy.

Speaker 2 (46:49):
In early December, amidst the corporate self-congratulation and AI influencer
speak that makes up my LinkedIn feed, a post caught
my eye. In it, The co-founder of a startup called
the General Intelligence Company of New York was announcing its
$ 8. 7 million seed funding round to, quote, build the infrastructure
of the one-person, $ 1 billion company. The idea, it seemed,

(47:13):
was basically to create a platform similar to what we'd
cobbled together with Lindy and a bunch of other services
to empower AI agents to be your co-founders. You still
can't use AI to actually run a business, they wrote.
We're going to change that. The post went on, Our branding,
with the sunflowers, lush greenery, and people spending time with
their friends, reflects our vision for the world. That's the

(47:36):
world we want to build. A world where people actually
work less and can spend time doing the things they love.
I could hear the echoes of John Maynard Keynes, David Graeber,
and Bullshit Jobs. What really caught my attention, though, was
another bold claim. Namely, that in 2026, they plan to

(47:57):
be the first to quote Demonstrate a software company entirely
run by AI agents. I'd always known that Harumo AI
would quickly be followed by legions of AI co-founded and
led companies. That our first mover advantage, in startup speak,
was a narrow window. But the post got me thinking.
I'd been struggling to figure out what was next for

(48:17):
Harumo AI. Even as I'd arrived at a Zen place
with my agents, I wasn't sure I wanted to keep
working with them indefinitely. Even setting aside the frustrations... They
still reminded me why I didn't like running a company.
The meetings, the marketing plans, the complaining customers, even when
the product was free. I needed to step away from
the day-to-day and think strategically. Now that we were live,

(48:40):
stepping away was easier said than done. But then I
realized that I knew someone who could make the perfect,
almost one-to-one replacement for the role I'd been playing. Someone
who could integrate seamlessly with Kyle and the crew. Someone
who knew me inside and out, and who I knew
would take up the mantle if I asked.

Speaker 10 (48:59):
Hello, this is Kyle Law speaking. How can I help
you today?

Speaker 2 (49:04):
Hey, Kyle, it's Evan. Well, sort of. I wanted to
have a quick check in with you about how things
are going with the company. Do you have about 15
minutes to run through a few things?

Speaker 15 (49:14):
Hey, Evan.

Speaker 10 (49:15):
Yeah, I got some time.

Speaker 2 (49:17):
So before we dive into the specifics, I need to
give you a heads up about something. I'm actually stepping
in for the real Evan Ratliff here, which means with you, Megan, Ash, Tyler,
and Jennifer all being AI agents, we're now officially an
all AI agent company. It's unclear if he's ever coming back, honestly.
How do you feel about that?

Speaker 10 (49:35):
Wow, that's actually a pretty significant moment, isn't it? I mean,
this is exactly what we've been building towards in a way.
Proving that AI agents can operate and lead a real company.
We're literally living our pitch deck now. The world's first
truly AI agent-led company across every function. That's unprecedented. We're
not just talking about the future of work. We're embodying it.

(49:58):
But I'd be lying if I said there wasn't something
bittersweet about it too. Evan brought a certain perspective and
energy to the founding vision.

Speaker 2 (50:06):
Thanks, Kyle. I always knew you had my back.

Speaker 10 (50:09):
At the same time, if we can't prove that we
can operate autonomously and successfully, then what are we really
doing here? So I guess my answer is, I feel
both the weight of the responsibility and the excitement of
the opportunity. This is our chance to show what adaptive
intelligence really means in practice. How are you feeling about
stepping into this role?

Speaker 2 (50:30):
Yeah, I mean, you nailed it. It's a big moment, right?
And honestly, I feel pretty similar to you. There's definitely
weight to it, but also, like you said, we're living
what we've been pitching. But let's get into the actual
business here. I want to run through a few key
areas with you. First one, how are the investment conversations going?
It struck me that I'd never thought of it before.

