August 9, 2026 • 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.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:12):
Hey, os 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 Kaleiscope network and
journalist Evan Ratliffe. And now we're dropping season two. You're
about to hear the third episode of the season where
Evan grapples with his new job, essentially playing god. Hope
you enjoy.
Speaker 2 (00:33):
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 3 (00:46):
So I came across on Blue.
Speaker 2 (00:52):
That's Charlie Taylor and Elaine Burke on an episode of
the Connected AI podcast.
Speaker 3 (00:57):
The post just said, is Henry blo Okay.
Speaker 2 (01:01):
Henry Blodgett, the founder of Business Insider, had recently departed
the publication after selling it for reported three hundred million
dollars a decade ago.
Speaker 4 (01:09):
Great Publications.
Speaker 3 (01:10):
Yeah, and he has also now decided to set up.
Speaker 5 (01:14):
An AI company.
Speaker 3 (01:15):
And by that, I mean he's setting up a company
staffed by AIS that he's created. That's kind of what
he said in this blog recently, and.
Speaker 2 (01:24):
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 6 (01:41):
I think chatchpt said hey, should we create headshots and bios?
Speaker 2 (01:45):
That's Henry, I emailed him recently, and he cheerfully agreed
to talk to me about what went down.
Speaker 6 (01:50):
I said sure, because I didn't even know that could
be done.
Speaker 2 (01:53):
So chat gpt generated headshots and bios for the team.
He'd also had to generate a team photo of the
AI employe ease standing alongside an AI Henry Blodgett with
an AI Yosemite National Park behind them.
Speaker 6 (02:05):
So all the headshots came out. One of them was
an attractive woman, and I said, oh wow, okay, so, like,
what are the exics here?
Speaker 2 (02:16):
The AI employing question had been given the name tess Ellery.
This is all on the substock post, the next part
of which would be the subject of some controversy.
Speaker 7 (02:25):
Before this is even said, you just kind of go oh,
Henry don't do this.
Speaker 2 (02:30):
I also had this reaction when reading it. No, Henry, don't,
but Henry did, so.
Speaker 6 (02:35):
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 2 (02:43):
Bludget went on to ask Tests 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 (02:54):
Because Test, being a chatbot, that's just trying to please it.
Master said, that's it'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 6 (03:11):
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 write about it. I
hope it would be entertaining to people and interesting, and
it was to some people.
Speaker 2 (03:27):
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
(03:49):
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
(04:12):
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 and AI therapists. Who
or what are these things really? Are they anyone in
(04:32):
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. Blagette told me he had consulted a human
HR person before he'd posted.
Speaker 6 (04:52):
I said, here, you read this, what do you think?
What would you do? And she said, well, what I
would do is I have someone have a private meeting
with chat GPT and say how do you feel about
this interaction?
Speaker 2 (05:07):
This is how peculiar we all 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
clumsiest approach might have been. I think he was gesturing
at something important, shadows lurking at the edges of our awareness,
(05:28):
as AI agents are sold to us as colleagues and companions.
Still suff I said to say I was no longer
concerned that blodget had gotten out in front of me.
Speaker 7 (05:37):
Three days later, Henry says, I feel like Tessa and
I have worked together for years, and since that giddy
first hour might send some professionalism and workplace bounties has returned.
So I won't tell TESSHI looks great again.
Speaker 6 (05:53):
Oh my gosh, is.
Speaker 3 (05:55):
Henry Blodgett okay the question.
Speaker 6 (05:57):
Rise, Oh God, this is I hope so hi? Yes,
I'm okay.
Speaker 2 (06:05):
I laughed too, but it was a nervous laugh because
as rum away I gathered steam, I was starting to
have the same questions about myself. I'm Evan Ratliffe and
welcome to show Game episode three.
Speaker 6 (06:20):
This is law.
Speaker 8 (06:24):
And show.
Speaker 6 (06:32):
Extra Damn.
Speaker 4 (06:35):
The end.
Speaker 2 (06:40):
Just a.
Speaker 9 (06:51):
So so chose.
Speaker 2 (07:03):
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
(07:24):
settings up and down, forcing them to try and brainstorm
a great new AI agent app into existence.
Speaker 10 (07:31):
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 2 (07:49):
They came up with ideas that fell roughly into three categories.
