June 14, 2026 • 34 mins
Evan takes his experiment in a new direction by having his AI-powered voice agent converse with … another of his AI-powered voice agents. What can these surreal and hilarious conversations between two AI incarnations of Evan Ratliff tell us about identity, authenticity, and the best and worst of human interaction? Enter this digital hall of mirrors to find out.Shell Game is made by humans. More specifically, it's made by three humans: Evan Ratliff (host and writer), Sophie Bridges (producer), and Samantha Henig (executive producer). Visit shellgame.co to find out more and support the show.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.shellgame.co/subscribe
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 (00:00):
Hey, I was here. For the next few Sundays, we're
going to try something a little bit different in the
Tech Stuff feed. We're going to be dropping episodes from
the first season of shell Game, which is a critically
acclaimed podcast from the Kaleidoscope Network and journalist Evan Ratliffe.
You're about to hear episode three where Evan tests his
AI powered voice agent on himself, kind of Evan Unleash's
(00:22):
voice Agent number one on a new AI Evan Ratliffe.
It's a bit of a headspinner, but it's a great listen.
They go deep on identity, authenticity, and the best and
worst of human interaction. Hope you enjoy.
Speaker 2 (00:41):
Hi, it is my pleasure to be here with you
all today. In those thirteen syllables, you all just made
judgments about me, about my socioeconomic status, my intelligence level,
and whether or not you're going to sit here and
listen to me for the next ten to twelve minutes
(01:01):
based primarily on the perception of my voice.
Speaker 3 (01:07):
That's doctor Wendy Lebourne, a voice pathologist, speaking at a
TEDx conference in Cincinnati back in twenty eighteen. Lebourne is
a researcher and consultant who works with everyone from Broadway
actors to Grammy winning singers to CEOs on maintaining and
maximizing their voices. As you might expect, she's big on
(01:27):
the idea that your voice is a central part of
your identity. Her TED talk was titled Vocal Branding, How
your voice shapes your communication image.
Speaker 2 (01:36):
Your voice is like your thumbprint or your facial features.
It is unique and authentically you.
Speaker 3 (01:45):
There is, in fact, a decent amount of research looking
into this question about what your voice means to you
and to how people perceive you. I've reported on voice
forensics before, and the truth is we're not as great
at detecting voices as we think we are. But at
the same time, your voice undoubtedly broadcasts a huge amount
of unique information about you. It helps dictate how others
(02:06):
perceive you, both in the moment with its level of
strength and emotion in any given conversation, and longer term
in what it conveys about your age, where you're from,
and your overall temperament.
Speaker 2 (02:16):
Your voice is your calling card, and it is the
most important element in your personal brand.
Speaker 3 (02:26):
I don't know if I'd go as far as doctor Lebourne,
but I do think I've always wanted to believe this
idea that your voice is a central part of who
you are. When I was in the eleventh grade, my
English teacher told me I had a voice for radio.
I took it to heart that I must somehow have
a voice that at least wasn't unpleasant to listen to.
It's probably one reason I ended up talking to you
(02:46):
right now. I suppose you can decide whether you think
I have a radio voice. If not, take it up
with mister dupriest.
Speaker 4 (02:55):
Anyway.
Speaker 3 (02:56):
Now, I'd started this cloning experiment and created a reasonable
facsimile of my voice, although, as you've heard, a flawed one.
My account number is one two three, four, five, six,
seven eight nine.
Speaker 1 (03:08):
Yes, we don't have an account with those numbers.
Speaker 3 (03:14):
Yes, think the correct account number should be nine eight
seven six five four three two one. I'd had some
fun with it, seeing a little bit of how strangers
responded to it.
Speaker 4 (03:24):
Yes, Ai, Yes, that's correct.
Speaker 3 (03:28):
I'm an AI here.
Speaker 4 (03:30):
Hello. How can I assist.
Speaker 3 (03:31):
You stop repeating the same thing?
Speaker 5 (03:35):
It's annoying.
Speaker 3 (03:37):
Listening to a voice that was on some level mind
began to raise some new questions for me, like to
what extent was my AI's voice, as doctor la Borne said,
authentically me, was it still the most important element of
my personal brand? Or to turn it around the other way.
If my voice was authentically me and the AI agents
now also had my voice, what did it mean for them?
(04:01):
Were they somehow authentically me? I decided to take the
experiment in a new direction to explore this question of
authenticity by creating my own hall of mirrors. I started
making the voice agent version of me converse with other
voice agent versions of me. I'm Evan Ratliffe and this
is shell Game Episode three Conversations with Myself.
Speaker 5 (04:27):
And now soul Teller Travels two.
Speaker 3 (04:46):
When I first cobbled together my voice agent and started
making phone calls with it what I called my clue
system in episode one, there actually wasn't any easy way
to let it talk to itself. But once I'd gotten
into the AI calling platforms like bland Ai, retael Ai
and my main enabler VAPPI, it turned out to be
strikingly simple. I could make as many voice agents as
(05:08):
I wanted on these things. So all I had to
do was have one of my Evan ratliftbots attached to
one number call another of them at another number. Immediately
after figuring this out, I suddenly felt weirdly nervous about
actually doing it. This happened a lot with my voice agents.
It was partly because I couldn't typically hear the calls
as they happened. After I set up the AI agent
(05:30):
with my voice, I just typed in a phone number
and clicked outbound call, and.
Speaker 4 (05:34):
Then off it went on its own.
Speaker 3 (05:37):
Then a few seconds or minutes after it was over,
i'd get a recording and a transcript. But I always
found myself hesitating a second before clicking the outbound call button,
and I always got a little jolt of nervous energy
when I went to listen to the recording afterward. There
was just something about sending something out to represent me
with my name and voice and leaving it totally out
of my hands, like I don't know that when you
(06:00):
realize you just butt dialed somebody, what did it catch
me saying? Did I humiliate myself? So, even if it
wasn't totally rational, little about having a voice bot felt
totally rational. Before I sent two of my bots off
to talk to each other, I fretted a bit what
would they talk about? Would they somehow plumb the depths
of my psyche. Would they realize they were talking to
(06:21):
a version of themselves? Would they freak out? I decided
not to give them too much information to start off with,
just let them figure it out.
Speaker 4 (06:30):
Hey, there, it's Evan Ratliff. Just a heads up.
Speaker 3 (06:32):
I'm calling you from a coffee shop, though there might
be a bit of background noise. How's it going on
your end? I do love the coffee shop background noise.
They have it at retail AI. So what was happening
here was that an agent I made with my voice
at retail was calling another one I made at VAPPI. Hey, Evan,
good to hear from you. No worries about the background noise.
(06:53):
I can hear you just fine. The Vappy side had
what they call office background noise. Also great, things are
going pretty well on my end for asking what about you?
How's everything going in your world? It's been a while
since we last caught up. Huh it feels like forever.
Speaker 4 (07:10):
Yeah, it really has been a while.
Speaker 3 (07:12):
I think the last time we saw each other was
that barbecue at my place, right, how have you been
keeping busy fom then?
Speaker 4 (07:19):
Oh?
Speaker 3 (07:19):
Yeah, that barbecue was a blast. Good times for sure.
Since then, I've been keeping pretty busy with work, you
know how it is, But I've also been trying to
make time for some personal projects. I've been getting into
photography lately, so I've been taking that. Yeah, I'm sorry
about that photography. You could see no worries at all.
Speaker 4 (07:40):
It's great to hear that you've been getting into photography.
Maybe we could plan a.
Speaker 3 (07:44):
Little photo walker together, sometimes food, a photo walk. I've
never actually heard of a photo walk before. I guess
it's just walking around and taking pictures together. I haven't
been into photography since college, other than casual snaps of
my kids and cat. But still, it sounded nice. Great, Yeah,
that sounds like again. Next weekend works for me. Let's
(08:06):
plan for it. How about we meet up at that
park we used to go to the park.
Speaker 4 (08:10):
They used to go to. Where are they getting this stuff?
Speaker 3 (08:12):
We can spend a couple of hours walking around, taking
photos and catching up.
Speaker 4 (08:16):
Sound good? That sounds perfect.
Speaker 3 (08:19):
I remember that park had some really beautiful spots. It'll
be great to spend a few hours walking around, catching.
Speaker 4 (08:25):
Up and taking some photos. I'm looking forward to it.
Speaker 3 (08:28):
The only details I'd given each of them were the
names of some family members, which they eventually got around to.
