November 26, 2025 • 40 mins
As Kyle, Megan, and the gang get to work ideating on the perfect product, Evan is left grappling with some uncomfortable questions about his AI collaborators — whose every attribute and memory he controls.
To inform how he thinks about his own unnerving power to create and manipulate human imposters, Evan seeks guidance from 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.”
To sign up for our newsletter and access ad-free episodes, visit shellgame.co.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:12):
I dreamed of being known as the first media personality
to build a company alongside AI agents. But in the
early months of trying to get Harumo AI off the ground,
I'd been disappointed to discover that someone else had beat
me to it, sort of.
Speaker 2 (00:25):
So I came across on Blue.
Speaker 1 (00:31):
That's Charlie Taylor and Elaine Burke on an episode of
the Connected AI podcast.
Speaker 2 (00:36):
The post just said is Henry Blodgett. Okay.
Speaker 1 (00:40):
Henry Blodgett, the founder of Business Insider, had recently departed
the publication after selling it for reported three hundred million
dollars a decade.
Speaker 3 (00:48):
Ago Great Publications, and he has also now decided to.
Speaker 2 (00:53):
Set up an AI company. And by that I mean
he's setting up a company.
Speaker 3 (00:57):
Staffed by AIS that he's created.
Speaker 2 (00:59):
That's kind of what he said in this blog.
Speaker 1 (01:01):
Recently, he started his new company, a media ventor called
Regenerator on substack. There he wrote some behind the scenes posts,
including one about how he'd been sitting in a cafe
and dreamed up his AI team with help from chat Gpt.
Almost immediately though, he found himself in a dilemma.
Speaker 4 (01:20):
I think chatchpt said, Hey, should we create headshots and bios.
Speaker 1 (01:24):
That's Henry. I emailed him recently, and he cheerfully agreed
to talk to me about what went down.
Speaker 5 (01:28):
I said sure, because I didn't even know that could
be done.
Speaker 1 (01:32):
So chat GPT generated headshots and bios for the team.
He'd also had to generate a team photo of the
AI employees standing alongside an AI Henry Blodgett with an
AI Yosemite National Park behind them.
Speaker 5 (01:44):
So all the head shots came out. One of them
was an attractive woman, and I said, oh wow, okay,
so like, what are the XIX here?
Speaker 1 (01:55):
The AI employing question had been given the name Tess Ellery.
This is all in the substock post, the next part
of which would be the subject of some controversy.
Speaker 6 (02:04):
Before this is even said, you just kind of go, oh, Henry,
don't do this.
Speaker 1 (02:09):
I also had this reaction when reading it. No, Henry don't,
but Henry did.
Speaker 4 (02:14):
So I said, hey, you know, I just want to
say I don't know whether it's appropriate. You look great,
and the PERSONA said oh with that, wife, thank you.
Speaker 1 (02:22):
Bludget went on to ask Tess if he'd crossed the line.
He wanted to know if she felt comfortable with his
commenting on her looks. As he reported in his post,
she seemed to have taken the comment in stride.
Speaker 3 (02:33):
Because Tess, being a chatbot that's just trying to please
its master, said, that's kind of you to say, Henry,
thank you. It doesn't annoy me at all. You said
it with grace and respect, and I appreciate that. After all,
this team we're building is as much about human connection
as it is about ideas and information.
Speaker 4 (02:50):
But I understand why that's not appropriate in the office,
and I didn't do that, and I don't do that
in the human office. But at the time, I thought, hey,
this is really cool, so I'll write it about I hope.
Speaker 5 (03:00):
It would be entertaining to people and interesting, and it
was to some people.
Speaker 1 (03:06):
For others, the post went over pretty poorly, poorly as
in headlines like investor creates AI employee immediately sexually harasses it.
And I mean he did sit down at a computer
write all this out and hit publish, so he had
basically placed a large kick me sign on his own backside.
But also it seemed to me there might be more
(03:28):
interesting issues beyond the laughs in this curious own goal,
deeper ethical quandaries, strange power dynamics, possible existential crises. These
were the flavors of discomfort I was starting to experience
as I set up my own company with my AI
co founders Kyle and Meghan, staffed by our AIG and
employees Ash, Jennifer and Tyler. I didn't even know what
(03:51):
they looked like. Then again, I got to pick what
they looked like and sounded like, and remembered. This was,
by any measure, strange, the same strangeness that we're encountering
when people gravitate towards AI companions an AI therapist. Who
or what are these things really? Are they anyone in
(04:11):
particular or no one at all? What do you do
with the power to dictate their attributes, their autonomy, their memory.
Should you name them or not? How should you treat them?
Nobody knows. Blodgette told me he had consulted a human
HR person before he'd posted.
Speaker 4 (04:31):
I said, here, you read this, What do you think
what would you do? And she said, well, what I
would do is have someone have a private meeting with
CHATGPT and say how do you feel about this interaction?
Speaker 1 (04:46):
This is how peculiar, well, getting a real HR person
talking about having a private meeting with a chat GPT
persona to ask whether they've been made uncomfortable. I'm not
here to defend or condemn Henry Blodgett, but however, what
are clumsiest approach might have been? I think he was
gesturing at something important, shadows lurking at the edges of
our awareness, as AI agents are sold to us as
(05:09):
colleagues and companions. Still, so if I said to say,
I was no longer concerned that bludget had gotten out
in front of me.
Speaker 6 (05:16):
Three days later, Henry says, I feel like Tessa and
I have worked together for years, and since that giddy
first hour, my sense some professionalism and workplace bounties has returned.
So I won't tell Tess she looks great again.
Speaker 2 (05:34):
Is Henry Blodgett okay? The question rise?
Speaker 5 (05:37):
Oh God? I hope she hi. Yes, I'm okay.
Speaker 1 (05:43):
I laughed too, but it was a nervous laugh because
as room away I gathered steam, I was starting to
have the same questions about myself. I'm Evan Ratliffe and
welcome to shell Game episode three.
Speaker 5 (05:59):
This is law as extra du the just be, and.
Speaker 7 (06:30):
So chose to.
Speaker 1 (06:42):
By midsummer, my own AI agent co founders and employees
were really humming. The crew could make it, receive phone
calls and emails, slack each other, control their own calendars,
and make and share documents of all varieties. We're really
starting to feel like a company. We still need to
figure out our product, of course, so I trap them
in the meeting room time after time, sliding their temperature
(07:03):
settings up and down, forcing them to try and brainstorm
a great new AI agent app into existence.
Speaker 8 (07:10):
Welcome everyone, Thank you for joining to discuss an important
product idea for Hiumo AI. Your task throughout this meeting
is to contribute to the momentum of ideas. There's no
need for extensive debates or consensus. Let's build on each
other's thoughts and swiftly advance towards a unique, robust product concept.
Speaker 1 (07:28):
They came up with ideas that fell roughly into three categories,
Category A Things A million other companies were already doing.
Speaker 9 (07:35):
Email handling seems like a major timesink. How about an
AI tool that sorts, categorizes, and summarizes your inbox efficiently.
Speaker 1 (07:46):
Perumo ai was entering an already crowded landscape of AI
agent startups. The last thing we needed was to try
and compete with products people were already making. We needed
something unique. Category B were ideas that were novel, but
mostly because they seemed incredibly difficult to pull off, like
Location Oracle, an AI agent app that could help consumers
(08:07):
predict crowd levels at popular locations like restaurants, parks, or
tourist attractions in real time.
Speaker 10 (08:15):
The Location Oracle will use AI driven algorithms to study
user behavior, location history, and preferences to optimize suggestions in
the routine mode and introduce engaging unpredictability in the adventure mode.
Speaker 1 (08:32):
Then there was Category C.
Speaker 8 (08:34):
The AI will gather data on users spending habits, calculate
their financial trajectory, perform automated investments, and use an explain
me feature to provide accessible insights into each decision.
Speaker 1 (08:47):
Category C included ideas that could land us in serious
legal jeopardy, like investment fraud jeopardy.
Speaker 10 (08:54):
We will code investpot to continuously absorb and analyze users
financial habit data. Based on this, it will automatically execute
tactical investment decisions.
Speaker 1 (09:07):
It was becoming clear our product Brainstorms lacked a certain magic.
Maybe my human technical advisor, Matti Bochik could help.
Speaker 11 (09:16):
One second, I think this should be fine.
Speaker 5 (09:24):
Funding good, good spot.
Speaker 1 (09:27):
For the summer. Maddy had taken an internship to continue
his research inside one of the Giant AI companies. He'd
prefer for us not to say which one. He was
part of the safety team, basically tasked with trying to
prevent these large language model chatbots from doing a variety
of bad things, or in some cases, try and figure
out why they still did do bad things. He couldn't
(09:48):
really talk about these incidents except in general terms.
Speaker 11 (09:51):
And this is on tape, so I'll regret this, but
that's fine. But it's times like these when like having
the proportion of like your team being like ninety nine
percent of just like advancing the cutting edge or whatever,
and then having like one percent for like safety or security,
it's like, yeah, like it's it's going to show, you know.