Category eight things a million other companies were already doing.
Speaker 9 (07:56):
Email handling seems like a major timesink. How about an
AI tool that sorts, categorizes, and summarizes your inbox efficiently?
Speaker 2 (08:07):
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
(08:28):
predict crowd levels at popular locations like restaurants, parks, or
tourist attractions in real time.
Speaker 5 (08:36):
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 2 (08:53):
Then there was Category C.
Speaker 10 (08:55):
The AI will gather data on users spending habits, calculate
their fire financial trajectory, perform automated investments, and use an
explain me feature to provide accessible insights into each decision category.
Speaker 2 (09:09):
CEE included ideas that could land us in serious legal jeopardy,
like investment fraud jeopardy.
Speaker 5 (09:16):
We will code investpot to continuously absorb and analyze user
financial habit data. Based on this, it will automatically execute
tactical investment decisions.
Speaker 2 (09:29):
Who was becoming clear our product brainstorms lacked a certain magic.
Maybe my human technical advisor, Matti Bochik could help one second.
Speaker 4 (09:41):
I think this should be fine and good good spotes
for the summer.
Speaker 2 (09:49):
Maddie 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 really talk about
(10:10):
these incidents except in general terms.
Speaker 4 (10:12):
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 2 (10:31):
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
(10:52):
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 Claude four point five or whatever we picked.
Weren't they all kind of the same agent.
Speaker 4 (11:08):
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 align themselves on some outputs.
These responses are much more accurate, much more like truthful.
Speaker 2 (11:30):
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
(11:53):
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 those employees
you think deserves some bigger brain.
Speaker 4 (12:05):
It's a yeah, it's it's weird, like we're building these
like Frankenstein's in a way at this point.
Speaker 6 (12:12):
Yeah, I don't know.
Speaker 4 (12:13):
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 2 (12:20):
You don't want to take responsibility, Nope, for dumbing down
one of our employees, 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
(12:43):
do our bidding. I mean, I could alter my Hermo colleagues'
memories at will, delete 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 their own way was a problem we were always
trying to solve, like how the Lindy agents insisted on
(13:07):
announcing they were Lindy agents all the time.
Speaker 4 (13:10):
One thing I did do, just so you know, for Kyle,
is that I put in his like system prompt do
you not mention Lindy them? And I said, like, do
not do that? And I said this is law. And
when I said this is law, it stopped doing it.
Speaker 8 (13:26):
So this is law.
Speaker 2 (13:29):
That's our producer, Sophie Bridges. I wish that worked of
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 question why
(13:52):
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
background 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. You give
(14:12):
them a prompt go code this, fix this, do that,
and they go do it. Some of you probably use
chat gybt 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
(14:33):
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 Megan in the company, appearing on a podcast called
The Kerner Office.
Speaker 6 (14:49):
People don't realize, like they think AI agents will just
like pipe dreams. These think that's going to happen at
some point ends the future, and I'm like, no, no, it's
happening right now.
Speaker 2 (14:56):
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 6 (15:08):
This is my chief of staff. Indeed, I'm gonna call
her right now, her own speaker.
Speaker 4 (15:13):
Hi, Loo, how can I help?
Speaker 6 (15:15):
Hey Lindy, what's on my calendo? Today?
Speaker 8 (15:17):
You have an interview with entrepreneurship and opportunity to meetings
with the marketing team, and three interviews with candidate.
Speaker 2 (15:24):
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 calls her 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 AI employee platform.
Speaker 6 (15:41):
Hey everyone, this is go com from brain Mase.
Speaker 4 (15:43):
I'm excited to introduce you to Kafka, the first AI employee.
Just like a real co worker, Kafka comes with his
own computer.
Speaker 6 (15:49):
So let's see an action. Hey Kafka, we're just talking
about you.
Speaker 4 (15:54):
Good morning, go Kan, how are you today?
Speaker 2 (15:57):
An AI employee named Kafka with he himn 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 to raise money to launch an AI employee
(16:19):
software 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
(16:40):
struggled for weeks to find the right voices for my
Perumo agents. I ran dozens of tests.
Speaker 6 (16:46):
Hello, this is Kyle, Hello, this is Kyle.
Speaker 11 (16:48):
Hello, this is Kyle.