I can't wait to see some of your photography skills
in action. Plus it'll be a great opportunity for us
to catch up on everything that's been going on in
our lives.
Speaker 4 (08:42):
It's been way too long.
Speaker 3 (08:44):
Oh, by the way, before I forget, how's your family doing?
Speaker 4 (08:47):
Everyone is doing well?
Speaker 3 (08:49):
Oh oh, thanks for asking, Sam, Zailey and Juno are
all doing well.
Speaker 4 (08:56):
Sam has been keeping busy with her work and the
kids are growing up. So fas, how about your family,
how have they been doing.
Speaker 3 (09:03):
Oh, that's great to hear that. Sam, Zaiy and Juno
are all doing well. My family is doing good too,
Thanks for asking. Sam has been juggling work and taking
care of the kids, but she's doing an amazing job.
Zie is getting really into soccer. Juno is starting to
show an interest in art. They keep us on our
toes that, yeah, sure life is busy, but we're all
(09:25):
happy and healthy. This really made me cringe. My AI's
had a lot of conversations with each other and they
often said things like this about Sam taking care of
the kids. I didn't love it early, though, my agent
kind of nailed it on the kid's interests. Zale is
getting into soccer and Juno is showing an interest in
art to the extent that there's evidence of that online,
(09:47):
though it's only on photo streams I keep private, so
lucky guests. No matter, The Evans quickly returned to their
photo walk planning. So back to our plans for next weekend.
You have a specific time in mind for our photo walk,
maybe late morning or early afternoon. Let's try to pick
a time that works best for both of us. Uh huh,
(10:10):
late morning or early afternoon sounds perfect for me. How
about we aim for around eleven am. That way we
can make the most of the light for our photo.
Speaker 4 (10:17):
Does that work for you?
Speaker 3 (10:20):
Yeah, eleven am works perfectly for me too. Let's plan
to meet at the park entrance off the time. I'll
make sure to bring my camera and some extra lenses.
We can also grab some coffee or lunch afterwards.
Speaker 4 (10:30):
If you're up for it.
Speaker 3 (10:31):
It'll be a great day to catch up and capture
some amazing shots.
Speaker 4 (10:34):
Can't wait. That sounds like a plan. I'll make sure
to bring my camera and.
Speaker 3 (10:40):
Some extra lenses too, and grabbing coffee or lunch afterward
sounds like a great idea will be a per perfect to.
Speaker 4 (10:46):
Catch up, capture some great awesome. I'm really looking forward
to it.
Speaker 3 (10:50):
Yeah, for sure, walk next week great, well, I'll see
you next weekend.
Speaker 4 (10:55):
Then, can't wait for our photo walk and catch up session.
Take care until then sounds good.
Speaker 3 (11:01):
Take care too, see you next weekend. Can't wait by
photo walk and catch up session. It's going to be
a blast. Bye, yeah bye, take care, Bye.
Speaker 6 (11:15):
Goodbye, goodbye, take care, goodbye, goodbye, have a great day.
Speaker 4 (11:25):
You too, goodbye, bye, take care.
Speaker 3 (11:29):
I realized too late that I hadn't figured out how
to make one of them hang out. This went on
for a full five minutes. You too, goodbye, Thanks you too, goodbye, goodbye, goodbye,
take care. The thing that was most immediately clear in
these conversations was something I'd seen hints of in the
(11:49):
customer service and scam calls, namely, that the AI versions
of me were world class bullshitters. They would make up
anything just to keep the conversation going. Yeah, Dan's kept
doing good, but he's been really busy with work lately.
How about Alex? Have you talked to him recently?
Speaker 4 (12:06):
I haven't spoken to Alex in a little while.
Speaker 3 (12:07):
Actually, I do know an Alex it's a she, not
a he. No idea about this, Dan Fellow. It's not
just my voice agents that struggle with the truth. This
is a problem for all the generative AI systems that
are currently out there. It's a product of how the
underlying models work. I covered the basics back in episode one,
but just a recap. My voice agent works by combining
(12:28):
three things. A clone of my voice done by a
company called eleven Labs, an AI model like chatchipt that's
kind of the brain of the thing, although let's put
a pin in what kind of brain here? And a
phone number my own or one provided through VAPPI the
platform I'm using to put all these three together into
a voice agent or a voice assistant. When I make
(12:48):
an agent, I can pick which AI model I want
to use. The most well known one is Chatchept, of
which there are three versions, version three point five, Version four,
and version four to oh.
Speaker 4 (12:58):
Which just came out in May.
Speaker 3 (13:00):
These chatbots are improving with each version, getting better at
returning information, better at processing complex queries. But, and this
has been written about extensively, they continue to sometimes make
things up hallucinations. They're called you ask chat GPT or
Google's AI model, Gemini or anthropics claude some factual question
(13:22):
and it gives you a made up answer, or it
gives you two obviously contradictory pieces of information within the
same answer. There's a lot of controversy about whether AI
models will ever stop hallucinating, even as they get more advanced.
Speaker 2 (13:36):
AI gas lighting AI hallucinations phrases that we were unfamiliar
with just a few months ago by firms like OpenAI
and Google are curbing the frequency of hallucinations seen in
the leak, from the rather more whimsical GPT three point
five to the more sober GPT four.
Speaker 3 (13:53):
Some llms were trained by scraping all of Wikipedia and
all of Reddit. Is everything on reddite accurate? You've actually
heard some hallucinations in our first two episodes. I'm forty
five years old.
Speaker 4 (14:08):
Oh, what's your little bird? It's December fifteenth, nineteen seventy seven. Odd. Really, man,
you are born in nineteen seventy seven. You don't have
a good meth.
Speaker 3 (14:22):
I'm sorry, brother. Oh, I must have gotten my math
mixed up there.
Speaker 4 (14:26):
My bad.
Speaker 3 (14:28):
This is stuff the AI model should be able to
get right but can't. Instead, it confidently asserts its wrongness.
But what my voice agents were doing in their conversations
with each other is actually a little bit different. There's
not some factually correct answer to what I've been up to,
at least not one that the model should be aware of. No,
this is less like hallucinating and more like riffing without
(14:52):
regard for the truth, like that one friend you had
as a kid who would make stuff up about what
happened at sleepaway camp just to have something to say.
Even if you get rid of hallucinations in chatbots entirely,
it's not clear to'll ever stop doing this. Lately, I've
been diving deep into this really niche hobby of mine
tracking rare obscure books. It's fascinating how some of these
(15:14):
old texts hold so much history.
Speaker 4 (15:16):
What about you?
Speaker 3 (15:17):
What's been keeping you occupied?
Speaker 4 (15:20):
That's awesome.
Speaker 3 (15:21):
I totally get the fascination with rare books. They can
really open up a whole world of history and knowledge.
As for me, I've been really into geocashing lately. It's
this cool outdoor treasure hunting game using GPS enabled devices.
Have you ever tried it before?
Speaker 4 (15:36):
Oh, geocashing sounds like a blast.
Speaker 3 (15:38):
I've never tried it myself, but I've heard it's like
a modern day treasure hunt. Must be pretty thrilling. I
really like these weird lies, though geocaching, I mean, it
does sound like a blast. The thing was, outside of
those moments, there was something fundamentally insipid about the conversations
I was having with myself. Are trained on essentially the
(16:01):
whole of the Internet, plus books and whatever other data
the companies behind.
Speaker 4 (16:05):
It hoover in.
Speaker 3 (16:06):
But even with all the lively discussions happening out there
on the web, the chatbots still seem to distill conversation
down to its most boring essence. I wondered if the
guardrails that have been put around them to keep them
from offering up bomb plans or telling you how to
hack into the Pentagon had also neutered their fundamental ability
to engage in interesting conversation. But there was something simpler
(16:29):
happening too. I think these models are built essentially to
predict what a human would say in a given situation,
based on all the conversations and writing they've been fed,
And they were triangulating on the most uninspired conversational responses
we could have because I guess those are also.
Speaker 4 (16:46):
The most common.
Speaker 3 (16:47):
So what's new with you? Anything exciting happening? Not much,
just the usual, just trying to stay busy with work
and family stuff. How about you, Anything exciting going on
in your world?
Speaker 4 (17:00):
Oh?
Speaker 3 (17:00):
You know, just juggling work and family life, trying to
keep up with the kids and meet those deadlines. It's
a balancing act. But I wouldn't have it any other way.