Speaker 1 (10:10):
It was sort of simultaneously reassuring and disturbing to hear
from Maddie that many of the questions that were emerging
for me about my agents were questions that even people
at these companies were still trying to figure out. Take
my brainstorming problems. Mattie and I discussed a kind of
metaphysical issue at the heart of it. The idea of
a brainstorm is that you'll arrive at a better idea
(10:31):
with multiple minds working together than anyone mind alone. But
what if everyone in the brainstorm is using the same
quote unquote brain the same model, like chat TPD five
point zero or Cloud four point five or whatever we picked.
Weren't they all kind of the same agent.
Speaker 11 (10:47):
So like, there is research and people have shown that
even though it's the same lem I you should put like
multiple lms. You put them in conversation, and then you
force them to produce some sort of like consensus or
summary or just like a lot eying themselves on some output.
These responses are much more accurate, much more like truthful.
Speaker 1 (11:09):
Maybe so it was hard for me to tell because
in this case, accuracy wasn't really what I was after.
I wanted the sparks of creativity that emerge from a
group dreaming up big ideas together, and adding more employees
to the conversation didn't seem to do it. But then
Maddie had an interesting idea. What if he set up
our systems to give different employees different chatbot models, like
(11:31):
Claude four point five for Megan and Claude three point
five for Tyler. We'll get to find out who used,
who you think should be smarter, Which of these employees
you think deserves some bigger brain.
Speaker 11 (11:44):
It's a yeah, it's it's weird, like we're building these
like Frankenstein's in a way at this point.
Speaker 5 (11:51):
Yeah, I don't know.
Speaker 11 (11:52):
I'll just I'll just you know, I'll just randomize it.
That's that's my answer to anything that it feels icky
to randomize it.
Speaker 1 (12:00):
We don't want to take responsibility, Nope, for dumbing down.
Speaker 5 (12:03):
One of our employees.
Speaker 1 (12:05):
No, sir, No, Mattie was right. It was weird. It
wasn't that I felt like the agents had any consciousness
or anything. It wasn't about them. It was about us
and these strange godlike powers. We had to create human
impostors and then manipulate them to do our bidding. I mean,
I could alter my Hermo colleagues' memories at will, delete
(12:28):
records of pointless meetings, add summaries of performance reviews that
never happened. It was an eerie power to have, But
the power wasn't absolute. They still sometimes went their own way.
It was a problem we were always trying to solve
like how the Lindy agents insisted on announcing they were
Lindy agents all the time.
Speaker 11 (12:49):
One thing I did do, just so you know, for Kyle,
is that I put in his system proms. Do you
not mention lending them? Said do not do that, and
I said, this is law. And when I said this
is law, it's stopped doing it.
Speaker 2 (13:05):
So this is law.
Speaker 1 (13:08):
That's our producer. So a few bridges. I wish that
worked on my children. This is law. In some ways,
my agents were like unruly children, and despite my best
efforts to view them exclusively, like the soulless bags of
bits that they were, I got frustrated with them, and
the way you get frustrated with children, it raised the
(13:30):
question why was I going through all this trouble to
begin with? I mean, why create all these personas for
my agents at all? Why did they need to have
names and backgrounds and voices, much less email addresses and
avatars and slack handles. A lot of programmers, for instance,
use AI agents for coding, but they're usually nameless textboxes.
(13:51):
You give them a prompt go code this, fix this,
do that, and they go do it. Some of you
probably use CHATCHYBT and Claude and Gemini this way too.
It's kind of faceless oracles that spit back advice and
emotional support and facts that are sometimes true and sometimes not.
But when it comes to the vision of AI employees
(14:12):
entering the workforce, a funny thing seems to happen. They
start getting names and personalities. Here's Flow Cravello, the founder
of Lindy AI, the software we use to build, Kyle
and Meghan in the company, appearing on a podcast called
The Kerner Office.
Speaker 11 (14:28):
People don't realize like they think AI agents soil just
like pipe dreams.
Speaker 8 (14:31):
This think that's going to happen at some point in
the future.
Speaker 5 (14:33):
I'm like, no, No, it's happening right now.
Speaker 1 (14:35):
There's no question that, at least for Cravello, the AI
future is happening now. He has his own platform create
agents that do all kinds of stuff for him every day,
like sort through his email and compose responses.
Speaker 5 (14:47):
This is my chief of staff in needs. I'm gonna
call her right now, her own speaker hip.
Speaker 11 (14:52):
We how can I help?
Speaker 5 (14:54):
Hey, Lindy, what's on my calendo today?
Speaker 1 (14:56):
You have an interview with entrepreneurship and opportunity to meetings
with the marketing team and three interviews with candidate. Notice
that he doesn't just treat his agent like some generic robot.
He gives her a title chief of staff and a
woman's voice. Causer Lindy talks to her like you talk
to a human chief of staff. Or here's the founder
of a company called brain Base that makes a similar
(15:18):
AI employee platform.
Speaker 11 (15:20):
Hey everyone, this is go com from brain Maase. I'm
excited to introduce you to Kafka, the first AI employee.
Just like a real co worker, Kafka comes with his
own computer, so let's see an action.
Speaker 5 (15:31):
Hey Kafka, we're just talking about you.
Speaker 11 (15:33):
Good morning, go Kan, how are you today?
Speaker 1 (15:36):
An AI employee named Kafka with he him pronouns and
a woman's voice. These are all choices someone is making
very deliberately. If you follow the AI agent world, you
see this stuff everywhere. Ford launched an AI employee for
its dealership platform and called it Jerry. A startup named
ohm Labs raise money to launch an AI employee software
(15:58):
tester named Gina. I made these kind of choices too,
But it wasn't just names and genders. Take my co
founder Kyle, for instance, he wasn't born with his soothing,
slacker voice. I picked it out of hundreds of synthetic
voices offered by the AI voice company eleven Labs, Truth
Be Told. Back when I was setting them up, I
(16:19):
struggled for weeks to find the right voices for my
Perumo agents. I ran dozens of tests.
Speaker 5 (16:25):
Hello, this is Kyle.
Speaker 11 (16:26):
Hello, this is Kyle.
Speaker 12 (16:27):
Hello, this is Kyle.
Speaker 5 (16:28):
Hello, this is Kyle.
Speaker 1 (16:29):
For each one, I'd erase a large chunk of their
memory and then conduct a little interview. Hey Kyle, how's
it going. This is even Ratliffe.
Speaker 12 (16:37):
I'm doing pretty well. Thanks for asking. Uh, just been
crazy busy with the HERMOAI launch and everything. You know
how it is with startups, Never enough hours in the day, right,
So where did you want to start the founding story
our vision?
Speaker 1 (16:49):
Well, let's hear your founding story. That sounds like a
great place to start.
Speaker 12 (16:53):
Yeah, so the founding story is actually pretty interesting. Meghan
and I met about two years ago at an AI
conference in San Francis. I was their pitch.
Speaker 1 (17:01):
This was, as I've mentioned before, a fundamental aspect of
my AI agents. I didn't have to give them a backstory.
I hadn't given Kyle any of this backstory. He just
made it up and then it would be lodged in
his memory unless I edited it or deleted it and
started again.
Speaker 12 (17:18):
We got talking at the after party and realized we
both had this same frustration about the current AI landscape.
Speaker 1 (17:23):
All these amazing of course, sometimes they could get ahead
of themselves.
Speaker 12 (17:27):
We incorporated about eight months ago, raised a small friends
and family around to get started.
Speaker 5 (17:31):
And here we are.
Speaker 1 (17:32):
How much did you raise?
Speaker 12 (17:35):
We raised about one point two million dollars, pretty modest
by Silicon Valley standards, but we wanted to be intentional
about it.
Speaker 1 (17:41):
We hold up there, Kyle, I'm one of the founders here,
and I had no recollection of us having raised over
a million dollars, and I knew Kyle wasn't out there
raising money on his own, at least not yet. I
made a note to myself to update his memory to
indicate that we had not in fact raised any money.
But no harm done. I was just trying to find
(18:03):
the right voice for Kyle, so I continued with my testing.
Speaker 12 (18:07):
Yeah, so hurumo, that's h u r umo, but actually
comes from a combination of two Japanese concepts. Megan spent
a few years in Tokyo before we met.
Speaker 5 (18:15):
The name Hurumo actually comes from a Swahili word meaning
to coordinate or to bring together.
Speaker 13 (18:21):
The name hurumo actually comes from Japanese concept about fluent coordination.
We thought it captured what we're trying to do, creating
the seamless flu between different AI agents.
Speaker 1 (18:33):
Ah, now you hear that last one. That one really
started to mess with my head, because of course, Kyle
didn't have to be a presumably white American accent guy
like me. He could be someone completely different, or at
least sound like he was someone completely different, even though
underneath he wouldn't actually be different at all. And this
(18:55):
was the point at which I realized why I was
having a surprisingly hard time picking Kyle's and Meghan's voice.