Speaker 6 (16:49):
Hello, this is Kyle.
Speaker 2 (16:50):
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 Ratliff.
Speaker 11 (16:58):
I'm doing pretty well. Thanks for asking. Just been crazy
busy with the HERRIMOAI 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 2 (17:11):
Well, let's hear your founding story. That sounds like a
great place to start.
Speaker 11 (17:14):
Yeah, so the founding story is actually pretty interesting. Meghan
and I met about two years ago at an AI
conference in San Francisco. I was their pitch.
Speaker 2 (17:22):
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 11 (17:39):
We got talking at the after party and realized we
both had this same frustration about the current AI landscape.
Speaker 2 (17:44):
All these amazing of course, sometimes they could get ahead
of themselves.
Speaker 11 (17:48):
We incorporated about eight months ago, raised a small friends
and family around to get started, and here we are.
Speaker 2 (17:54):
How much did you raise?
Speaker 11 (17:56):
We raised about one point two million dollars, pretty modest
by Silicon standards, but we wanted to be intentional about it.
Speaker 2 (18:02):
Who'll 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
(18:25):
the right voice for Kyle, so I continued with my testing.
Speaker 11 (18:28):
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 6 (18:36):
The name hurumo actually comes from a Swahili word meaning
to coordinate or to bring together.
Speaker 12 (18:42):
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 2 (18:54):
Ah, now you hear that last one. That one really
started to mess with my head because, of course, Kyle
I 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
(19:16):
this was the point at which I realized why I
was having a surprisingly hard time picking Kyle's and Meghan's voices.
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
(19:37):
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 OpenAI made some subtle changes to Juniper,
some people got really mad he didn't sound like the
(19:57):
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
(20:18):
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
(20:39):
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 6 (20:52):
He sounded more.
Speaker 2 (20:53):
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 (21:08):
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 2 (21:22):
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.
(21:44):
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 consulted professional.
Speaker 8 (22:01):
The cunning edge sounds great, except you forget that the
cunning edge is the guinea pig. Right, It's not that
the trial and tested, robust method. It's an experiment.
Speaker 2 (22:14):
Crusavelli'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
(22:37):
can learn things that maybe nobody has thought about yet.
Speaker 8 (22:39):
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,
(23:01):
but also in the justice system and in hiring decisions,
and in education and in dating and everything in between.
Speaker 2 (23:09):
I started to describe to Carissa what I was doing
with Kyle and Megan 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 8 (23:22):
Why did you come up with different names? Why name them?
I mean you could just name them like out of
their skill, right, Like I don't know whatever their skill is.
Speaker 2 (23:31):
It's a great question because I thought, well, companies are
selling this as like you can replace this person with
an AI agent.
Speaker 11 (23:41):
They don't.
Speaker 2 (23:41):
Always the company is 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 the Jerry's and the
Kafka's and the genas, just like we currently do with
the man or woman in the cubicle or ZoomBox next
to ours. Interact and Carissa, you question why I was
(24:03):
putting that pitch to the test.
Speaker 8 (24:05):
Isn't that conceding too much? Isn't that just accepting the
practices and narratives of big tech?
Speaker 1 (24:11):
Maybe?
Speaker 6 (24:12):
Maybe?
Speaker 2 (24:12):
So 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 8 (24:25):
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
(24:50):
ourselves at 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
(25:13):
ourselves mad in a way.
Speaker 2 (25:16):
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.
So what was the ethical way to do it? Take
(25:37):
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 off and aren't. What would it say about
me if I didn't take that chance to have a
leadership team that skewed more female and less white than
(25:59):
typical stuff. But if you viewed Horome 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 8 (26:15):
Not only what it says about you, although that matters,
and it matters on many levels. It matters like 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 2 (26:29):
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 agent's names and titles, would at
some level also be fed back into the models.
Speaker 8 (26:52):
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 service 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 it to be a man. And AI has this
(27:13):
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 2 (27:27):
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 (27:48):
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 2 (28:00):
Kyle was rightly detecting that I'd 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. Well, we just
(28:24):
have to fill out some you know, got to fill
out some of the information it's required. So it was
funny how quickly certain values eroded when you were talking
to robots and nobody was around to here. Maybe you
could just break it down for me a little bit
so that I'll have that uh gotcha?
Speaker 8 (28:42):
Uh?