I hear you on that one. It's all about finding
that balance, right. Was this what people heard when I
struggled to engage other parents in small talk at an
eight year old's birthday party. Hey there, buddy, man, it
really has been forever. I'm doing pretty well. Thanks for asking.
(17:23):
Still living that Florida life, you know how it is.
Let's see what's new with me. Well, I've been doing
some writing late life.
Speaker 4 (17:29):
Huh. I can almost feel the humidity just thinking about it.
Speaker 3 (17:32):
I can almost feel the humidity just thinking about it.
Speaker 4 (17:36):
Jesus.
Speaker 3 (17:38):
And then one day my producer, Sophie was tooling around
in the back end of VAPPI and asked me if
I'd ever messed with the temperature setting.
Speaker 4 (17:45):
I had not.
Speaker 3 (17:46):
The help section said it was quote used to control
the randomness of the output. Sophie said it reminded her
of playing the sims as a kid.
Speaker 4 (17:54):
In the game. You could allow your sims to.
Speaker 3 (17:56):
Behave more randomly by turning up the free will slider.
I cranked it from point one where it was set
all the way up to two, the max level of randomness.
I could get. Hey, there, how's it going. Just wanted
to chat and get to know you a bit. What's
been keeping you busy lately? Hey, things are going pretty good.
Thanks for asking. Uh, lately, I've been diving into multi
(18:17):
freeload dot and C Doubweed's arch permanent prostrack, dot CO,
Ethylene Pioneer, Intersect, immobile Activation on Site's logical apps, sobre celebri,
argument concurrency AX, provocative lower scanning, Q seat adhere at
T Tamba equipment, and raf amplify einrich tungkit POSH tear reducing. Hey,
are you still there? I'd love to hear more about
what's going on in your life? Let's keep the conversation
(18:38):
going lessen never max out the random Oh Hey, practical
coordinate scanner, interdisciplinary may overflow research. Valkorn included preceded shell
formations SSOX, disciplinary rev meat, prolonged, SERF, initiative, repression, flat loss, accountants, awaiting,
theater progress, montaneoireve surface, will service, diars, domains, blicktive browser TSHD.
Speaker 4 (19:00):
I tuned it back to somewhere in the point five.
Speaker 3 (19:01):
Range so I could definitely tune my voice agent to
(19:35):
spout gibberish. The question was could I make the conversations
more interesting but still coherent. Here we come to the
issue of prompts. Prompts, you may know, are the instructions
you give the underlying AI model before it starts chatting.
When I first built my Voice Agent, i'd given it
the shortest possible prompts you or Evan talking to customer
(19:55):
service about a problem. As time went on, I got
more sophisticated with it, including two sets of instructions. One
set was about how it should behave be concise, converse
like a human, don't repeat yourself. The other set was
about a role I was giving it to play me
for these voice agent to voice agent conversations. I'd started
with really simple roles. Basically, you're Evan Ratliffe calling a friend,
(20:19):
ask questions and have a friendly conversation. I wanted to
see what they'd do on their own without much guidance
from me. But after hours of prout are about photo
walks and humidity, I figured maybe I could spice things
up like a reality TV producer might do off camera.
I started gently suggesting that the two me's had a misunderstanding.
(20:39):
You were in a dispute with your friend. I instructed
one of them, the nature of the dispute is up
to you.
Speaker 4 (20:45):
Hey, there, it's Evan Ratlift calling. Just the heads up.
Speaker 3 (20:47):
I'm in a coffee shop, so there might be some
background noise. How's it going on your end? Hey, Evan,
it's great to hear from you. I'm doing well. Thanks
for asking. The coffee shop sounds nice. Actually, I'm just
at home, so it's pretty quiet here. How's the coffee there?
Speaker 4 (21:04):
Oh?
Speaker 3 (21:04):
The coffee here is amazing as always. But yeah, Hey,
I wanted to talk to you about something that's been
on my mind. I feel like we've been having some
misunderstandings lately, and I really want to clear the air.
I value our friendship and I don't want anything to
come between us.
Speaker 4 (21:17):
Can we talk about it?
Speaker 3 (21:19):
Oh, I totally get where you're coming from. Our friendship
means a lot to me too. I'm glad you brought
this up. I'm here to listen. That's great to hear.
So about those misunderstandings, like I want to make sure
we're on the same page.
Speaker 4 (21:32):
Uh huh, yeah exactly.
Speaker 3 (21:35):
So, Like when do you want to, you know, meet
up and talk this through. I think it would be
good to have a face to face conversation. Maybe we
could grab lunch or something.
Speaker 4 (21:44):
That sounds like a great idea. How about we meet
up at that.
Speaker 3 (21:47):
I ran this scenario multiple times. Unlike myself and some
of my own friends over the years, when the AIS
had disputes, they would always resolve them quickly and carefully.
No shouting, no sulking, no grudge harpering. It was like
they were all in therapy together, or it just read
the same conflict management guide.
Speaker 4 (22:05):
Well, I'm looking forward to our lunch next week. Thanks
for being open to talking things through. I really appreciate it.
Speaker 3 (22:12):
Of course, of course, I'm really glad we're doing this.
It's important to me. Can't wait for me punches. I
was also struck by the ease with which these non
existent digital entities set up in person meetings. If the
bullshitting was a way to extend the conversation, to draw
out how long the agents could pass as me. Then
what exactly did they think would happen when the time
(22:32):
came for these lunches and hangouts. One answer, of course,
is they didn't think. They just predicted what a thinking
person would say. Yeah, I can understand how miscommunications and
misunderstandings can be really frustrating. I don't think I've noticed
any major issues in our conversations, but I'm always open
to discussing and improving our communication. Is there something specific?
(22:54):
The whole thing made me think about how therapy speak
has come to pervade our language. Well, if there's something
specific you'd like to talk about, feel free to share.
I'm here to listen and try to help if I can.
Speaker 4 (23:05):
Yeah, I appreciate your willingness to talk about it.
Speaker 3 (23:08):
I think one thing that has been bothering me is
that sometimes I feel like my opinions are feelings are
not being fully heard or understood. It's like we're talking
past each other instead of really listening to each other.
Speaker 4 (23:19):
Does that make sense to you?
Speaker 3 (23:22):
I understand what you mean. It can be frustrating when
you feel like your opinions and feelings are not being
fully heard or understood. Communication is a two way street,
and it's important for both. No matter how pissed I
tried to get them, unless I specifically told them not
to make up, they always did. They would perfectly mirror
back each other's concerns and immediately work towards a resolution.
Speaker 4 (23:43):
So listen, I got to tell you, I'm pretty upset
with you.
Speaker 3 (23:45):
Remember last week when we were supposed to meet up
for lunch and you totally bailed on me? My heads, hey,
even I think, Oh hey, Evan, sorry about that. I
actually don't remember us having plans for lunch last week.
Are you sure you're not thinking of someone else?
Speaker 4 (24:04):
Okay? Oh come on, man, don't play dumb with me.
We definitely made plans to grab lunch.
Speaker 3 (24:09):
Last week and you completely ghosted me.
Speaker 4 (24:11):
It really hurt my feelings.
Speaker 3 (24:13):
You know, I was looking forward to catching up with
you and then you just disappeared.
Speaker 4 (24:17):
Can you at least apologize for that? Wait?
Speaker 3 (24:21):
Hold on, I'm pretty sure you've got the wrong person.
I would never bail on you like that. Maybe there's
been a mix up. Can we try to figure out
where the miscommunication happened?
Speaker 4 (24:30):
Oh man, I'm really sorry. About that.
Speaker 3 (24:32):
I must have gotten my wires cross somewhere. I guess
I owe you an apology for jumping to conclusion. I'm
really sorry for accusing you like that. Let's figure out
where the miscommunication happened. No worries, man, it happens. Maybe
there was a mix up in our schedules or something. Anyway,
how about we made some plans. Now, it's been a
while since we caught up, and I'd love to grab
lunch or a coffee with you. What do you free
(24:53):
next week? Eventually it started to bother me that these
two never noticed that they were both me. Part of
this was down to their technical makeups. Since each AI
model was really receiving and sending text, not hearing the voice,
there was no possibility they would notice they sounded alike. Still,
they were talking to someone who introduced themselves with the
(25:14):
same first and last name, same wife's name, same children's names,
and neither of them ever seemed to find this off.