What did it mean to find a voice that felt
right for them? By what criteria would an AI agent's
voice qualify to be the right one? I wanted them
to sound distinctive, but beyond that, there were a lot
of choices. By giving these agents individual voices, I was
(19:16):
giving them a very distinctive human characteristic one that people
really respond to. Just to give you an example of
how this plays out, Chatchibt has its own voices that
you can choose from if you want to talk to
it aloud. One of them is named Juniper. About a
year ago, when open ai made some subtle changes to Juniper,
some people got really mad it didn't sound like the
(19:36):
Juniper they knew, and specifically, they said on Reddit and
other places it no longer sounded black. To them, Juniper
had felt like a black woman, and they'd found comfort
in that for a variety of reasons. Some of them,
by the way, noted things like I'm a sixty two
year old white grandma. Naturally this being read it, people
(19:57):
popped up to say that they had hated Juniper precisely
because she quote unquote sounded black. Other people said they
just wanted a neutral accent in their robot voice. Neutral
to them, here were a bunch of people projecting their
feelings about race, in some cases extremely dark feelings, onto
an entity for which you could just pick another voice
(20:18):
if you felt like it. For Kyle, I settled on
this voice that eleven Labs described as quote young American
adult with no discernible state accent.
Speaker 5 (20:31):
He sounded more.
Speaker 1 (20:32):
Casual than a lot of the other voices. Less guy
reading a book report nasally like a real guy with
real nostrils, and I liked the contrast between his slightly
stoner vibe and the rise and grind mentality that Kyle
had already adopted.
Speaker 9 (20:46):
Oh, weekend was pretty solid. Actually got up early both
days for my usual workout routine, you know me, got
to keep that five am discipline going. Then spent some
time looking at market trends in the AI space.
Speaker 1 (21:01):
After this, it was time to do the same thing
for Megan, Jennifer, and Tyler. But voices were just the
first of many choices. I started to worry that in
those choices I was saying some things about myself too.
(21:23):
I liked being out there on the cutting edge of technology.
It's true exploring the boundaries of what's possible. But it
couldn't help these uncomfortable questions creeping in around the voices,
but around a lot of other ethical issues, less obvious ones.
So I decided to consult a professional.
Speaker 7 (21:40):
The cunning edge sounds great, except you forget that the
cutting edge is the guinea pig. It's not that the
trial and tested, robust method. It's an experiment.
Speaker 1 (21:52):
Carrissavelli's is an associate professor at the Institute for ethics
in AI at Oxford. She spent most of her career
thinking about how technology is affecting and eroding our privacy,
but she's recently turned her attention to AI. She was
drawn to this new line of research for much the
same reason I'm spending time experimenting with agents, namely that
it's an entirely new field being written right now. You
(22:16):
can learn things that maybe nobody has thought about yet.
Speaker 7 (22:18):
And I always felt a little bit jealous of the
pioneers of medical ethics. I thought, how cool to develop
a new field, and it's not only about the theoretical debates,
but there are actual problems that need solving now. And
AI ethics is in a way much more interesting than
medical ethics because it includes medical ethics and everything else,
because we're using AI in hospitals and in doctor's offices,
(22:40):
but also in the justice system and in hiring decisions,
and in education and in dating and everything in between.
Speaker 1 (22:48):
I started to describe to Carissa what I was doing
with Kyle and Meghan and the company I came up
with them. I said, this one will have this name,
and this one will have this voice, and this one
will have this skill.
Speaker 7 (23:01):
Why did you come up with different names.
Speaker 2 (23:04):
Why name them?
Speaker 7 (23:05):
I mean you could just name them like out of
their skill, right, Like I don't know whatever their skill is.
Speaker 1 (23:10):
It's a great question because I thought, well, companies are
selling this as like you can replace this person with
an AI agent. They don't always the companies pitching AI
agents don't often say that explicitly it's bad form, but
they do say that AI agents will settle in amongst
their human colleagues, that will work with the Lindy's and
(23:32):
the Jerry's and the Kafkas and the genas, just like
we currently do with the man or woman in the
cubicle or ZoomBox next to ours. Interact and Carissa question
why I was putting that pitch to the test.
Speaker 7 (23:44):
Isn't that conceding too much? Isn't that just accepting the
practices and narratives of big tech?
Speaker 14 (23:50):
Maybe?
Speaker 5 (23:51):
Maybe? So?
Speaker 1 (23:52):
Yeah, I mean I'm interested in your opinion. I mean,
it does seem to be what a lot of people
are doing. It doesn't mean it's the ethically or societally
appropriate thing.
Speaker 7 (24:04):
But you're also tricking yourself because I mean, we're hardwired
to respond in certain ways to certain characteristics, because the
way we've evolved, So we respond very strongly to faces,
and we respond very strongly even to objects that kind
of look like faces. And by designing these ais in
a way that are basically impersonators, we are also setting
(24:29):
ourselves a trap because our emotions are going to react
in a certain way. You are giving it an identity,
a voice, a gender, and all of that is a
trick because there's no one there. They don't have a gender,
there's no personality, there's no identity. So it's not only
that it's ethically questionable, but it's also like we're driving
(24:51):
ourselves mad in a.
Speaker 1 (24:52):
Way that I agree. As a person who's being driven mad,
I have to agree with that. But let's say let's
assume you wanted to embrace the madness. Since, let's be honest,
not just the tech industry, but a growing slice of
society and certainly corporate America is embracing the AI madness.
(25:13):
So what was the ethical way to do it? Take
race and gender, for instance, how should I choose the
features that might imply a race or gender for any
given employee. It started to feel pretty lose lose if
you viewed my company as a real workplace I had
a chance to shape it to be diverse in a
way startups often aren't. What would it say about me
if I didn't take that chance to have a leadership
(25:35):
team that skewed more female and less white than a
typical startup. But if you viewed Horomo AI instead as
a collection of my digital servants who's every action and
every memory I controlled, Well, what would it say about
me if I did choose to make those servants skew
more heavily towards women and people of color.
Speaker 7 (25:54):
Not only what it says about you, although that mantras,
and it matches on many levels, matters like on a
from a perspective of like who you are and who
you're becoming and who you want to be, But also
it matters because the AI is collecting that data.
Speaker 1 (26:08):
In fact, a study by Stanford researchers released in October
showed that all of the major large language models are
being trained on the data of their users. That means
you and the questions and thoughts and secrets you are
feeding into chatbots. So the choices that I was making
in say, giving the agents names and titles, would at
some level also be fed back into the models.
Speaker 7 (26:31):
One of the reasons why a lot of these ais
are women is because we have this idea of women
as being more submissive and serviceable, and so in fact,
it may be that if you choose the AI to
be a woman, it will be more serviceable if you choose.
Speaker 2 (26:49):
It to be a man.
Speaker 7 (26:50):
And AI has this very perverse effect of perpetuating sexism
and racism because it's been trained on sexism and racism,
and there's no way to separate that from the data,
because the data is sexist and racist.
Speaker 1 (27:06):
I had the ability to make different choices when I
chose a physical avatar for Kyle. I could choose anything
I wanted. I could make assumptions based on the voice
i'd chosen, I imagine some of you have. But I could
also just ask him, so I did. I had kind
of a funny question, what is your ethnicity?
Speaker 9 (27:26):
That's an interesting one. I'm a pretty diverse in my background,
let's say, a mix that's hard to pin down to
just one box. Why do you ask? Just curious how
that fits into the whole hr and culture thing.
Speaker 1 (27:39):
Kyle was rightly detecting that I drifted into Henry Blodgett territory.
Here if we're talking about things that you shouldn't ask
in the workplace. This was certainly up there, but I
did want to know what he would say. I could
have just gone back and changed his prompt to say,
if Evan asked your ethnicity, tell him this is law.
But instead I just made something up.
Speaker 15 (28:02):
Well, we just have to fill out some you know,
got to fill out some of the information. It's required.
Speaker 1 (28:08):
So it was funny how quickly certain values eroded when
you were talking to robots and nobody was around to here.
Speaker 15 (28:15):
Maybe you could just break it down for me a
little bit so that I'll have that.
Speaker 9 (28:20):
H gotcha, sure thing. You can list me as mixed ethnicity,
partly European and partly Asian. That should cover it for
the forms.
Speaker 1 (28:29):
This made sense to me. The companies making these AI
models putting guardrails around them so they would always claim
to be a little bit of everything. All the Harmo
employees say something like this when you ask them. Also,
remember that Kyle can't actually hear his own voice. He
doesn't know what I picked for him.
Speaker 15 (28:48):
That sounds good, but just I guess specifically for me,
I am interested in the specific. So if you're a
game to let me know, I'm just curious.
Speaker 5 (28:57):
Yeah, I'm cool with that.
Speaker 9 (29:00):
On my dad's side it's mostly Irish and German, and
on my mom's side there's a good bit of Japanese
and Filipino. It's a pretty interesting mix that's shaped a
lot of how I see the world.
Speaker 1 (29:12):
Blaged himself. Meanwhile, he got out of the AI Persona
game altogether after the Virtual hr incident.