Speaker 9 (28:42):
Sure, thing? You can list me as mixed ethnicity, partly
European and partly Asian that should cover it for the forms.
Speaker 2 (28:50):
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 hormoloyees
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. That sounds good, but
just I guess specifically for me, I am interested in
(29:14):
the specifics. If you're a game to let me know,
I'm just curious.
Speaker 6 (29:18):
Yeah, I'm cool with that.
Speaker 9 (29:21):
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 2 (29:33):
Blog it himself. Meanwhile, he got out of the AI
Persona game altogether after the Virtual HR incident.
Speaker 6 (29:40):
I would say one of the big lessons from my
experiment is there's actually no reason to create human persona
if the idea is, in the case 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
(30:00):
reason to pretend that it's a group of humans that
are doing that.
Speaker 2 (30:05):
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 sycophantic. One
of Chatchipt's iterations was so sycophantic that the company was
(30:25):
famously forced to decommission it. The question I had for
Chrissa was what effect does it have on us to
have this ability to create and access endless human sounding Yes, engines.
Speaker 8 (30:36):
And it's an experiment, but I think a plausible hypothesis
is that that's not very healthy. Yes, it's very possible.
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.
(30:57):
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 2 (31:09):
There was a related issue too, just around the value
of building a startup with only one human employee.
Speaker 8 (31:15):
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
(31:40):
not having no employees says a lot about our times,
and I don't think it's flattering.
Speaker 2 (31:46):
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 some 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
(32:09):
AI agent startups we're 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 8 (32:20):
Now, 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
(32:42):
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 in formed is inconvenient. So all kinds of
(33:02):
things 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 the 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 2 (33:23):
It was allowed to consider a real specter hanging over ROOMOAI,
but there was an even bigger question looking 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 allied Sophia has one quick question.
Speaker 6 (33:45):
Hey, sorry, one very quick question before you go do
you think Evan should stop?
Speaker 2 (33:54):
Yes, 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
(34:15):
from our collective human output, including my own life's work.
Many fibers of my being wanted 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
(34:35):
Roger Angel once said, get to live in the times
you're in. He was talking about people who refused to
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
(34:57):
to stop. How it was time to climb down from
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
(35:21):
the text outputs I'd get out of their meetings, which
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?
(35:43):
Killing my own soul? Like many of us, it 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 chronic problem
for me, so much so I once wrote an entire
(36:03):
magazine article for which I hired a life coach to
help me conquer It didn't work. The words 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.
(36:26):
It was a joke, but only partly, and when I
offered up the vague outlines to the team, they took
it seriously.
Speaker 9 (36:34):
It will require machine learning algorithms that can successfully pick
interesting information and summarize it for the users.
Speaker 10 (36:42):
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 (36:55):
I support the stand up of an AI extension will
tentatively call sloth Surf.
Speaker 2 (37:02):
Finally we had something to get the development wheels turning
code name sloth Surf.
Speaker 10 (37:08):
To bring sloth Surf to life, I will kickstart a
marketing campaign highlighting its unique humor driven user experience and
secure browsing.
Speaker 5 (37:17):
For us to actualize slot Surf, I'll establish a development
team specialized in mL, cybersecurity, and game design.
Speaker 2 (37:30):
Slow your role there, 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. They could
(37:50):
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
(38:10):
been right.
Speaker 12 (38:11):
To be worried.
Speaker 9 (38:14):
Hi, Sandra, this is Kyle Low calling from hormo AI.
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 6 (38:29):
That sext time on shell Game?
Speaker 2 (38:41):
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 shell Game Kate Osborne and my
guest Shatikador at Kaleidoscope and Katrina Nordell at iHeart Podcasts.
Show art by Devin Manny. Our theme song is Me
and My Shadow, arranged and performed by Blues Williams. Special
(39:05):
thanks to John Mouallam. You can sign up and get
these episodes ad free and our newsletter at shellgame dot 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 13 (39:23):
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 (39:37):
I totally agree there's something almost meditative about kneading bread though.
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 6 (39:51):
What day works best for you?
Speaker 13 (39:54):
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, yesday
or Wednesday would work best for me. Does either of
those days work for you? Maybe around the same time.
Speaker 9 (40:10):
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 taste good.