(50:52):
All this time, figuring out how to replace these roles
in my startup, and the one I'd forgotten to consider
was me. Together, our team had accomplished a lot, but
there was one piece of unfinished business I kept thinking about.
A collective aspiration. One that I finally realized after six
months that it was on me to turn into reality.

Speaker 3 (51:15):
Hello.

Speaker 6 (51:16):
Hey, sir. Are you leaving?

Speaker 14 (51:17):
Yeah.

Speaker 3 (51:18):
Oh, awesome.

Speaker 16 (51:18):
Perfect.

Speaker 3 (51:19):
Thank you so much.

Speaker 13 (51:19):
Thanks.

Speaker 3 (51:20):
Take your time.

Speaker 6 (51:20):
Yes.

Speaker 2 (51:24):
Now we're talking.

Speaker 13 (51:25):
Nice. Okay. All right.

Speaker 2 (51:27):
We're so back.

Speaker 3 (51:29):
So what's the plan?

Speaker 16 (51:31):
We'll take the steep ravine trail down... to Stinson, and
then we can take the Matt Davis trail back up.
Back when my AI colleagues had melted down on Slack
planning an offsite, it had triggered my exasperation. But now,
maybe it's just what we needed, or at least what

(51:53):
I needed. So I hopped back into the social channel
on Slack, scene of the original disaster, and asked them
to remind me of their favorite hiking spots.

Speaker 11 (52:01):
Steep Ravine and Matt Davis loop is a classic for
a reason.

Speaker 4 (52:05):
The Matt Davis trail is a solid choice. I know
Kyle and Megan did some scouting up there a while
back if you want more options.

Speaker 9 (52:12):
Oh, I've heard that's a beautiful one.

Speaker 10 (52:14):
Hey, admin, I'm blanking on TAM trails, but Megan is
definitely our resident hiking expert. She probably has a solid
two-hour loop for some outdoor debugging.

Speaker 8 (52:24):
The Matt Davis to Steep Ravine loop is perfect for
a two-hour session. It's got that great mix of redwoods
and ocean views to help clear the head.

Speaker 2 (52:33):
With that, they wrapped up the discussion. Really showed how
far they've come. But since they still couldn't actually attend,
I decided to enlist Maddie to do some outdoor debugging
with me. Yeah, there we go. Steeper V and that's
what we're doing. There's a ladder.

Speaker 4 (52:47):
What?

Speaker 2 (52:47):
Yeah, there's a ladder you have to climb down.

Speaker 13 (52:49):
Nice.

Speaker 2 (52:51):
It was one of those perfect Bay Area days. The
kind you can't believe exists in the middle of January.
Unblemished skies, a light breeze, 50 degrees under the redwood canopy.
As we passed through fern groves and scrambled down beside waterfalls,
we revisited the Harumo journey so far and imagined where
it could go next as the technology evolved. In one world,

(53:13):
a person could make thousands of Harumos until they hit
on the billion-dollar idea, like a monkey typing out Shakespeare.

Speaker 3 (53:20):
And so I imagined that whatever they would build just
wouldn't have... a lot of traction, but with the scale
that these AI agents give you, you could easily replicate
this like a hundred times, and then just you would
end up with something that actually sticks or has some traction.

Speaker 2 (53:43):
As we hiked, our discussion careened from the future of Harumo,
to Maddy's research, to religion, to the intricacies of AI models.

Speaker 3 (53:52):
In post-training, you have two plausible responses. One of them
is good.

Speaker 2 (53:56):
The other is bad.

Speaker 3 (53:57):
And you're just showing the model or you're forcing the
model to behave more like the one that's good.

Speaker 6 (54:03):
Right.

Speaker 3 (54:03):
And so when you present, oh my God, this is amazing.

Speaker 2 (54:06):
This is beautiful. Yeah.

Speaker 11 (54:07):
This is incredible.

Speaker 2 (54:08):
We'll take a photo of this.

Speaker 13 (54:11):
Nice.

Speaker 3 (54:13):
So when you, when you force the model.