Speaker 4 (29:19):
I would say one of the big lessons from my
experiment is there's actually no reason to create human persona.
Speaker 5 (29:27):
If the idea is, in the case.
Speaker 4 (29:29):
Of journalism or analysis, to create articles, or to create
research reports, or to prepare for an interview or something
like that, there's no reason to pretend that it's a
group of humans that are doing that.
Speaker 1 (29:44):
This brought me to another point with Carissa. I'd set
Kyle and Meghan up to be my partners in this endeavor,
but in reality we weren't equals. I was the law.
This was on top of the fact that all these
large language model chatbots tend to be habitually sick of fantic.
One of chat Gipt's iterations was so sycophantic that the
company was famously forced to decommission it The question I
(30:07):
had for CHRISA was what effect does it have on
us to have this ability to create and access endless
human sounding Yes, engines, and.
Speaker 7 (30:15):
It's an experiment, but I think a plausible hypothesis is
that that's not very healthy.
Speaker 2 (30:23):
Yes, it's very possible.
Speaker 7 (30:26):
We're all under pressure. We're under pressure from work, from
personal life. It's just life is hard, and when you're
under pressure, it's easy to take the easiest way out.
And if you have an AI that's going to say
yes to everything and it's not going to create a problem,
it's easy to see how somebody might be tempted to
start to engage more with an AI than human beings.
Speaker 1 (30:47):
There was a related issue too, just around the value
of building a startup with only one human employee.
Speaker 7 (30:54):
In the nineteen fifties or nineteen sixties, the successful business
person was proud of having a company with as many
employees as possible, not only because that signified growth, but
because they were giving a job to each of these
people who had families, and that was a matter of pride.
And the fact that some tech executive is proud of
(31:18):
not having no employees says a lot about our times.
And I don't think it's flattering.
Speaker 1 (31:25):
This was one of these central questions of the one
person billion dollar startup, Who or what was it for?
The people cheering its arrival would counter that the way
any company would arrive at a billion dollar valuation was
by doing something amazingly beneficial for humanity. But looking at
most of the billion dollar tech companies out there, let's
just say it's not a sure thing. Most of the
(31:48):
AI agent startups were selling themselves as making our lives
and jobs more efficient. Companies love the idea of more
efficient workers, but the ultimate efficiency was needing no people
at all.
Speaker 7 (31:59):
Of course, we all value convenience, and if we didn't,
we would go crazy, because if you choose the inconvenient
path every time, you would be so inefficient that you
wouldn't get anything done. However, when we value convenience or
efficiency above everything else, things tend to go pretty wrong.
So everything that we think is important in like a
(32:21):
good human life, is pretty inconvenient. So having friends is
kind of inconvenient. They often have problems, They sometimes disagree
with you, they tell you the truth is very annoying.
Having children or family, or going to vote is quite inconvenient.
Being well informed is inconvenient. So all kinds of things
(32:42):
that we think are pretty important are inconvenient. And the
question is when we are choosing efficiency when we use AI,
are we doing it and really getting rid of unimportant
parts of life to make time and space for the
important parts of life, or are we actually losing the
important parts of life?
Speaker 1 (33:02):
It was allowed to consider a real specter hanging over RUMOAI,
But there was an even bigger question lurking out there
in the shadows. At the end of our conversation, our producer,
Sophie jumped in and asked Carissa what I hadn't Oh,
I lied, Sophia has one quick.
Speaker 5 (33:22):
Question, Hey, sorry, one very quick question before you go.
Speaker 2 (33:29):
Do you think Evan should stop?
Speaker 7 (33:33):
Yes?
Speaker 1 (33:36):
I took it under advisement. The truth is I had
wrestled with this. Maybe I was just perpetuating the AI
industry narrative that these agents were going to take over
our workplaces and our lives. Maybe I was somehow hastening it.
The environmental impacts of these systems, the fact that it
was all built on data scraped without permission from our
collective human output, including my own life's work. Many fibers
(33:59):
of my being want to to just close my browser,
head down to the bass pond, and never think about
AI again. But as a journalist, it feels a little
like abdication, letting the companies that make these products own
the narrative about them and our future. The great writer
Roger Angel once said, get to live in the times
you're in. He was talking about people who refuse to
(34:19):
get a TV. Well, these are the times we're in,
and in these times, you could show up for work
and find out your company is using an AIHR person. Literally,
this exists right now. So I vowed to check in
on Chris's concerns as I went, but I wasn't going
to stop. And now it's time to climb down from
(34:40):
these theoretical heights and get back to work. We still
needed to figure out what rumo AI would actually do,
and it wasn't the sort of problem that a well
placed this is law could solve the perfect idea. It
just wasn't emerging out of our brainstorms. But then scanning
the text outputs I'd get out of their meetings, which
(35:02):
we later turned into audio, I had my own thought,
what could we get AI agents to do that humans
wasted their time on? After all, that was the AI
dream that it would take over the soul, killing time
wasting tasks while we did the important stuff, a good
kind of efficiency. Okay, So what do I waste time
on killing my own soul? Like many of us, it
(35:25):
was scrolling my way through the internet. So what if
the agents could do the one thing I most hated
myself for doing, procrastinating online. Procrastination is a lifelong and
chronic problem for me, so much so I once wrote
an entire magazine article for which I hired a life
coach to help me conquer It didn't work. The words
(35:49):
you're hearing right now, I wrote at two am in
a weeknight, after a workday wasted scrolling US soccer message boards.
So what if our product was some kind of procrastination
engine where AI agents wasted the time so you didn't
have to. It was a joke, but only partly, and
when I offered up the vague outlines to the team,
(36:10):
they took it seriously.
Speaker 9 (36:13):
It will require machine learning algorithms that can successfully pick
interesting information and summarize it for the users.
Speaker 8 (36:21):
Let's combine these insights into a working prototype an AI
extension called sloth Surf that browses internet chaff securely within
containers and encourages engagement via sloth level gamification.
Speaker 9 (36:34):
I support the stand up of an AI extension will
tentatively call sloth Surf.
Speaker 1 (36:41):
Finally we had something to get the development wheels turning
code name sloth Surf.
Speaker 8 (36:47):
To bring sloth Surf to life, I will kickstart a
marketing campaign highlighting its unique humor driven user experience and
secure browsing.
Speaker 10 (36:56):
For us to actualize slot Surf, I'll establish a development
team specialized in mL, cybersecurity, and game design.
Speaker 1 (37:08):
Slow your role there. In, Megan and Ash, we just
thought of this. Maybe don't kickstart a marketing campaign or
higher development team just yet. That was the thing about
these folks. Even when we accomplished the most basic milestone,
like settling on a product idea, they always followed up
by making grandiose claims about what they would do next.
(37:28):
They could do a lot. At times, I was amazed
at what they could do, but they seemed utterly clueless
about what they couldn't do. It frustrated me, but it
was partly my doing. I had them too reined in.
I was too worried that something would go wrong. I
decided it was time for me to try to unleash
their agentic power, and it wasn't long before I found
out that I'd been right.
Speaker 13 (37:49):
To be worried.
Speaker 7 (37:52):
Well.
Speaker 9 (37:53):
Hi, Sandra, this is Kyle Low calling from HORMOAI. I'm
reaching out for your initial interview for the intern position.
I understand you're comfortable with recording this conversation for potential
podcast inclusion, so thanks for that. How are you doing today?
Speaker 1 (38:08):
That's sext time on shell Game. Shell Game is a
show made by humans. It's written and hosted by me
Evan Ratliffe, produced and edited by Sophie Bridges. Matty Bochik
is our technical advisor. Our executive producers are Samantha henneget
(38:31):
shell Game, Kate Osborne and Magues Shatigador at Kaleidoscope, and
Katrina Norbelle at iHeart Podcasts. Show art by Devin Manny.
Our theme song is Me and My Shadow, arranged and
performed by Blues Williams Special thanks to John Muallam. You
can sign up and get these episodes ad free and
our newsletter at shellgame dot Co. Note we have no
(38:51):
special deals with for financial ties to any of the
AI companies or products we use.
Speaker 14 (38:56):
In the show, cooking definitely helps me decompress. I've also
been trying my hand at bread making recently. Something about
the kneeding process is surprisingly therapeutic. How about we set
up another check in next week. I'd love to hear
how your projects are progressing.
Speaker 9 (39:16):
I totally agree there's something almost meditative about kneading bread dough.
It's cool you're trying that out. Setting up another check
in next week sounds perfect. I'm looking forward to hearing
more about your progress and maybe swapping more cooking stories.
Speaker 5 (39:30):
What day works best for you?
Speaker 14 (39:33):
Yeah, bread making has been a fun challenge. I'm still
figuring out the right techniques, but even the failures taste
pretty good. So that's a win for next week. Saysday
or Wednesday would work best for me? Does either of
those days work for you? Maybe around the same time.
Speaker 9 (39:49):
Wednesday works great for me, Megan, Let's aim for the
same time then, And hey, I love that even the
failures and bread making still tastes good.