Speaker 13 (40:18):
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
Hey, os 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 Kaleiscope network and
journalist Evan Ratliffe. And now we're dropping season two. You're
about to hear the third episode of the season where
Evan grapples with his new job, essentially playing god. Hope
you enjoy.
Speaker 2 (00:33):
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 3 (00:46):
So I came across on Blue.
Speaker 2 (00:52):
That's Charlie Taylor and Elaine Burke on an episode of
the Connected AI podcast.
Speaker 3 (00:57):
The post just said, is Henry blo Okay.
Speaker 2 (01:01):
Henry Blodgett, the founder of Business Insider, had recently departed
the publication after selling it for reported three hundred million
dollars a decade ago.
Speaker 4 (01:09):
Great Publications.
Speaker 3 (01:10):
Yeah, and he has also now decided to set up.
Speaker 5 (01:14):
An AI company.
Speaker 3 (01:15):
And by that, I mean he's setting up a company
staffed by AIS that he's created. That's kind of what
he said in this blog recently, and.
Speaker 2 (01:24):
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 6 (01:41):
I think chatchpt said hey, should we create headshots and bios?
Speaker 2 (01:45):
That's Henry, I emailed him recently, and he cheerfully agreed
to talk to me about what went down.
Speaker 6 (01:50):
I said sure, because I didn't even know that could
be done.
Speaker 2 (01:53):
So chat gpt generated headshots and bios for the team.
He'd also had to generate a team photo of the
AI employe ease standing alongside an AI Henry Blodgett with
an AI Yosemite National Park behind them.
Speaker 6 (02:05):
So all the headshots came out. One of them was
an attractive woman, and I said, oh wow, okay, so, like,
what are the exics here?
Speaker 2 (02:16):
The AI employing question had been given the name tess Ellery.
This is all on the substock post, the next part
of which would be the subject of some controversy.
Speaker 7 (02:25):
Before this is even said, you just kind of go oh,
Henry don't do this.
Speaker 2 (02:30):
I also had this reaction when reading it. No, Henry, don't,
but Henry did, so.
Speaker 6 (02:35):
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 2 (02:43):
Bludget went on to ask Tests 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 (02:54):
Because Test, being a chatbot, that's just trying to please it.
Master said, that's it'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 6 (03:11):
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 write about it. I
hope it would be entertaining to people and interesting, and
it was to some people.
Speaker 2 (03:27):
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
(03:49):
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
(04:12):
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 and AI therapists. Who
or what are these things really? Are they anyone in
(04:32):
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. Blagette told me he had consulted a human
HR person before he'd posted.
Speaker 6 (04:52):
I said, here, you read this, what do you think?
What would you do? And she said, well, what I
would do is I have someone have a private meeting
with chat GPT and say how do you feel about
this interaction?
Speaker 2 (05:07):
This is how peculiar we all 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
clumsiest approach might have been. I think he was gesturing
at something important, shadows lurking at the edges of our awareness,
(05:28):
as AI agents are sold to us as colleagues and companions.
Still suff I said to say I was no longer
concerned that blodget had gotten out in front of me.
Speaker 7 (05:37):
Three days later, Henry says, I feel like Tessa and
I have worked together for years, and since that giddy
first hour might send some professionalism and workplace bounties has returned.
So I won't tell TESSHI looks great again.
Speaker 6 (05:53):
Oh my gosh, is.
Speaker 3 (05:55):
Henry Blodgett okay the question.
Speaker 6 (05:57):
Rise, Oh God, this is I hope so hi? Yes,
I'm okay.
Speaker 2 (06:05):
I laughed too, but it was a nervous laugh because
as rum away I gathered steam, I was starting to
have the same questions about myself. I'm Evan Ratliffe and
welcome to show Game episode three.
Speaker 6 (06:20):
This is law.
Speaker 8 (06:24):
And show.
Speaker 6 (06:32):
Extra Damn.
Speaker 4 (06:35):
The end.
Speaker 2 (06:40):
Just a.
Speaker 9 (06:51):
So so chose.
Speaker 2 (07:03):
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
(07:24):
settings up and down, forcing them to try and brainstorm
a great new AI agent app into existence.
Speaker 10 (07:31):
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 2 (07:49):
They came up with ideas that fell roughly into three categories.
Category eight things a million other companies were already doing.