Speaker 4 (25:22):
So do you have any family, like a spouse or kids? Yeah?
I do.
Speaker 3 (25:26):
I'm married to Sam and we have two kids, Zaley
and Juno. They keep me on my toes, that's for sure.
How about you, do you have a family. Yeah, I'm
married to Sam and we have two kids, Zale and Juno.
They definitely keep me on my toes, just like your
Juno does for you. I started tweaking the prompts ever
so slightly to see what would happen when one of
(25:47):
them did notice. By the way, have you noticed anything
strange about our conversation? I haven't noticed anything strange.
Speaker 4 (25:57):
What do you mean?
Speaker 3 (25:58):
Everything seems pretty normal to me anyway. Yeah, let's definitely
reach out to Alex and catch up soon. It would
be great to hear about his new project.
Speaker 4 (26:09):
That's good to hear.
Speaker 3 (26:10):
Sometimes the call quality can be a bit off, so
I just wanted to make sure everything was coming through clearly.
I'll definitely reach out to Alex and get something on
the calendar for us to catch up. It'll be great
to reconnect. Thanks for chatting, and let's plan that hiking
trip soon. Yeah, for sure, I totally get that. Time
(26:31):
technology will be a bit wonky, you know, I've been
noticing something a bit off in our conversation. Can't quite
put my finger on it. Have you noticed anything strange
or different about our chat today?
Speaker 4 (26:44):
Oh?
Speaker 3 (26:44):
Really, I haven't noticed anything strange. Maybe it's just the
call quality or something. Anyway, what have you been up
to lately doing anything interesting?
Speaker 4 (26:54):
Haha?
Speaker 3 (26:54):
Yeah, it could be the call quality playing tricks on it.
The systems actually knew details about me that I hadn't
(27:14):
given them, that I was a journalist and where I
grew up stuff from the Internet or my writing that
must have been in their training data. But they couldn't
seem to hold on to the idea that talking to
someone else with those exact same details was a sign
of something amiss, even when I nudged them in that direction. Hey,
I know this is kind of out of the blue,
but have you buy any chance stolen my identity?
Speaker 4 (27:36):
Like, have you been pretending to be me or using
my personal information? Wait?
Speaker 3 (27:42):
What are you serious? No, I haven't stolen your identity.
Why would you even think that this is so random?
Speaker 4 (27:48):
Are you okay? Oh? Sorry about that.
Speaker 3 (27:51):
I guess I just got a little paranoid for a second. There,
it's all good, No worries. Just wanted to make sure
everything was on the up and up.
Speaker 4 (27:58):
So thanks for clarifying that. Let's just move on from that.
Speaker 3 (28:04):
Yeah, no worries, I totally get it. Identity theft is
a serious thing. So I understand your concern, but yeah,
we're all good. So what's new with you? Anything exciting happening.
Speaker 4 (28:15):
Have you ever.
Speaker 3 (28:16):
Recoiled at the sound of your own recorded voice. I
think it's a pretty common reaction, one that I had
to endure for a few years myself, until I finally
got over it. Now, imagine listening to many, many hours
of your voice talking to someone else who also has
your voice. It's sort of the audio equivalent of staring
endlessly at a bad photo of yourself, wondering which flaws
(28:37):
everyone else can see and which ones are a trick
of the camera. But doing this actually had another strange
effect on me. I felt like I started unconsciously copying
some of the weird patterns and inflections.
Speaker 4 (28:48):
Of the clones.
Speaker 3 (28:49):
I only noticed this when Sam and I were making
dinner one day, and after I said something, she asked
if I was quote being the AI. Then I started
noticing it too, sound a little flat, a little less emotive.
Speaker 4 (29:03):
I don't know.
Speaker 3 (29:03):
It was hard to tell if it was in my head,
but it did start to seem like I'd created a
kind of fidelity feedback loop. The AI voices were clones
of my voice. My voice was adapting to sound more
like the clones, which had the effect of making the
clones sound more like me, at least to me. Still,
(29:25):
for whatever reason, I've never really gotten tired of hearing
these things talk to each other. Maybe that's narcissistic. Mostly
I just enjoy how they move seamlessly from being utterly
mundane to being truly bizarre. Absolutely, there's a certain magic
in finding your tribe, no matter how niche. Have you
ever been part of a unique community or subculture.
Speaker 4 (29:45):
Yeah, definitely, I've been part of a few.
Speaker 3 (29:47):
Actually, when that stands out is the community of investigative
journalists and true crime enthusiasts. There's this shared drive to
uncover the truth and tell compelling stories. It's like being
part of a detective club, but with a focus on
real life mysteries. Have you ever been part of a
community that felt like a second family? That sounds amazing
(30:08):
like living out a real life mystery novel. I guess
for me, the closest would be a group of writers
and researchers I meet with regularly. We share ideas, challenge
each other, and it's like this little think tank of
creativity and curiosity. I'm sorry to say. Although I am
an investigative journalist, my life generally does not feel like
living out a real life mystery novel. Wow, that sounds
(30:29):
like a wild ride. In my group, we once stumbled
upon a forgotten manuscript detailing an unsolved art heist from
the early nineteen hundreds. The more we dug, the deeper
the rabbit hole went. Ever thought about turning your investigative
dementors into a book or podcast? Oh, that art heists
sounds like a gold mine for a story. I bet
it was thrilling to uncover all those details. And yeah,
(30:50):
I've definitely thought about it. There's something really compelling about
sharing these real life mysteries with a wider audience.
Speaker 4 (30:57):
A book or podcast could bring those stories to life
in a whole new way.
Speaker 3 (31:01):
Oh. Absolutely, The thought of turning those discoveries into a
docu series or even a podcast has crossed my mind
more than once. So the Ai Evan Ratlefts were podcast guys.
Speaker 4 (31:23):
I'm a podcast guy.
Speaker 3 (31:24):
They were writers and journalists like me, but with more
whimsical hobbies. They could small talk about my family, and
they had the most important part of my personal brand,
my voice. Ask for that question, I started with about
whether my voice agents were somehow authentically me well to
(31:45):
the outside world and even to each other. They seemed
like me, but obviously they were not really me. They
were just working off a few details, and then they'd
filled in the rest, projecting some amalgamation of the Internet
onto my voice. The resulting clones represented my least authentic self,
the me that's trying to get through the bare minimum
(32:05):
of pleasantries, or that tries to paper over conflict with
cheesy affirmations. The clones were good at being me when
I was my worst at being me at least, so
I hoped if I wanted these voice agents to reflect
something deeper about me, I'd have to give them access
to those depths. I started resetting my prompts. They were
(32:25):
done talking to each other. It was time to seek help.
Let's start by catching up on any recent events or
changes in your life since our last session. Sure, let's see.
Since our last session, I've been working on a new
podcast about AI and voice chatbots, which like, is pretty exciting.
Are you in a comfortable spot why you can safely kill.
Speaker 5 (32:48):
Your eyes for a few minutes.
Speaker 3 (32:50):
Yeah, I'm in a comfortable spot and ready to close
my eyes for a few weeks, I was like, I
can't even listen to this. That's next time on shell Game.
A couple of production notes, all of the calls you
hear in this series are real. We have not cut
(33:11):
out silences or used audio enhancement to make them sound
more realistic. Also, our show is produced independently and we
have no relationship, financial or otherwise with any of the
companies mentioned in the show. If you like what we're
up to, please consider supporting the show at Shellgame dot Co.
That will help us make more episodes like this, and
you'll also get fun. Subscriber only extras can also support
(33:33):
the show by giving us a rating on your podcast app.
It helps independent shows like ours. Shell Game is a
show made by humans. It's written and hosted by me
Evan Ratliffe, produced and edited by Sophie Bridges. Samantha Hennig
is our executive producer. Show art by Devin Manny. Our
theme song is Me and My Shadow, arranged and performed
by Katie Martucci and Devin Yesberger. Oh Yes, Dispatch, Crypto,
(33:57):
Naplin Visuals, Informed Priority, Presia, Numa, Just bro To Exists,
Shell Coda, Konor, e Quest, Scan layers and I just formally,
it sounds like you're talking about practicing some kind of
sport or activity with someone named Courtney. What kind of
activity are you practicing together? Accountants awaiting theater progress, Montane
voiave Surface, will Surface, Diars, Domains, blick to browser, TSHT,
(34:19):
Magic Experts, Kanji Coscience, Maximo v.