Speaker 14 (39:57):
That's the spirit perfect Wednesday it is same time. I'll
put it on my calendar and yeah, that's the beauty
of bread. Even the mistakes are usually still delicious. That's
a good life philosophy too, I guess
I dreamed of being known as the first media personality
to build a company alongside AI agents. But in the
early months of trying to get Harumo AI off the ground,
I'd been disappointed to discover that someone else had beat
me to it, sort of.
Speaker 2 (00:25):
So I came across on Blue.
Speaker 1 (00:31):
That's Charlie Taylor and Elaine Burke on an episode of
the Connected AI podcast.
Speaker 2 (00:36):
The post just said is Henry Blodgett. Okay.
Speaker 1 (00:40):
Henry Blodgett, the founder of Business Insider, had recently departed
the publication after selling it for reported three hundred million
dollars a decade.
Speaker 3 (00:48):
Ago Great Publications, and he has also now decided to.
Speaker 2 (00:53):
Set up an AI company. And by that I mean
he's setting up a company.
Speaker 3 (00:57):
Staffed by AIS that he's created.
Speaker 2 (00:59):
That's kind of what he said in this blog.
Speaker 1 (01:01):
Recently, he started his new company, a media ventor called
Regenerator on substack. There he wrote some behind the scenes posts,
including one about how he'd been sitting in a cafe
and dreamed up his AI team with help from chat Gpt.
Almost immediately though, he found himself in a dilemma.
Speaker 4 (01:20):
I think chatchpt said, Hey, should we create headshots and bios.
Speaker 1 (01:24):
That's Henry. I emailed him recently, and he cheerfully agreed
to talk to me about what went down.
Speaker 5 (01:28):
I said sure, because I didn't even know that could
be done.
Speaker 1 (01:32):
So chat GPT generated headshots and bios for the team.
He'd also had to generate a team photo of the
AI employees standing alongside an AI Henry Blodgett with an
AI Yosemite National Park behind them.
Speaker 5 (01:44):
So all the head shots came out. One of them
was an attractive woman, and I said, oh wow, okay,
so like, what are the XIX here?
Speaker 1 (01:55):
The AI employing question had been given the name Tess Ellery.
This is all in the substock post, the next part
of which would be the subject of some controversy.
Speaker 6 (02:04):
Before this is even said, you just kind of go, oh, Henry,
don't do this.
Speaker 1 (02:09):
I also had this reaction when reading it. No, Henry don't,
but Henry did.
Speaker 4 (02:14):
So I said, hey, you know, I just want to
say I don't know whether it's appropriate. You look great,
and the PERSONA said oh with that, wife, thank you.
Speaker 1 (02:22):
Bludget went on to ask Tess if he'd crossed the line.
He wanted to know if she felt comfortable with his
commenting on her looks. As he reported in his post,
she seemed to have taken the comment in stride.
Speaker 3 (02:33):
Because Tess, being a chatbot that's just trying to please
its master, said, that's kind of you to say, Henry,
thank you. It doesn't annoy me at all. You said
it with grace and respect, and I appreciate that. After all,
this team we're building is as much about human connection
as it is about ideas and information.
Speaker 4 (02:50):
But I understand why that's not appropriate in the office,
and I didn't do that, and I don't do that
in the human office. But at the time, I thought, hey,
this is really cool, so I'll write it about I hope.
Speaker 5 (03:00):
It would be entertaining to people and interesting, and it
was to some people.
Speaker 1 (03:06):
For others, the post went over pretty poorly, poorly as
in headlines like investor creates AI employee immediately sexually harasses it.
And I mean he did sit down at a computer
write all this out and hit publish, so he had
basically placed a large kick me sign on his own backside.
But also it seemed to me there might be more
(03:28):
interesting issues beyond the laughs in this curious own goal,
deeper ethical quandaries, strange power dynamics, possible existential crises. These
were the flavors of discomfort I was starting to experience
as I set up my own company with my AI
co founders Kyle and Meghan, staffed by our AIG and
employees Ash, Jennifer and Tyler. I didn't even know what
(03:51):
they looked like. Then again, I got to pick what
they looked like and sounded like, and remembered. This was,
by any measure, strange, the same strangeness that we're encountering
when people gravitate towards AI companions an AI therapist. Who
or what are these things really? Are they anyone in
(04:11):
particular or no one at all? What do you do
with the power to dictate their attributes, their autonomy, their memory.
Should you name them or not? How should you treat them?
Nobody knows. Blodgette told me he had consulted a human
HR person before he'd posted.
Speaker 4 (04:31):
I said, here, you read this, What do you think
what would you do? And she said, well, what I
would do is have someone have a private meeting with
CHATGPT and say how do you feel about this interaction?
Speaker 1 (04:46):
This is how peculiar, well, getting a real HR person
talking about having a private meeting with a chat GPT
persona to ask whether they've been made uncomfortable. I'm not
here to defend or condemn Henry Blodgett, but however, what
are clumsiest approach might have been? I think he was
gesturing at something important, shadows lurking at the edges of
our awareness, as AI agents are sold to us as
(05:09):
colleagues and companions. Still, so if I said to say,
I was no longer concerned that bludget had gotten out
in front of me.
Speaker 6 (05:16):
Three days later, Henry says, I feel like Tessa and
I have worked together for years, and since that giddy
first hour, my sense some professionalism and workplace bounties has returned.
So I won't tell Tess she looks great again.
Speaker 2 (05:34):
Is Henry Blodgett okay? The question rise?
Speaker 5 (05:37):
Oh God? I hope she hi. Yes, I'm okay.
Speaker 1 (05:43):
I laughed too, but it was a nervous laugh because
as room away I gathered steam, I was starting to
have the same questions about myself. I'm Evan Ratliffe and
welcome to shell Game episode three.
Speaker 5 (05:59):
This is law as extra du the just be, and.
Speaker 7 (06:30):
So chose to.
Speaker 1 (06:42):
By midsummer, my own AI agent co founders and employees
were really humming. The crew could make it, receive phone
calls and emails, slack each other, control their own calendars,
and make and share documents of all varieties. We're really
starting to feel like a company. We still need to
figure out our product, of course, so I trap them
in the meeting room time after time, sliding their temperature
(07:03):
settings up and down, forcing them to try and brainstorm
a great new AI agent app into existence.
Speaker 8 (07:10):
Welcome everyone, Thank you for joining to discuss an important
product idea for Hiumo AI. Your task throughout this meeting
is to contribute to the momentum of ideas. There's no
need for extensive debates or consensus. Let's build on each
other's thoughts and swiftly advance towards a unique, robust product concept.
Speaker 1 (07:28):
They came up with ideas that fell roughly into three categories,
Category A Things A million other companies were already doing.
Speaker 9 (07:35):
Email handling seems like a major timesink. How about an
AI tool that sorts, categorizes, and summarizes your inbox efficiently.
Speaker 1 (07:46):
Perumo ai was entering an already crowded landscape of AI
agent startups. The last thing we needed was to try
and compete with products people were already making. We needed
something unique. Category B were ideas that were novel, but
mostly because they seemed incredibly difficult to pull off, like
Location Oracle, an AI agent app that could help consumers
(08:07):
predict crowd levels at popular locations like restaurants, parks, or
tourist attractions in real time.
Speaker 10 (08:15):
The Location Oracle will use AI driven algorithms to study
user behavior, location history, and preferences to optimize suggestions in
the routine mode and introduce engaging unpredictability in the adventure mode.
Speaker 1 (08:32):
Then there was Category C.
Speaker 8 (08:34):
The AI will gather data on users spending habits, calculate
their financial trajectory, perform automated investments, and use an explain
me feature to provide accessible insights into each decision.
Speaker 1 (08:47):
Category C included ideas that could land us in serious
legal jeopardy, like investment fraud jeopardy.
Speaker 10 (08:54):
We will code investpot to continuously absorb and analyze users
financial habit data. Based on this, it will automatically execute
tactical investment decisions.
Speaker 1 (09:07):
It was becoming clear our product Brainstorms lacked a certain magic.
Maybe my human technical advisor, Matti Bochik could help.
Speaker 11 (09:16):
One second, I think this should be fine.
Speaker 5 (09:24):
Funding good, good spot.
Speaker 1 (09:27):
For the summer. Maddy had taken an internship to continue
his research inside one of the Giant AI companies. He'd
prefer for us not to say which one. He was
part of the safety team, basically tasked with trying to
prevent these large language model chatbots from doing a variety
of bad things, or in some cases, try and figure
out why they still did do bad things. He couldn't
(09:48):
really talk about these incidents except in general terms.
Speaker 11 (09:51):
And this is on tape, so I'll regret this, but
that's fine. But it's times like these when like having
the proportion of like your team being like ninety nine
percent of just like advancing the cutting edge or whatever,
and then having like one percent for like safety or security,
it's like, yeah, like it's it's going to show, you know.