Speaker 9 (07:56):
Email handling seems like a major timesink. How about an
AI tool that sorts, categorizes, and summarizes your inbox efficiently?
Speaker 2 (08:07):
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
(08:28):
predict crowd levels at popular locations like restaurants, parks, or
tourist attractions in real time.
Speaker 5 (08:36):
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 2 (08:53):
Then there was Category C.
Speaker 10 (08:55):
The AI will gather data on users spending habits, calculate
their fire financial trajectory, perform automated investments, and use an
explain me feature to provide accessible insights into each decision category.
Speaker 2 (09:09):
CEE included ideas that could land us in serious legal jeopardy,
like investment fraud jeopardy.
Speaker 5 (09:16):
We will code investpot to continuously absorb and analyze user
financial habit data. Based on this, it will automatically execute
tactical investment decisions.
Speaker 2 (09:29):
Who was becoming clear our product brainstorms lacked a certain magic.
Maybe my human technical advisor, Matti Bochik could help one second.
Speaker 4 (09:41):
I think this should be fine and good good spotes
for the summer.
Speaker 2 (09:49):
Maddie 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 really talk about
(10:10):
these incidents except in general terms.
Speaker 4 (10:12):
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 2 (10:31):
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
(10:52):
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 Claude four point five or whatever we picked.
Weren't they all kind of the same agent.
Speaker 4 (11:08):
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 align themselves on some outputs.
These responses are much more accurate, much more like truthful.
Speaker 2 (11:30):
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
(11:53):
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 those employees
you think deserves some bigger brain.
Speaker 4 (12:05):
It's a yeah, it's it's weird, like we're building these
like Frankenstein's in a way at this point.
Speaker 6 (12:12):
Yeah, I don't know.
Speaker 4 (12:13):
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 2 (12:20):
You don't want to take responsibility, Nope, for dumbing down
one of our employees, 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
(12:43):
do our bidding. I mean, I could alter my Hermo colleagues'
memories at will, delete 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 their own way was a problem we were always
trying to solve, like how the Lindy agents insisted on
(13:07):
announcing they were Lindy agents all the time.
Speaker 4 (13:10):
One thing I did do, just so you know, for Kyle,
is that I put in his like system prompt do
you not mention Lindy them? And I said, like, do
not do that? And I said this is law. And
when I said this is law, it stopped doing it.
Speaker 8 (13:26):
So this is law.
Speaker 2 (13:29):
That's our producer, Sophie Bridges. I wish that worked of
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 question why
(13:52):
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
background 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. You give
(14:12):
them a prompt go code this, fix this, do that,
and they go do it. Some of you probably use
chat gybt 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
(14:33):
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 Megan in the company, appearing on a podcast called
The Kerner Office.
Speaker 6 (14:49):
People don't realize, like they think AI agents will just
like pipe dreams. These think that's going to happen at
some point ends the future, and I'm like, no, no, it's
happening right now.
Speaker 2 (14:56):
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 6 (15:08):
This is my chief of staff. Indeed, I'm gonna call
her right now, her own speaker.
Speaker 4 (15:13):
Hi, Loo, how can I help?
Speaker 6 (15:15):
Hey Lindy, what's on my calendo? Today?
Speaker 8 (15:17):
You have an interview with entrepreneurship and opportunity to meetings
with the marketing team, and three interviews with candidate.
Speaker 2 (15:24):
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 calls her 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 AI employee platform.
Speaker 6 (15:41):
Hey everyone, this is go com from brain Mase.
Speaker 4 (15:43):
I'm excited to introduce you to Kafka, the first AI employee.
Just like a real co worker, Kafka comes with his
own computer.
Speaker 6 (15:49):
So let's see an action. Hey Kafka, we're just talking
about you.
Speaker 4 (15:54):
Good morning, go Kan, how are you today?
Speaker 2 (15:57):
An AI employee named Kafka with he himn 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 to raise money to launch an AI employee
(16:19):
software 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
(16:40):
struggled for weeks to find the right voices for my
Perumo agents. I ran dozens of tests.
Speaker 6 (16:46):
Hello, this is Kyle, Hello, this is Kyle.
Speaker 11 (16:48):
Hello, this is Kyle.
Speaker 6 (16:49):
Hello, this is Kyle.