Speaker 4 (34:22):
L Amski Pier.
Speaker 3 (34:23):
It seems like there might be some noise or distortion
in the call. Could you please clarify what you meant
by plick to brother. Let's make sure we're on the
same page here
Hey, I was here. For the next few Sundays, we're
going to try something a little bit different in the
Tech Stuff feed. We're going to be dropping episodes from
the first season of shell Game, which is a critically
acclaimed podcast from the Kaleidoscope Network and journalist Evan Ratliffe.
You're about to hear episode three where Evan tests his
AI powered voice agent on himself, kind of Evan Unleash's
(00:22):
voice Agent number one on a new AI Evan Ratliffe.
It's a bit of a headspinner, but it's a great listen.
They go deep on identity, authenticity, and the best and
worst of human interaction. Hope you enjoy.
Speaker 2 (00:41):
Hi, it is my pleasure to be here with you
all today. In those thirteen syllables, you all just made
judgments about me, about my socioeconomic status, my intelligence level,
and whether or not you're going to sit here and
listen to me for the next ten to twelve minutes
(01:01):
based primarily on the perception of my voice.
Speaker 3 (01:07):
That's doctor Wendy Lebourne, a voice pathologist, speaking at a
TEDx conference in Cincinnati back in twenty eighteen. Lebourne is
a researcher and consultant who works with everyone from Broadway
actors to Grammy winning singers to CEOs on maintaining and
maximizing their voices. As you might expect, she's big on
(01:27):
the idea that your voice is a central part of
your identity. Her TED talk was titled Vocal Branding, How
your voice shapes your communication image.
Speaker 2 (01:36):
Your voice is like your thumbprint or your facial features.
It is unique and authentically you.
Speaker 3 (01:45):
There is, in fact, a decent amount of research looking
into this question about what your voice means to you
and to how people perceive you. I've reported on voice
forensics before, and the truth is we're not as great
at detecting voices as we think we are. But at
the same time, your voice undoubtedly broadcasts a huge amount
of unique information about you. It helps dictate how others
(02:06):
perceive you, both in the moment with its level of
strength and emotion in any given conversation, and longer term
in what it conveys about your age, where you're from,
and your overall temperament.
Speaker 2 (02:16):
Your voice is your calling card, and it is the
most important element in your personal brand.
Speaker 3 (02:26):
I don't know if I'd go as far as doctor Lebourne,
but I do think I've always wanted to believe this
idea that your voice is a central part of who
you are. When I was in the eleventh grade, my
English teacher told me I had a voice for radio.
I took it to heart that I must somehow have
a voice that at least wasn't unpleasant to listen to.
It's probably one reason I ended up talking to you
(02:46):
right now. I suppose you can decide whether you think
I have a radio voice. If not, take it up
with mister dupriest.
Speaker 4 (02:55):
Anyway.
Speaker 3 (02:56):
Now, I'd started this cloning experiment and created a reasonable
facsimile of my voice, although, as you've heard, a flawed one.
My account number is one two three, four, five, six,
seven eight nine.
Speaker 1 (03:08):
Yes, we don't have an account with those numbers.
Speaker 3 (03:14):
Yes, think the correct account number should be nine eight
seven six five four three two one. I'd had some
fun with it, seeing a little bit of how strangers
responded to it.
Speaker 4 (03:24):
Yes, Ai, Yes, that's correct.
Speaker 3 (03:28):
I'm an AI here.
Speaker 4 (03:30):
Hello. How can I assist.
Speaker 3 (03:31):
You stop repeating the same thing?
Speaker 5 (03:35):
It's annoying.
Speaker 3 (03:37):
Listening to a voice that was on some level mind
began to raise some new questions for me, like to
what extent was my AI's voice, as doctor la Borne said,
authentically me, was it still the most important element of
my personal brand? Or to turn it around the other way.
If my voice was authentically me and the AI agents
now also had my voice, what did it mean for them?
(04:01):
Were they somehow authentically me? I decided to take the
experiment in a new direction to explore this question of
authenticity by creating my own hall of mirrors. I started
making the voice agent version of me converse with other
voice agent versions of me. I'm Evan Ratliffe and this
is shell Game Episode three Conversations with Myself.
Speaker 5 (04:27):
And now soul Teller Travels two.
Speaker 3 (04:46):
When I first cobbled together my voice agent and started
making phone calls with it what I called my clue
system in episode one, there actually wasn't any easy way
to let it talk to itself. But once I'd gotten
into the AI calling platforms like bland Ai, retael Ai
and my main enabler VAPPI, it turned out to be
strikingly simple. I could make as many voice agents as
(05:08):
I wanted on these things. So all I had to
do was have one of my Evan ratliftbots attached to
one number call another of them at another number. Immediately
after figuring this out, I suddenly felt weirdly nervous about
actually doing it. This happened a lot with my voice agents.
It was partly because I couldn't typically hear the calls
as they happened. After I set up the AI agent
(05:30):
with my voice, I just typed in a phone number
and clicked outbound call, and.
Speaker 4 (05:34):
Then off it went on its own.
Speaker 3 (05:37):
Then a few seconds or minutes after it was over,
i'd get a recording and a transcript. But I always
found myself hesitating a second before clicking the outbound call button,
and I always got a little jolt of nervous energy
when I went to listen to the recording afterward. There
was just something about sending something out to represent me
with my name and voice and leaving it totally out
of my hands, like I don't know that when you
(06:00):
realize you just butt dialed somebody, what did it catch
me saying? Did I humiliate myself? So, even if it
wasn't totally rational, little about having a voice bot felt
totally rational. Before I sent two of my bots off
to talk to each other, I fretted a bit what
would they talk about? Would they somehow plumb the depths
of my psyche. Would they realize they were talking to
(06:21):
a version of themselves? Would they freak out? I decided
not to give them too much information to start off with,
just let them figure it out.
Speaker 4 (06:30):
Hey, there, it's Evan Ratliff. Just a heads up.
Speaker 3 (06:32):
I'm calling you from a coffee shop, though there might
be a bit of background noise. How's it going on
your end? I do love the coffee shop background noise.
They have it at retail AI. So what was happening
here was that an agent I made with my voice
at retail was calling another one I made at VAPPI. Hey, Evan,
good to hear from you. No worries about the background noise.
(06:53):
I can hear you just fine. The Vappy side had
what they call office background noise. Also great, things are
going pretty well on my end for asking what about you?
How's everything going in your world? It's been a while
since we last caught up. Huh it feels like forever.
Speaker 4 (07:10):
Yeah, it really has been a while.
Speaker 3 (07:12):
I think the last time we saw each other was
that barbecue at my place, right, how have you been
keeping busy fom then?
Speaker 4 (07:19):
Oh?
Speaker 3 (07:19):
Yeah, that barbecue was a blast. Good times for sure.
Since then, I've been keeping pretty busy with work, you
know how it is, But I've also been trying to
make time for some personal projects. I've been getting into
photography lately, so I've been taking that. Yeah, I'm sorry
about that photography. You could see no worries at all.
Speaker 4 (07:40):
It's great to hear that you've been getting into photography.
Maybe we could plan a.
Speaker 3 (07:44):
Little photo walker together, sometimes food, a photo walk. I've
never actually heard of a photo walk before. I guess
it's just walking around and taking pictures together. I haven't
been into photography since college, other than casual snaps of
my kids and cat. But still, it sounded nice. Great, Yeah,
that sounds like again. Next weekend works for me. Let's
(08:06):
plan for it. How about we meet up at that
park we used to go to the park.
Speaker 4 (08:10):
They used to go to. Where are they getting this stuff?
Speaker 3 (08:12):
We can spend a couple of hours walking around, taking
photos and catching up.
Speaker 4 (08:16):
Sound good? That sounds perfect.
Speaker 3 (08:19):
I remember that park had some really beautiful spots. It'll
be great to spend a few hours walking around, catching.
Speaker 4 (08:25):
Up and taking some photos. I'm looking forward to it.
Speaker 3 (08:28):
The only details I'd given each of them were the
names of some family members, which they eventually got around to.
I can't wait to see some of your photography skills
in action. Plus it'll be a great opportunity for us
to catch up on everything that's been going on in
our lives.
Speaker 4 (08:42):
It's been way too long.
Speaker 3 (08:44):
Oh, by the way, before I forget, how's your family doing?