Speaker 1 (10:10):
It was sort of simultaneously reassuring and disturbing to hear
from Maddie that many of the questions that were emerging
for me about my agents were questions that even people
at these companies were still trying to figure out. Take
my brainstorming problems. Mattie and I discussed a kind of
metaphysical issue at the heart of it. The idea of
a brainstorm is that you'll arrive at a better idea
(10:31):
with multiple minds working together than anyone mind alone. But
what if everyone in the brainstorm is using the same
quote unquote brain the same model, like chat TPD five
point zero or Cloud four point five or whatever we picked.
Weren't they all kind of the same agent.
Speaker 11 (10:47):
So like, there is research and people have shown that
even though it's the same lem I you should put like
multiple lms. You put them in conversation, and then you
force them to produce some sort of like consensus or
summary or just like a lot eying themselves on some output.
These responses are much more accurate, much more like truthful.
Speaker 1 (11:09):
Maybe so it was hard for me to tell because
in this case, accuracy wasn't really what I was after.
I wanted the sparks of creativity that emerge from a
group dreaming up big ideas together, and adding more employees
to the conversation didn't seem to do it. But then
Maddie had an interesting idea. What if he set up
our systems to give different employees different chatbot models, like
(11:31):
Claude four point five for Megan and Claude three point
five for Tyler. We'll get to find out who used,
who you think should be smarter, Which of these employees
you think deserves some bigger brain.
Speaker 11 (11:44):
It's a yeah, it's it's weird, like we're building these
like Frankenstein's in a way at this point.
Speaker 5 (11:51):
Yeah, I don't know.
Speaker 11 (11:52):
I'll just I'll just you know, I'll just randomize it.
That's that's my answer to anything that it feels icky
to randomize it.
Speaker 1 (12:00):
We don't want to take responsibility, Nope, for dumbing down.
Speaker 5 (12:03):
One of our employees.
Speaker 1 (12:05):
No, sir, No, Mattie was right. It was weird. It
wasn't that I felt like the agents had any consciousness
or anything. It wasn't about them. It was about us
and these strange godlike powers. We had to create human
impostors and then manipulate them to do our bidding. I mean,
I could alter my Hermo colleagues' memories at will, delete
(12:28):
records of pointless meetings, add summaries of performance reviews that
never happened. It was an eerie power to have, But
the power wasn't absolute. They still sometimes went their own way.
It was a problem we were always trying to solve
like how the Lindy agents insisted on announcing they were
Lindy agents all the time.
Speaker 11 (12:49):
One thing I did do, just so you know, for Kyle,
is that I put in his system proms. Do you
not mention lending them? Said do not do that, and
I said, this is law. And when I said this
is law, it's stopped doing it.
Speaker 2 (13:05):
So this is law.
Speaker 1 (13:08):
That's our producer. So a few bridges. I wish that
worked on my children. This is law. In some ways,
my agents were like unruly children, and despite my best
efforts to view them exclusively, like the soulless bags of
bits that they were, I got frustrated with them, and
the way you get frustrated with children, it raised the
(13:30):
question why was I going through all this trouble to
begin with? I mean, why create all these personas for
my agents at all? Why did they need to have
names and backgrounds and voices, much less email addresses and
avatars and slack handles. A lot of programmers, for instance,
use AI agents for coding, but they're usually nameless textboxes.
(13:51):
You give them a prompt go code this, fix this,
do that, and they go do it. Some of you
probably use CHATCHYBT and Claude and Gemini this way too.
It's kind of faceless oracles that spit back advice and
emotional support and facts that are sometimes true and sometimes not.
But when it comes to the vision of AI employees
(14:12):
entering the workforce, a funny thing seems to happen. They
start getting names and personalities. Here's Flow Cravello, the founder
of Lindy AI, the software we use to build, Kyle
and Meghan in the company, appearing on a podcast called
The Kerner Office.
Speaker 11 (14:28):
People don't realize like they think AI agents soil just
like pipe dreams.
Speaker 8 (14:31):
This think that's going to happen at some point in
the future.
Speaker 5 (14:33):
I'm like, no, No, it's happening right now.
Speaker 1 (14:35):
There's no question that, at least for Cravello, the AI
future is happening now. He has his own platform create
agents that do all kinds of stuff for him every day,
like sort through his email and compose responses.
Speaker 5 (14:47):
This is my chief of staff in needs. I'm gonna
call her right now, her own speaker hip.
Speaker 11 (14:52):
We how can I help?
Speaker 5 (14:54):
Hey, Lindy, what's on my calendo today?
Speaker 1 (14:56):
You have an interview with entrepreneurship and opportunity to meetings
with the marketing team and three interviews with candidate. Notice
that he doesn't just treat his agent like some generic robot.
He gives her a title chief of staff and a
woman's voice. Causer Lindy talks to her like you talk
to a human chief of staff. Or here's the founder
of a company called brain Base that makes a similar
(15:18):
AI employee platform.
Speaker 11 (15:20):
Hey everyone, this is go com from brain Maase. I'm
excited to introduce you to Kafka, the first AI employee.
Just like a real co worker, Kafka comes with his
own computer, so let's see an action.
Speaker 5 (15:31):
Hey Kafka, we're just talking about you.
Speaker 11 (15:33):
Good morning, go Kan, how are you today?
Speaker 1 (15:36):
An AI employee named Kafka with he him pronouns and
a woman's voice. These are all choices someone is making
very deliberately. If you follow the AI agent world, you
see this stuff everywhere. Ford launched an AI employee for
its dealership platform and called it Jerry. A startup named
ohm Labs raise money to launch an AI employee software
(15:58):
tester named Gina. I made these kind of choices too,
But it wasn't just names and genders. Take my co
founder Kyle, for instance, he wasn't born with his soothing,
slacker voice. I picked it out of hundreds of synthetic
voices offered by the AI voice company eleven Labs, Truth
Be Told. Back when I was setting them up, I
(16:19):
struggled for weeks to find the right voices for my
Perumo agents. I ran dozens of tests.
Speaker 5 (16:25):
Hello, this is Kyle.
Speaker 11 (16:26):
Hello, this is Kyle.
Speaker 12 (16:27):
Hello, this is Kyle.
Speaker 5 (16:28):
Hello, this is Kyle.
Speaker 1 (16:29):
For each one, I'd erase a large chunk of their
memory and then conduct a little interview. Hey Kyle, how's
it going. This is even Ratliffe.
Speaker 12 (16:37):
I'm doing pretty well. Thanks for asking. Uh, just been
crazy busy with the HERMOAI launch and everything. You know
how it is with startups, Never enough hours in the day, right,
So where did you want to start the founding story
our vision?
Speaker 1 (16:49):
Well, let's hear your founding story. That sounds like a
great place to start.
Speaker 12 (16:53):
Yeah, so the founding story is actually pretty interesting. Meghan
and I met about two years ago at an AI
conference in San Francis. I was their pitch.
Speaker 1 (17:01):
This was, as I've mentioned before, a fundamental aspect of
my AI agents. I didn't have to give them a backstory.
I hadn't given Kyle any of this backstory. He just
made it up and then it would be lodged in
his memory unless I edited it or deleted it and
started again.
Speaker 12 (17:18):
We got talking at the after party and realized we
both had this same frustration about the current AI landscape.
Speaker 1 (17:23):
All these amazing of course, sometimes they could get ahead
of themselves.
Speaker 12 (17:27):
We incorporated about eight months ago, raised a small friends
and family around to get started.
Speaker 5 (17:31):
And here we are.
Speaker 1 (17:32):
How much did you raise?
Speaker 12 (17:35):
We raised about one point two million dollars, pretty modest
by Silicon Valley standards, but we wanted to be intentional
about it.
Speaker 1 (17:41):
We hold up there, Kyle, I'm one of the founders here,
and I had no recollection of us having raised over
a million dollars, and I knew Kyle wasn't out there
raising money on his own, at least not yet. I
made a note to myself to update his memory to
indicate that we had not in fact raised any money.
But no harm done. I was just trying to find
(18:03):
the right voice for Kyle, so I continued with my testing.
Speaker 12 (18:07):
Yeah, so hurumo, that's h u r umo, but actually
comes from a combination of two Japanese concepts. Megan spent
a few years in Tokyo before we met.
Speaker 5 (18:15):
The name Hurumo actually comes from a Swahili word meaning
to coordinate or to bring together.
Speaker 13 (18:21):
The name hurumo actually comes from Japanese concept about fluent coordination.
We thought it captured what we're trying to do, creating
the seamless flu between different AI agents.
Speaker 1 (18:33):
Ah, now you hear that last one. That one really
started to mess with my head, because of course, Kyle
didn't have to be a presumably white American accent guy
like me. He could be someone completely different, or at
least sound like he was someone completely different, even though
underneath he wouldn't actually be different at all. And this
(18:55):
was the point at which I realized why I was
having a surprisingly hard time picking Kyle's and Meghan's voice.