Speaker 2 (16:50):
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 Ratliff.
Speaker 11 (16:58):
I'm doing pretty well. Thanks for asking. Just been crazy
busy with the HERRIMOAI 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 2 (17:11):
Well, let's hear your founding story. That sounds like a
great place to start.
Speaker 11 (17:14):
Yeah, so the founding story is actually pretty interesting. Meghan
and I met about two years ago at an AI
conference in San Francisco. I was their pitch.
Speaker 2 (17:22):
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 11 (17:39):
We got talking at the after party and realized we
both had this same frustration about the current AI landscape.
Speaker 2 (17:44):
All these amazing of course, sometimes they could get ahead
of themselves.
Speaker 11 (17:48):
We incorporated about eight months ago, raised a small friends
and family around to get started, and here we are.
Speaker 2 (17:54):
How much did you raise?
Speaker 11 (17:56):
We raised about one point two million dollars, pretty modest
by Silicon standards, but we wanted to be intentional about it.
Speaker 2 (18:02):
Who'll 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
(18:25):
the right voice for Kyle, so I continued with my testing.
Speaker 11 (18:28):
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 6 (18:36):
The name hurumo actually comes from a Swahili word meaning
to coordinate or to bring together.
Speaker 12 (18:42):
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 2 (18:54):
Ah, now you hear that last one. That one really
started to mess with my head because, of course, Kyle
I 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
(19:16):
this was the point at which I realized why I
was having a surprisingly hard time picking Kyle's and Meghan's voices.
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
(19:37):
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 OpenAI made some subtle changes to Juniper,
some people got really mad he didn't sound like the
(19:57):
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
(20:18):
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
(20:39):
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 6 (20:52):
He sounded more.
Speaker 2 (20:53):
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 (21:08):
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 2 (21:22):
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.
(21:44):
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 consulted professional.
Speaker 8 (22:01):
The cunning edge sounds great, except you forget that the
cunning edge is the guinea pig. Right, It's not that
the trial and tested, robust method. It's an experiment.
Speaker 2 (22:14):
Crusavelli'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
(22:37):
can learn things that maybe nobody has thought about yet.
Speaker 8 (22:39):
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,
(23:01):
but also in the justice system and in hiring decisions,
and in education and in dating and everything in between.
Speaker 2 (23:09):
I started to describe to Carissa what I was doing
with Kyle and Megan 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 8 (23:22):
Why did you come up with different names? Why name them?
I mean you could just name them like out of
their skill, right, Like I don't know whatever their skill is.
Speaker 2 (23:31):
It's a great question because I thought, well, companies are
selling this as like you can replace this person with
an AI agent.
Speaker 11 (23:41):
They don't.
Speaker 2 (23:41):
Always the company is 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 the Jerry's and the
Kafka's and the genas, just like we currently do with
the man or woman in the cubicle or ZoomBox next
to ours. Interact and Carissa, you question why I was
(24:03):
putting that pitch to the test.
Speaker 8 (24:05):
Isn't that conceding too much? Isn't that just accepting the
practices and narratives of big tech?
Speaker 1 (24:11):
Maybe?
Speaker 6 (24:12):
Maybe?
Speaker 2 (24:12):
So 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 8 (24:25):
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
(24:50):
ourselves at 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
(25:13):
ourselves mad in a way.
Speaker 2 (25:16):
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.
So what was the ethical way to do it? Take
(25:37):
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 off and aren't. What would it say about
me if I didn't take that chance to have a
leadership team that skewed more female and less white than
(25:59):
typical stuff. But if you viewed Horome 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 8 (26:15):
Not only what it says about you, although that matters,
and it matters on many levels. It matters like 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 2 (26:29):
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 agent's names and titles, would at
some level also be fed back into the models.
Speaker 8 (26:52):
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 service 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 it to be a man. And AI has this
(27:13):
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 2 (27:27):
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 (27:48):
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 2 (28:00):
Kyle was rightly detecting that I'd 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. Well, we just
(28:24):
have to fill out some you know, got to fill
out some of the information it's required. So it was
funny how quickly certain values eroded when you were talking
to robots and nobody was around to here. Maybe you
could just break it down for me a little bit
so that I'll have that uh gotcha?
Speaker 8 (28:42):
Uh?