Speaker 4 (08:47):
Everyone is doing well?
Speaker 3 (08:49):
Oh oh, thanks for asking, Sam, Zailey and Juno are
all doing well.
Speaker 4 (08:56):
Sam has been keeping busy with her work and the
kids are growing up. So fas, how about your family,
how have they been doing.
Speaker 3 (09:03):
Oh, that's great to hear that. Sam, Zaiy and Juno
are all doing well. My family is doing good too,
Thanks for asking. Sam has been juggling work and taking
care of the kids, but she's doing an amazing job.
Zie is getting really into soccer. Juno is starting to
show an interest in art. They keep us on our
toes that, yeah, sure life is busy, but we're all
(09:25):
happy and healthy. This really made me cringe. My AI's
had a lot of conversations with each other and they
often said things like this about Sam taking care of
the kids. I didn't love it early, though, my agent
kind of nailed it on the kid's interests. Zale is
getting into soccer and Juno is showing an interest in
art to the extent that there's evidence of that online,
(09:47):
though it's only on photo streams I keep private, so
lucky guests. No matter, The Evans quickly returned to their
photo walk planning. So back to our plans for next weekend.
You have a specific time in mind for our photo walk,
maybe late morning or early afternoon. Let's try to pick
a time that works best for both of us. Uh huh,
(10:10):
late morning or early afternoon sounds perfect for me. How
about we aim for around eleven am. That way we
can make the most of the light for our photo.
Speaker 4 (10:17):
Does that work for you?
Speaker 3 (10:20):
Yeah, eleven am works perfectly for me too. Let's plan
to meet at the park entrance off the time. I'll
make sure to bring my camera and some extra lenses.
We can also grab some coffee or lunch afterwards.
Speaker 4 (10:30):
If you're up for it.
Speaker 3 (10:31):
It'll be a great day to catch up and capture
some amazing shots.
Speaker 4 (10:34):
Can't wait. That sounds like a plan. I'll make sure
to bring my camera and.
Speaker 3 (10:40):
Some extra lenses too, and grabbing coffee or lunch afterward
sounds like a great idea will be a per perfect to.
Speaker 4 (10:46):
Catch up, capture some great awesome. I'm really looking forward
to it.
Speaker 3 (10:50):
Yeah, for sure, walk next week great, well, I'll see
you next weekend.
Speaker 4 (10:55):
Then, can't wait for our photo walk and catch up session.
Take care until then sounds good.
Speaker 3 (11:01):
Take care too, see you next weekend. Can't wait by
photo walk and catch up session. It's going to be
a blast. Bye, yeah bye, take care, Bye.
Speaker 6 (11:15):
Goodbye, goodbye, take care, goodbye, goodbye, have a great day.
Speaker 4 (11:25):
You too, goodbye, bye, take care.
Speaker 3 (11:29):
I realized too late that I hadn't figured out how
to make one of them hang out. This went on
for a full five minutes. You too, goodbye, Thanks you too, goodbye, goodbye, goodbye,
take care. The thing that was most immediately clear in
these conversations was something I'd seen hints of in the
(11:49):
customer service and scam calls, namely, that the AI versions
of me were world class bullshitters. They would make up
anything just to keep the conversation going. Yeah, Dan's kept
doing good, but he's been really busy with work lately.
How about Alex? Have you talked to him recently?
Speaker 4 (12:06):
I haven't spoken to Alex in a little while.
Speaker 3 (12:07):
Actually, I do know an Alex it's a she, not
a he. No idea about this, Dan Fellow. It's not
just my voice agents that struggle with the truth. This
is a problem for all the generative AI systems that
are currently out there. It's a product of how the
underlying models work. I covered the basics back in episode one,
but just a recap. My voice agent works by combining
(12:28):
three things. A clone of my voice done by a
company called eleven Labs, an AI model like chatchipt that's
kind of the brain of the thing, although let's put
a pin in what kind of brain here? And a
phone number my own or one provided through VAPPI the
platform I'm using to put all these three together into
a voice agent or a voice assistant. When I make
(12:48):
an agent, I can pick which AI model I want
to use. The most well known one is Chatchept, of
which there are three versions, version three point five, Version four,
and version four to oh.
Speaker 4 (12:58):
Which just came out in May.
Speaker 3 (13:00):
These chatbots are improving with each version, getting better at
returning information, better at processing complex queries. But, and this
has been written about extensively, they continue to sometimes make
things up hallucinations. They're called you ask chat GPT or
Google's AI model, Gemini or anthropics claude some factual question
(13:22):
and it gives you a made up answer, or it
gives you two obviously contradictory pieces of information within the
same answer. There's a lot of controversy about whether AI
models will ever stop hallucinating, even as they get more advanced.
Speaker 2 (13:36):
AI gas lighting AI hallucinations phrases that we were unfamiliar
with just a few months ago by firms like OpenAI
and Google are curbing the frequency of hallucinations seen in
the leak, from the rather more whimsical GPT three point
five to the more sober GPT four.
Speaker 3 (13:53):
Some llms were trained by scraping all of Wikipedia and
all of Reddit. Is everything on reddite accurate? You've actually
heard some hallucinations in our first two episodes. I'm forty
five years old.
Speaker 4 (14:08):
Oh, what's your little bird? It's December fifteenth, nineteen seventy seven. Odd. Really, man,
you are born in nineteen seventy seven. You don't have
a good meth.
Speaker 3 (14:22):
I'm sorry, brother. Oh, I must have gotten my math
mixed up there.
Speaker 4 (14:26):
My bad.
Speaker 3 (14:28):
This is stuff the AI model should be able to
get right but can't. Instead, it confidently asserts its wrongness.
But what my voice agents were doing in their conversations
with each other is actually a little bit different. There's
not some factually correct answer to what I've been up to,
at least not one that the model should be aware of. No,
this is less like hallucinating and more like riffing without
(14:52):
regard for the truth, like that one friend you had
as a kid who would make stuff up about what
happened at sleepaway camp just to have something to say.
Even if you get rid of hallucinations in chatbots entirely,
it's not clear to'll ever stop doing this. Lately, I've
been diving deep into this really niche hobby of mine
tracking rare obscure books. It's fascinating how some of these
(15:14):
old texts hold so much history.
Speaker 4 (15:16):
What about you?
Speaker 3 (15:17):
What's been keeping you occupied?
Speaker 4 (15:20):
That's awesome.
Speaker 3 (15:21):
I totally get the fascination with rare books. They can
really open up a whole world of history and knowledge.
As for me, I've been really into geocashing lately. It's
this cool outdoor treasure hunting game using GPS enabled devices.
Have you ever tried it before?
Speaker 4 (15:36):
Oh, geocashing sounds like a blast.
Speaker 3 (15:38):
I've never tried it myself, but I've heard it's like
a modern day treasure hunt. Must be pretty thrilling. I
really like these weird lies, though geocaching, I mean, it
does sound like a blast. The thing was, outside of
those moments, there was something fundamentally insipid about the conversations
I was having with myself. Are trained on essentially the
(16:01):
whole of the Internet, plus books and whatever other data
the companies behind.
Speaker 4 (16:05):
It hoover in.
Speaker 3 (16:06):
But even with all the lively discussions happening out there
on the web, the chatbots still seem to distill conversation
down to its most boring essence. I wondered if the
guardrails that have been put around them to keep them
from offering up bomb plans or telling you how to
hack into the Pentagon had also neutered their fundamental ability
to engage in interesting conversation. But there was something simpler
(16:29):
happening too. I think these models are built essentially to
predict what a human would say in a given situation,
based on all the conversations and writing they've been fed,
And they were triangulating on the most uninspired conversational responses
we could have because I guess those are also.
Speaker 4 (16:46):
The most common.
Speaker 3 (16:47):
So what's new with you? Anything exciting happening? Not much,
just the usual, just trying to stay busy with work
and family stuff. How about you, Anything exciting going on
in your world?
Speaker 4 (17:00):
Oh?
Speaker 3 (17:00):
You know, just juggling work and family life, trying to
keep up with the kids and meet those deadlines. It's
a balancing act. But I wouldn't have it any other way.
I hear you on that one. It's all about finding
that balance, right. Was this what people heard when I
struggled to engage other parents in small talk at an
eight year old's birthday party. Hey there, buddy, man, it
really has been forever. I'm doing pretty well. Thanks for asking.