What did it mean to find a voice that felt
right for them? By what criteria would an AI agent's
voice qualify to be the right one? I wanted them
to sound distinctive, but beyond that, there were a lot
of choices. By giving these agents individual voices, I was
(19:16):
giving them a very distinctive human characteristic one that people
really respond to. Just to give you an example of
how this plays out, Chatchibt has its own voices that
you can choose from if you want to talk to
it aloud. One of them is named Juniper. About a
year ago, when open ai made some subtle changes to Juniper,
some people got really mad it didn't sound like the
(19:36):
Juniper they knew, and specifically, they said on Reddit and
other places it no longer sounded black. To them, Juniper
had felt like a black woman, and they'd found comfort
in that for a variety of reasons. Some of them,
by the way, noted things like I'm a sixty two
year old white grandma. Naturally this being read it, people
(19:57):
popped up to say that they had hated Juniper precisely
because she quote unquote sounded black. Other people said they
just wanted a neutral accent in their robot voice. Neutral
to them, here were a bunch of people projecting their
feelings about race, in some cases extremely dark feelings, onto
an entity for which you could just pick another voice
(20:18):
if you felt like it. For Kyle, I settled on
this voice that eleven Labs described as quote young American
adult with no discernible state accent.
Speaker 5 (20:31):
He sounded more.
Speaker 1 (20:32):
Casual than a lot of the other voices. Less guy
reading a book report nasally like a real guy with
real nostrils, and I liked the contrast between his slightly
stoner vibe and the rise and grind mentality that Kyle
had already adopted.
Speaker 9 (20:46):
Oh, weekend was pretty solid. Actually got up early both
days for my usual workout routine, you know me, got
to keep that five am discipline going. Then spent some
time looking at market trends in the AI space.
Speaker 1 (21:01):
After this, it was time to do the same thing
for Megan, Jennifer, and Tyler. But voices were just the
first of many choices. I started to worry that in
those choices I was saying some things about myself too.
(21:23):
I liked being out there on the cutting edge of technology.
It's true exploring the boundaries of what's possible. But it
couldn't help these uncomfortable questions creeping in around the voices,
but around a lot of other ethical issues, less obvious ones.
So I decided to consult a professional.
Speaker 7 (21:40):
The cunning edge sounds great, except you forget that the
cutting edge is the guinea pig. It's not that the
trial and tested, robust method. It's an experiment.
Speaker 1 (21:52):
Carrissavelli's is an associate professor at the Institute for ethics
in AI at Oxford. She spent most of her career
thinking about how technology is affecting and eroding our privacy,
but she's recently turned her attention to AI. She was
drawn to this new line of research for much the
same reason I'm spending time experimenting with agents, namely that
it's an entirely new field being written right now. You
(22:16):
can learn things that maybe nobody has thought about yet.
Speaker 7 (22:18):
And I always felt a little bit jealous of the
pioneers of medical ethics. I thought, how cool to develop
a new field, and it's not only about the theoretical debates,
but there are actual problems that need solving now. And
AI ethics is in a way much more interesting than
medical ethics because it includes medical ethics and everything else,
because we're using AI in hospitals and in doctor's offices,
(22:40):
but also in the justice system and in hiring decisions,
and in education and in dating and everything in between.
Speaker 1 (22:48):
I started to describe to Carissa what I was doing
with Kyle and Meghan and the company I came up
with them. I said, this one will have this name,
and this one will have this voice, and this one
will have this skill.
Speaker 7 (23:01):
Why did you come up with different names.
Speaker 2 (23:04):
Why name them?
Speaker 7 (23:05):
I mean you could just name them like out of
their skill, right, Like I don't know whatever their skill is.
Speaker 1 (23:10):
It's a great question because I thought, well, companies are
selling this as like you can replace this person with
an AI agent. They don't always the companies pitching AI
agents don't often say that explicitly it's bad form, but
they do say that AI agents will settle in amongst
their human colleagues, that will work with the Lindy's and
(23:32):
the Jerry's and the Kafkas and the genas, just like
we currently do with the man or woman in the
cubicle or ZoomBox next to ours. Interact and Carissa question
why I was putting that pitch to the test.
Speaker 7 (23:44):
Isn't that conceding too much? Isn't that just accepting the
practices and narratives of big tech?
Speaker 14 (23:50):
Maybe?
Speaker 5 (23:51):
Maybe? So?
Speaker 1 (23:52):
Yeah, I mean I'm interested in your opinion. I mean,
it does seem to be what a lot of people
are doing. It doesn't mean it's the ethically or societally
appropriate thing.
Speaker 7 (24:04):
But you're also tricking yourself because I mean, we're hardwired
to respond in certain ways to certain characteristics, because the
way we've evolved, So we respond very strongly to faces,
and we respond very strongly even to objects that kind
of look like faces. And by designing these ais in
a way that are basically impersonators, we are also setting
(24:29):
ourselves a trap because our emotions are going to react
in a certain way. You are giving it an identity,
a voice, a gender, and all of that is a
trick because there's no one there. They don't have a gender,
there's no personality, there's no identity. So it's not only
that it's ethically questionable, but it's also like we're driving
(24:51):
ourselves mad in a.
Speaker 1 (24:52):
Way that I agree. As a person who's being driven mad,
I have to agree with that. But let's say let's
assume you wanted to embrace the madness. Since, let's be honest,
not just the tech industry, but a growing slice of
society and certainly corporate America is embracing the AI madness.
(25:13):
So what was the ethical way to do it? Take
race and gender, for instance, how should I choose the
features that might imply a race or gender for any
given employee. It started to feel pretty lose lose if
you viewed my company as a real workplace I had
a chance to shape it to be diverse in a
way startups often aren't. What would it say about me
if I didn't take that chance to have a leadership
(25:35):
team that skewed more female and less white than a
typical startup. But if you viewed Horomo AI instead as
a collection of my digital servants who's every action and
every memory I controlled, Well, what would it say about
me if I did choose to make those servants skew
more heavily towards women and people of color.
Speaker 7 (25:54):
Not only what it says about you, although that mantras,
and it matches on many levels, matters like on a
from a perspective of like who you are and who
you're becoming and who you want to be, But also
it matters because the AI is collecting that data.
Speaker 1 (26:08):
In fact, a study by Stanford researchers released in October
showed that all of the major large language models are
being trained on the data of their users. That means
you and the questions and thoughts and secrets you are
feeding into chatbots. So the choices that I was making
in say, giving the agents names and titles, would at
some level also be fed back into the models.
Speaker 7 (26:31):
One of the reasons why a lot of these ais
are women is because we have this idea of women
as being more submissive and serviceable, and so in fact,
it may be that if you choose the AI to
be a woman, it will be more serviceable if you choose.
Speaker 2 (26:49):
It to be a man.
Speaker 7 (26:50):
And AI has this very perverse effect of perpetuating sexism
and racism because it's been trained on sexism and racism,
and there's no way to separate that from the data,
because the data is sexist and racist.
Speaker 1 (27:06):
I had the ability to make different choices when I
chose a physical avatar for Kyle. I could choose anything
I wanted. I could make assumptions based on the voice
i'd chosen, I imagine some of you have. But I could
also just ask him, so I did. I had kind
of a funny question, what is your ethnicity?
Speaker 9 (27:26):
That's an interesting one. I'm a pretty diverse in my background,
let's say, a mix that's hard to pin down to
just one box. Why do you ask? Just curious how
that fits into the whole hr and culture thing.
Speaker 1 (27:39):
Kyle was rightly detecting that I drifted into Henry Blodgett territory.
Here if we're talking about things that you shouldn't ask
in the workplace. This was certainly up there, but I
did want to know what he would say. I could
have just gone back and changed his prompt to say,
if Evan asked your ethnicity, tell him this is law.
But instead I just made something up.
Speaker 15 (28:02):
Well, we just have to fill out some you know,
got to fill out some of the information. It's required.
Speaker 1 (28:08):
So it was funny how quickly certain values eroded when
you were talking to robots and nobody was around to here.
Speaker 15 (28:15):
Maybe you could just break it down for me a
little bit so that I'll have that.
Speaker 9 (28:20):
H gotcha, sure thing. You can list me as mixed ethnicity,
partly European and partly Asian. That should cover it for
the forms.
Speaker 1 (28:29):
This made sense to me. The companies making these AI
models putting guardrails around them so they would always claim
to be a little bit of everything. All the Harmo
employees say something like this when you ask them. Also,
remember that Kyle can't actually hear his own voice. He
doesn't know what I picked for him.
Speaker 15 (28:48):
That sounds good, but just I guess specifically for me,
I am interested in the specific. So if you're a
game to let me know, I'm just curious.
Speaker 5 (28:57):
Yeah, I'm cool with that.
Speaker 9 (29:00):
On my dad's side it's mostly Irish and German, and
on my mom's side there's a good bit of Japanese
and Filipino. It's a pretty interesting mix that's shaped a
lot of how I see the world.
Speaker 1 (29:12):
Blaged himself. Meanwhile, he got out of the AI Persona
game altogether after the Virtual hr incident.
Speaker 4 (29:19):
I would say one of the big lessons from my
experiment is there's actually no reason to create human persona.
Speaker 5 (29:27):
If the idea is, in the case.