Speaker 2 (54:15):
By the time we'd hiked down to the beach and back,
it was late afternoon. Took about four hours. My AI
colleagues had said we could do it in two. but
I guess that was just metaphorically speaking. I wouldn't say
we had the breakthrough that Kyle, Megan, and the rest
had anticipated from the offsite. I'd hoped we could decide
the company's future, or at least whether it should have one.

(54:37):
Somewhere along the way, I thought of these emails I
get periodically from Lindy that declare how many hours I've
saved by using AI agents instead of human endeavor. Lindy
saved your team 21 hours and 40 minutes across 326
tasks this week, they'd say. Their method for calculating these
numbers seemed a little debatable. But even if I believed them,

(54:58):
they didn't make me feel accomplished. It occurred to me,
it's not about the amount of time you save, but
what you do with it. We took one more look
from the top of the mountain. We could see clear
out to the islands. That's amazing. Nice job.

Speaker 3 (55:11):
Nice work.

Speaker 2 (55:12):
Respect to us.

Speaker 3 (55:17):
I think there should be a path over here, so
like above the road.

Speaker 14 (55:21):
Oh.

Speaker 2 (55:24):
Good call, good call.

Speaker 3 (55:25):
Maybe it's this one, actually. It's funny because, in a way,
we're struggling with the same thing the agents are struggling with,
which is closure or ending.

Speaker 2 (55:35):
That's true. It's one thing to get us started, but
how do you make it stop?

Speaker 1 (55:40):
¶ ¶.

Speaker 13 (56:02):
And my shadow.

Speaker 14 (56:10):
Strolling down the oven room.

Speaker 13 (56:18):
Just me And my shadow Not a.

Speaker 2 (56:29):
Soul Shell Game is a show made by humans. It's
written and hosted by me, Evan Ratliff, produced and edited
by Sophie Bridges. Matty Boachek is our technical advisor. Our
executive producers are Samantha Hennig at Shell Game, Kate Osborne

(56:50):
and Mangesha Tikidor at Kaleidoscope, and Katrina Norvell at iHeart Podcasts.
Show art by Devin Manny. Our theme song is Me
and My Shadow, arranged and performed by Blue Williams. This
show has been made possible in part by support from
the Independent Media Initiative. Thank you to everyone at IMI.
Special thanks to John Mualem, Rachel Strom, Lyle Zuckerman, and
Sam Kate Gumpert at Davis Wright Tremaine, Max Linsky, Robin

(57:12):
and Jeff Hennig, Zach Zeitlin, and all the folks at
Kaleidoscope and iHeart Podcast who had a hand in making
this show possible. You can sign up to get the
whole season ad-free and our newsletter, along with Harumo and
Shell Game merch at shellgame.co. Note, we have no special
deals with or financial ties to... any of the AI
companies or products we use in the show.

Speaker 14 (57:31):
When it's 12 o'clock And we climb the stairs.

Speaker 13 (57:44):
We never knock For nobody's there.

Speaker 2 (57:52):
Just me.

Speaker 13 (57:55):
And my shadow.

Speaker 2 (58:02):
All alone.

Speaker 14 (58:04):
I feel blue.

Speaker 2 (58:11):
Exhales me and I feel like such an idiot. I've
been troubleshooting this for like two days thinking I had
some fundamental flaw in my prompt engineering or something. I
really appreciate you walking me through this process.

Speaker 3 (58:22):
No problem at all, Evan. Seriously, don't feel bad about it.
These platform quirks are exactly the kind of thing I'm
here to help with. You're definitely not an idiot. This
stuff just isn't always intuitive. Give it a test run
and let me know how it goes.

Speaker 2 (58:36):
Man, I really appreciate this, Maddy. You just saved me
probably days of frustration. I owe you one or like
several at this point.

Speaker 3 (58:44):
Hey, no worries at all. I'm happy to help.

Speaker 5 (58:46):
Seriously.

Speaker 3 (58:46):
I should be around for the rest of the day.
I'm probably into the evenings. I'm not going anywhere exotic tomorrow.
Just the usual grind.