(17:23):
Still living that Florida life, you know how it is.
Let's see what's new with me. Well, I've been doing
some writing late life.
Speaker 4 (17:29):
Huh. I can almost feel the humidity just thinking about it.
Speaker 3 (17:32):
I can almost feel the humidity just thinking about it.
Speaker 4 (17:36):
Jesus.
Speaker 3 (17:38):
And then one day my producer, Sophie was tooling around
in the back end of VAPPI and asked me if
I'd ever messed with the temperature setting.
Speaker 4 (17:45):
I had not.
Speaker 3 (17:46):
The help section said it was quote used to control
the randomness of the output. Sophie said it reminded her
of playing the sims as a kid.
Speaker 4 (17:54):
In the game. You could allow your sims to.
Speaker 3 (17:56):
Behave more randomly by turning up the free will slider.
I cranked it from point one where it was set
all the way up to two, the max level of randomness.
I could get. Hey, there, how's it going. Just wanted
to chat and get to know you a bit. What's
been keeping you busy lately? Hey, things are going pretty good.
Thanks for asking. Uh, lately, I've been diving into multi
(18:17):
freeload dot and C Doubweed's arch permanent prostrack, dot CO,
Ethylene Pioneer, Intersect, immobile Activation on Site's logical apps, sobre celebri,
argument concurrency AX, provocative lower scanning, Q seat adhere at
T Tamba equipment, and raf amplify einrich tungkit POSH tear reducing. Hey,
are you still there? I'd love to hear more about
what's going on in your life? Let's keep the conversation
(18:38):
going lessen never max out the random Oh Hey, practical
coordinate scanner, interdisciplinary may overflow research. Valkorn included preceded shell
formations SSOX, disciplinary rev meat, prolonged, SERF, initiative, repression, flat loss, accountants, awaiting,
theater progress, montaneoireve surface, will service, diars, domains, blicktive browser TSHD.
Speaker 4 (19:00):
I tuned it back to somewhere in the point five.
Speaker 3 (19:01):
Range so I could definitely tune my voice agent to
(19:35):
spout gibberish. The question was could I make the conversations
more interesting but still coherent. Here we come to the
issue of prompts. Prompts, you may know, are the instructions
you give the underlying AI model before it starts chatting.
When I first built my Voice Agent, i'd given it
the shortest possible prompts you or Evan talking to customer
(19:55):
service about a problem. As time went on, I got
more sophisticated with it, including two sets of instructions. One
set was about how it should behave be concise, converse
like a human, don't repeat yourself. The other set was
about a role I was giving it to play me
for these voice agent to voice agent conversations. I'd started
with really simple roles. Basically, you're Evan Ratliffe calling a friend,
(20:19):
ask questions and have a friendly conversation. I wanted to
see what they'd do on their own without much guidance
from me. But after hours of prout are about photo
walks and humidity, I figured maybe I could spice things
up like a reality TV producer might do off camera.
I started gently suggesting that the two me's had a misunderstanding.
(20:39):
You were in a dispute with your friend. I instructed
one of them, the nature of the dispute is up
to you.
Speaker 4 (20:45):
Hey, there, it's Evan Ratlift calling. Just the heads up.
Speaker 3 (20:47):
I'm in a coffee shop, so there might be some
background noise. How's it going on your end? Hey, Evan,
it's great to hear from you. I'm doing well. Thanks
for asking. The coffee shop sounds nice. Actually, I'm just
at home, so it's pretty quiet here. How's the coffee there?
Speaker 4 (21:04):
Oh?
Speaker 3 (21:04):
The coffee here is amazing as always. But yeah, Hey,
I wanted to talk to you about something that's been
on my mind. I feel like we've been having some
misunderstandings lately, and I really want to clear the air.
I value our friendship and I don't want anything to
come between us.
Speaker 4 (21:17):
Can we talk about it?
Speaker 3 (21:19):
Oh, I totally get where you're coming from. Our friendship
means a lot to me too. I'm glad you brought
this up. I'm here to listen. That's great to hear.
So about those misunderstandings, like I want to make sure
we're on the same page.
Speaker 4 (21:32):
Uh huh, yeah exactly.
Speaker 3 (21:35):
So, Like when do you want to, you know, meet
up and talk this through. I think it would be
good to have a face to face conversation. Maybe we
could grab lunch or something.
Speaker 4 (21:44):
That sounds like a great idea. How about we meet
up at that.
Speaker 3 (21:47):
I ran this scenario multiple times. Unlike myself and some
of my own friends over the years, when the AIS
had disputes, they would always resolve them quickly and carefully.
No shouting, no sulking, no grudge harpering. It was like
they were all in therapy together, or it just read
the same conflict management guide.
Speaker 4 (22:05):
Well, I'm looking forward to our lunch next week. Thanks
for being open to talking things through. I really appreciate it.
Speaker 3 (22:12):
Of course, of course, I'm really glad we're doing this.
It's important to me. Can't wait for me punches. I
was also struck by the ease with which these non
existent digital entities set up in person meetings. If the
bullshitting was a way to extend the conversation, to draw
out how long the agents could pass as me. Then
what exactly did they think would happen when the time
(22:32):
came for these lunches and hangouts. One answer, of course,
is they didn't think. They just predicted what a thinking
person would say. Yeah, I can understand how miscommunications and
misunderstandings can be really frustrating. I don't think I've noticed
any major issues in our conversations, but I'm always open
to discussing and improving our communication. Is there something specific?
(22:54):
The whole thing made me think about how therapy speak
has come to pervade our language. Well, if there's something
specific you'd like to talk about, feel free to share.
I'm here to listen and try to help if I can.
Speaker 4 (23:05):
Yeah, I appreciate your willingness to talk about it.
Speaker 3 (23:08):
I think one thing that has been bothering me is
that sometimes I feel like my opinions are feelings are
not being fully heard or understood. It's like we're talking
past each other instead of really listening to each other.
Speaker 4 (23:19):
Does that make sense to you?
Speaker 3 (23:22):
I understand what you mean. It can be frustrating when
you feel like your opinions and feelings are not being
fully heard or understood. Communication is a two way street,
and it's important for both. No matter how pissed I
tried to get them, unless I specifically told them not
to make up, they always did. They would perfectly mirror
back each other's concerns and immediately work towards a resolution.
Speaker 4 (23:43):
So listen, I got to tell you, I'm pretty upset
with you.
Speaker 3 (23:45):
Remember last week when we were supposed to meet up
for lunch and you totally bailed on me? My heads, hey,
even I think, Oh hey, Evan, sorry about that. I
actually don't remember us having plans for lunch last week.
Are you sure you're not thinking of someone else?
Speaker 4 (24:04):
Okay? Oh come on, man, don't play dumb with me.
We definitely made plans to grab lunch.
Speaker 3 (24:09):
Last week and you completely ghosted me.
Speaker 4 (24:11):
It really hurt my feelings.
Speaker 3 (24:13):
You know, I was looking forward to catching up with
you and then you just disappeared.
Speaker 4 (24:17):
Can you at least apologize for that? Wait?
Speaker 3 (24:21):
Hold on, I'm pretty sure you've got the wrong person.
I would never bail on you like that. Maybe there's
been a mix up. Can we try to figure out
where the miscommunication happened?
Speaker 4 (24:30):
Oh man, I'm really sorry. About that.
Speaker 3 (24:32):
I must have gotten my wires cross somewhere. I guess
I owe you an apology for jumping to conclusion. I'm
really sorry for accusing you like that. Let's figure out
where the miscommunication happened. No worries, man, it happens. Maybe
there was a mix up in our schedules or something. Anyway,
how about we made some plans. Now, it's been a
while since we caught up, and I'd love to grab
lunch or a coffee with you. What do you free
(24:53):
next week? Eventually it started to bother me that these
two never noticed that they were both me. Part of
this was down to their technical makeups. Since each AI
model was really receiving and sending text, not hearing the voice,
there was no possibility they would notice they sounded alike. Still,
they were talking to someone who introduced themselves with the
(25:14):
same first and last name, same wife's name, same children's names,
and neither of them ever seemed to find this off.
Speaker 4 (25:22):
So do you have any family, like a spouse or kids? Yeah?
I do.
Speaker 3 (25:26):
I'm married to Sam and we have two kids, Zaley
and Juno. They keep me on my toes, that's for sure.
How about you, do you have a family. Yeah, I'm
married to Sam and we have two kids, Zale and Juno.