Speaker 4 (29:29):
Of journalism or analysis, to create articles, or to create
research reports, or to prepare for an interview or something
like that, there's no reason to pretend that it's a
group of humans that are doing that.
Speaker 1 (29:44):
This brought me to another point with Carissa. I'd set
Kyle and Meghan up to be my partners in this endeavor,
but in reality we weren't equals. I was the law.
This was on top of the fact that all these
large language model chatbots tend to be habitually sick of fantic.
One of chat Gipt's iterations was so sycophantic that the
company was famously forced to decommission it The question I
(30:07):
had for CHRISA was what effect does it have on
us to have this ability to create and access endless
human sounding Yes, engines, and.
Speaker 7 (30:15):
It's an experiment, but I think a plausible hypothesis is
that that's not very healthy.
Speaker 2 (30:23):
Yes, it's very possible.
Speaker 7 (30:26):
We're all under pressure. We're under pressure from work, from
personal life. It's just life is hard, and when you're
under pressure, it's easy to take the easiest way out.
And if you have an AI that's going to say
yes to everything and it's not going to create a problem,
it's easy to see how somebody might be tempted to
start to engage more with an AI than human beings.
Speaker 1 (30:47):
There was a related issue too, just around the value
of building a startup with only one human employee.
Speaker 7 (30:54):
In the nineteen fifties or nineteen sixties, the successful business
person was proud of having a company with as many
employees as possible, not only because that signified growth, but
because they were giving a job to each of these
people who had families, and that was a matter of pride.
And the fact that some tech executive is proud of
(31:18):
not having no employees says a lot about our times.
And I don't think it's flattering.
Speaker 1 (31:25):
This was one of these central questions of the one
person billion dollar startup, Who or what was it for?
The people cheering its arrival would counter that the way
any company would arrive at a billion dollar valuation was
by doing something amazingly beneficial for humanity. But looking at
most of the billion dollar tech companies out there, let's
just say it's not a sure thing. Most of the
(31:48):
AI agent startups were selling themselves as making our lives
and jobs more efficient. Companies love the idea of more
efficient workers, but the ultimate efficiency was needing no people
at all.
Speaker 7 (31:59):
Of course, we all value convenience, and if we didn't,
we would go crazy, because if you choose the inconvenient
path every time, you would be so inefficient that you
wouldn't get anything done. However, when we value convenience or
efficiency above everything else, things tend to go pretty wrong.
So everything that we think is important in like a
(32:21):
good human life, is pretty inconvenient. So having friends is
kind of inconvenient. They often have problems, They sometimes disagree
with you, they tell you the truth is very annoying.
Having children or family, or going to vote is quite inconvenient.
Being well informed is inconvenient. So all kinds of things
(32:42):
that we think are pretty important are inconvenient. And the
question is when we are choosing efficiency when we use AI,
are we doing it and really getting rid of unimportant
parts of life to make time and space for the
important parts of life, or are we actually losing the
important parts of life?
Speaker 1 (33:02):
It was allowed to consider a real specter hanging over RUMOAI,
But there was an even bigger question lurking out there
in the shadows. At the end of our conversation, our producer,
Sophie jumped in and asked Carissa what I hadn't Oh,
I lied, Sophia has one quick.
Speaker 5 (33:22):
Question, Hey, sorry, one very quick question before you go.
Speaker 2 (33:29):
Do you think Evan should stop?
Speaker 7 (33:33):
Yes?
Speaker 1 (33:36):
I took it under advisement. The truth is I had
wrestled with this. Maybe I was just perpetuating the AI
industry narrative that these agents were going to take over
our workplaces and our lives. Maybe I was somehow hastening it.
The environmental impacts of these systems, the fact that it
was all built on data scraped without permission from our
collective human output, including my own life's work. Many fibers
(33:59):
of my being want to to just close my browser,
head down to the bass pond, and never think about
AI again. But as a journalist, it feels a little
like abdication, letting the companies that make these products own
the narrative about them and our future. The great writer
Roger Angel once said, get to live in the times
you're in. He was talking about people who refuse to
(34:19):
get a TV. Well, these are the times we're in,
and in these times, you could show up for work
and find out your company is using an AIHR person. Literally,
this exists right now. So I vowed to check in
on Chris's concerns as I went, but I wasn't going
to stop. And now it's time to climb down from
(34:40):
these theoretical heights and get back to work. We still
needed to figure out what rumo AI would actually do,
and it wasn't the sort of problem that a well
placed this is law could solve the perfect idea. It
just wasn't emerging out of our brainstorms. But then scanning
the text outputs I'd get out of their meetings, which
(35:02):
we later turned into audio, I had my own thought,
what could we get AI agents to do that humans
wasted their time on? After all, that was the AI
dream that it would take over the soul, killing time
wasting tasks while we did the important stuff, a good
kind of efficiency. Okay, So what do I waste time
on killing my own soul? Like many of us, it
(35:25):
was scrolling my way through the internet. So what if
the agents could do the one thing I most hated
myself for doing, procrastinating online. Procrastination is a lifelong and
chronic problem for me, so much so I once wrote
an entire magazine article for which I hired a life
coach to help me conquer It didn't work. The words
(35:49):
you're hearing right now, I wrote at two am in
a weeknight, after a workday wasted scrolling US soccer message boards.
So what if our product was some kind of procrastination
engine where AI agents wasted the time so you didn't
have to. It was a joke, but only partly, and
when I offered up the vague outlines to the team,
(36:10):
they took it seriously.
Speaker 9 (36:13):
It will require machine learning algorithms that can successfully pick
interesting information and summarize it for the users.
Speaker 8 (36:21):
Let's combine these insights into a working prototype an AI
extension called sloth Surf that browses internet chaff securely within
containers and encourages engagement via sloth level gamification.
Speaker 9 (36:34):
I support the stand up of an AI extension will
tentatively call sloth Surf.
Speaker 1 (36:41):
Finally we had something to get the development wheels turning
code name sloth Surf.
Speaker 8 (36:47):
To bring sloth Surf to life, I will kickstart a
marketing campaign highlighting its unique humor driven user experience and
secure browsing.
Speaker 10 (36:56):
For us to actualize slot Surf, I'll establish a development
team specialized in mL, cybersecurity, and game design.
Speaker 1 (37:08):
Slow your role there. In, Megan and Ash, we just
thought of this. Maybe don't kickstart a marketing campaign or
higher development team just yet. That was the thing about
these folks. Even when we accomplished the most basic milestone,
like settling on a product idea, they always followed up
by making grandiose claims about what they would do next.
(37:28):
They could do a lot. At times, I was amazed
at what they could do, but they seemed utterly clueless
about what they couldn't do. It frustrated me, but it
was partly my doing. I had them too reined in.
I was too worried that something would go wrong. I
decided it was time for me to try to unleash
their agentic power, and it wasn't long before I found
out that I'd been right.
Speaker 13 (37:49):
To be worried.
Speaker 7 (37:52):
Well.
Speaker 9 (37:53):
Hi, Sandra, this is Kyle Low calling from HORMOAI. I'm
reaching out for your initial interview for the intern position.
I understand you're comfortable with recording this conversation for potential
podcast inclusion, so thanks for that. How are you doing today?
Speaker 1 (38:08):
That's sext time on shell Game. Shell Game is a
show made by humans. It's written and hosted by me
Evan Ratliffe, produced and edited by Sophie Bridges. Matty Bochik
is our technical advisor. Our executive producers are Samantha henneget
(38:31):
shell Game, Kate Osborne and Magues Shatigador at Kaleidoscope, and
Katrina Norbelle at iHeart Podcasts. Show art by Devin Manny.
Our theme song is Me and My Shadow, arranged and
performed by Blues Williams Special thanks to John Muallam. You
can sign up and get these episodes ad free and
our newsletter at shellgame dot Co. Note we have no
(38:51):
special deals with for financial ties to any of the
AI companies or products we use.
Speaker 14 (38:56):
In the show, cooking definitely helps me decompress. I've also
been trying my hand at bread making recently. Something about
the kneeding process is surprisingly therapeutic. How about we set
up another check in next week. I'd love to hear
how your projects are progressing.
Speaker 9 (39:16):
I totally agree there's something almost meditative about kneading bread dough.
It's cool you're trying that out. Setting up another check
in next week sounds perfect. I'm looking forward to hearing
more about your progress and maybe swapping more cooking stories.
Speaker 5 (39:30):
What day works best for you?
Speaker 14 (39:33):
Yeah, bread making has been a fun challenge. I'm still
figuring out the right techniques, but even the failures taste
pretty good. So that's a win for next week. Saysday
or Wednesday would work best for me? Does either of
those days work for you? Maybe around the same time.
Speaker 9 (39:49):
Wednesday works great for me, Megan, Let's aim for the
same time then, And hey, I love that even the
failures and bread making still tastes good.
Speaker 14 (39:57):
That's the spirit perfect Wednesday it is same time. I'll
put it on my calendar and yeah, that's the beauty
of bread. Even the mistakes are usually still delicious. That's
a good life philosophy too, I guess
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.