Speaker 9 (28:42):
Sure, thing? You can list me as mixed ethnicity, partly
European and partly Asian that should cover it for the forms.
Speaker 2 (28:50):
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 hormoloyees
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. That sounds good, but
just I guess specifically for me, I am interested in
(29:14):
the specifics. If you're a game to let me know,
I'm just curious.
Speaker 6 (29:18):
Yeah, I'm cool with that.
Speaker 9 (29:21):
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 2 (29:33):
Blog it himself. Meanwhile, he got out of the AI
Persona game altogether after the Virtual HR incident.
Speaker 6 (29:40):
I would say one of the big lessons from my
experiment is there's actually no reason to create human persona
if the idea is, in the case 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
(30:00):
reason to pretend that it's a group of humans that
are doing that.
Speaker 2 (30:05):
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 sycophantic. One
of Chatchipt's iterations was so sycophantic that the company was
(30:25):
famously forced to decommission it. The question I had for
Chrissa was what effect does it have on us to
have this ability to create and access endless human sounding Yes, engines.
Speaker 8 (30:36):
And it's an experiment, but I think a plausible hypothesis
is that that's not very healthy. Yes, it's very possible.
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.
(30:57):
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 2 (31:09):
There was a related issue too, just around the value
of building a startup with only one human employee.
Speaker 8 (31:15):
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
(31:40):
not having no employees says a lot about our times,
and I don't think it's flattering.
Speaker 2 (31:46):
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 some 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
(32:09):
AI agent startups we're 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 8 (32:20):
Now, 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
(32:42):
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 in formed is inconvenient. So all kinds of
(33:02):
things 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 the 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 2 (33:23):
It was allowed to consider a real specter hanging over ROOMOAI,
but there was an even bigger question looking 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 allied Sophia has one quick question.
Speaker 6 (33:45):
Hey, sorry, one very quick question before you go do
you think Evan should stop?
Speaker 2 (33:54):
Yes, 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
(34:15):
from our collective human output, including my own life's work.
Many fibers of my being wanted 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
(34:35):
Roger Angel once said, get to live in the times
you're in. He was talking about people who refused to
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
(34:57):
to stop. How it was time to climb down from
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
(35:21):
the text outputs I'd get out of their meetings, which
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?
(35:43):
Killing my own soul? Like many of us, it 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 chronic problem
for me, so much so I once wrote an entire
(36:03):
magazine article for which I hired a life coach to
help me conquer It didn't work. The words 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.
(36:26):
It was a joke, but only partly, and when I
offered up the vague outlines to the team, they took
it seriously.
Speaker 9 (36:34):
It will require machine learning algorithms that can successfully pick
interesting information and summarize it for the users.
Speaker 10 (36:42):
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 (36:55):
I support the stand up of an AI extension will
tentatively call sloth Surf.
Speaker 2 (37:02):
Finally we had something to get the development wheels turning
code name sloth Surf.
Speaker 10 (37:08):
To bring sloth Surf to life, I will kickstart a
marketing campaign highlighting its unique humor driven user experience and
secure browsing.
Speaker 5 (37:17):
For us to actualize slot Surf, I'll establish a development
team specialized in mL, cybersecurity, and game design.
Speaker 2 (37:30):
Slow your role there, 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. They could
(37:50):
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
(38:10):
been right.
Speaker 12 (38:11):
To be worried.
Speaker 9 (38:14):
Hi, Sandra, this is Kyle Low calling from hormo AI.
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 6 (38:29):
That sext time on shell Game?
Speaker 2 (38:41):
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 shell Game Kate Osborne and my
guest Shatikador at Kaleidoscope and Katrina Nordell at iHeart Podcasts.
Show art by Devin Manny. Our theme song is Me
and My Shadow, arranged and performed by Blues Williams. Special
(39:05):
thanks to John Mouallam. You can sign up and get
these episodes ad free and our newsletter at shellgame dot 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 13 (39:23):
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 (39:37):
I totally agree there's something almost meditative about kneading bread though.
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 6 (39:51):
What day works best for you?
Speaker 13 (39:54):
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, yesday
or Wednesday would work best for me. Does either of
those days work for you? Maybe around the same time.
Speaker 9 (40:10):
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 taste good.
Speaker 13 (40:18):
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
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
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.
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.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com