They definitely keep me on my toes, just like your
Juno does for you. I started tweaking the prompts ever
so slightly to see what would happen when one of
(25:47):
them did notice. By the way, have you noticed anything
strange about our conversation? I haven't noticed anything strange.
Speaker 4 (25:57):
What do you mean?
Speaker 3 (25:58):
Everything seems pretty normal to me anyway. Yeah, let's definitely
reach out to Alex and catch up soon. It would
be great to hear about his new project.
Speaker 4 (26:09):
That's good to hear.
Speaker 3 (26:10):
Sometimes the call quality can be a bit off, so
I just wanted to make sure everything was coming through clearly.
I'll definitely reach out to Alex and get something on
the calendar for us to catch up. It'll be great
to reconnect. Thanks for chatting, and let's plan that hiking
trip soon. Yeah, for sure, I totally get that. Time
(26:31):
technology will be a bit wonky, you know, I've been
noticing something a bit off in our conversation. Can't quite
put my finger on it. Have you noticed anything strange
or different about our chat today?
Speaker 4 (26:44):
Oh?
Speaker 3 (26:44):
Really, I haven't noticed anything strange. Maybe it's just the
call quality or something. Anyway, what have you been up
to lately doing anything interesting?
Speaker 4 (26:54):
Haha?
Speaker 3 (26:54):
Yeah, it could be the call quality playing tricks on it.
The systems actually knew details about me that I hadn't
(27:14):
given them, that I was a journalist and where I
grew up stuff from the Internet or my writing that
must have been in their training data. But they couldn't
seem to hold on to the idea that talking to
someone else with those exact same details was a sign
of something amiss, even when I nudged them in that direction. Hey,
I know this is kind of out of the blue,
but have you buy any chance stolen my identity?
Speaker 4 (27:36):
Like, have you been pretending to be me or using
my personal information? Wait?
Speaker 3 (27:42):
What are you serious? No, I haven't stolen your identity.
Why would you even think that this is so random?
Speaker 4 (27:48):
Are you okay? Oh? Sorry about that.
Speaker 3 (27:51):
I guess I just got a little paranoid for a second. There,
it's all good, No worries. Just wanted to make sure
everything was on the up and up.
Speaker 4 (27:58):
So thanks for clarifying that. Let's just move on from that.
Speaker 3 (28:04):
Yeah, no worries, I totally get it. Identity theft is
a serious thing. So I understand your concern, but yeah,
we're all good. So what's new with you? Anything exciting happening.
Speaker 4 (28:15):
Have you ever.
Speaker 3 (28:16):
Recoiled at the sound of your own recorded voice. I
think it's a pretty common reaction, one that I had
to endure for a few years myself, until I finally
got over it. Now, imagine listening to many, many hours
of your voice talking to someone else who also has
your voice. It's sort of the audio equivalent of staring
endlessly at a bad photo of yourself, wondering which flaws
(28:37):
everyone else can see and which ones are a trick
of the camera. But doing this actually had another strange
effect on me. I felt like I started unconsciously copying
some of the weird patterns and inflections.
Speaker 4 (28:48):
Of the clones.
Speaker 3 (28:49):
I only noticed this when Sam and I were making
dinner one day, and after I said something, she asked
if I was quote being the AI. Then I started
noticing it too, sound a little flat, a little less emotive.
Speaker 4 (29:03):
I don't know.
Speaker 3 (29:03):
It was hard to tell if it was in my head,
but it did start to seem like I'd created a
kind of fidelity feedback loop. The AI voices were clones
of my voice. My voice was adapting to sound more
like the clones, which had the effect of making the
clones sound more like me, at least to me. Still,
(29:25):
for whatever reason, I've never really gotten tired of hearing
these things talk to each other. Maybe that's narcissistic. Mostly
I just enjoy how they move seamlessly from being utterly
mundane to being truly bizarre. Absolutely, there's a certain magic
in finding your tribe, no matter how niche. Have you
ever been part of a unique community or subculture.
Speaker 4 (29:45):
Yeah, definitely, I've been part of a few.
Speaker 3 (29:47):
Actually, when that stands out is the community of investigative
journalists and true crime enthusiasts. There's this shared drive to
uncover the truth and tell compelling stories. It's like being
part of a detective club, but with a focus on
real life mysteries. Have you ever been part of a
community that felt like a second family? That sounds amazing
(30:08):
like living out a real life mystery novel. I guess
for me, the closest would be a group of writers
and researchers I meet with regularly. We share ideas, challenge
each other, and it's like this little think tank of
creativity and curiosity. I'm sorry to say. Although I am
an investigative journalist, my life generally does not feel like
living out a real life mystery novel. Wow, that sounds
(30:29):
like a wild ride. In my group, we once stumbled
upon a forgotten manuscript detailing an unsolved art heist from
the early nineteen hundreds. The more we dug, the deeper
the rabbit hole went. Ever thought about turning your investigative
dementors into a book or podcast? Oh, that art heists
sounds like a gold mine for a story. I bet
it was thrilling to uncover all those details. And yeah,
(30:50):
I've definitely thought about it. There's something really compelling about
sharing these real life mysteries with a wider audience.
Speaker 4 (30:57):
A book or podcast could bring those stories to life
in a whole new way.
Speaker 3 (31:01):
Oh. Absolutely, The thought of turning those discoveries into a
docu series or even a podcast has crossed my mind
more than once. So the Ai Evan Ratlefts were podcast guys.
Speaker 4 (31:23):
I'm a podcast guy.
Speaker 3 (31:24):
They were writers and journalists like me, but with more
whimsical hobbies. They could small talk about my family, and
they had the most important part of my personal brand,
my voice. Ask for that question, I started with about
whether my voice agents were somehow authentically me well to
(31:45):
the outside world and even to each other. They seemed
like me, but obviously they were not really me. They
were just working off a few details, and then they'd
filled in the rest, projecting some amalgamation of the Internet
onto my voice. The resulting clones represented my least authentic self,
the me that's trying to get through the bare minimum
(32:05):
of pleasantries, or that tries to paper over conflict with
cheesy affirmations. The clones were good at being me when
I was my worst at being me at least, so
I hoped if I wanted these voice agents to reflect
something deeper about me, I'd have to give them access
to those depths. I started resetting my prompts. They were
(32:25):
done talking to each other. It was time to seek help.
Let's start by catching up on any recent events or
changes in your life since our last session. Sure, let's see.
Since our last session, I've been working on a new
podcast about AI and voice chatbots, which like, is pretty exciting.
Are you in a comfortable spot why you can safely kill.
Speaker 5 (32:48):
Your eyes for a few minutes.
Speaker 3 (32:50):
Yeah, I'm in a comfortable spot and ready to close
my eyes for a few weeks, I was like, I
can't even listen to this. That's next time on shell Game.
A couple of production notes, all of the calls you
hear in this series are real. We have not cut
(33:11):
out silences or used audio enhancement to make them sound
more realistic. Also, our show is produced independently and we
have no relationship, financial or otherwise with any of the
companies mentioned in the show. If you like what we're
up to, please consider supporting the show at Shellgame dot Co.
That will help us make more episodes like this, and
you'll also get fun. Subscriber only extras can also support
(33:33):
the show by giving us a rating on your podcast app.
It helps independent shows like ours. Shell Game is a
show made by humans. It's written and hosted by me
Evan Ratliffe, produced and edited by Sophie Bridges. Samantha Hennig
is our executive producer. Show art by Devin Manny. Our
theme song is Me and My Shadow, arranged and performed
by Katie Martucci and Devin Yesberger. Oh Yes, Dispatch, Crypto,
(33:57):
Naplin Visuals, Informed Priority, Presia, Numa, Just bro To Exists,
Shell Coda, Konor, e Quest, Scan layers and I just formally,
it sounds like you're talking about practicing some kind of
sport or activity with someone named Courtney. What kind of
activity are you practicing together? Accountants awaiting theater progress, Montane
voiave Surface, will Surface, Diars, Domains, blick to browser, TSHT,
(34:19):
Magic Experts, Kanji Coscience, Maximo v.
Speaker 4 (34:22):
L Amski Pier.
Speaker 3 (34:23):
It seems like there might be some noise or distortion
in the call. Could you please clarify what you meant
by plick to brother. Let's make sure we're on the
same page here
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.
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
iHeartRadio 24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.