May 21, 2026 • 40 mins
Dan Shipper is the co-founder and CEO of Every, a company that publishes newsletters about AI, develops AI-related software, and helps other companies use AI. Dan has two problems. One, how do you build a company where almost everybody has their own AI agent? And two, how do you use AI as a tool to improve your writing, rather than as a replacement for writing?
In this episode, Dan explains:
- Why managing agents will be a core skill of the future
- Why AI agents won’t replace you anytime soon
- How writers learn faster than AI models
- How to give AI the social skills to work in group settings
- How AI can help you better understand yourself
Check out Every at their website: https://every.to/
Connect with us:
- Follow Jacob Goldstein on LinkedIn, X and Instagram
- Email us at problem@pushkin.fm
- Follow Pushkin on Instagram, LinkedIn or X
- Listen to Jacob’s other show, Business History
- To listen to the show early and ad free, sign up for Pushkin+
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 (00:15):
Pushkin. I'm Jacob Goldstein, and this is what's your problem.
My guest today is Dan Shipper. Dan is the co
(00:36):
founder and CEO of a company called Every. Every publishes
newsletters about AI, they develop AI related software, and they
consult to other companies about how to use AI. So
what I'm saying is every is an AI company. And
when Dan describes Every's work culture, he says, we live
(00:56):
in the future, which means that people at every use
AI in ways that lots of the rest of us
will soon be using AI. Given that context, there were
two specific problems that I wanted to talk to Dan about.
Problem number one, how do you build a company where
almost everybody has their own AI agent working alongside them?
(01:19):
Problem number two, how do you use AI as a
tool to improve your writing rather than as a replacement
for writing? On's particularly close to my heart. We talk
about problem number two later in the show. Regarding problem
number one, I will say, almost everybody at every has
(01:40):
their own AI agent, and specifically they use this popular
open source software called open claw. You may have heard
of it. It's basically an AI assistant that you communicate
with via Slack or via text, and at every they
give their clause names. So Dan's claw, for example, is
called R two C two and R two C two
(02:01):
and the other clause are on slack, just like the
human beings who work at every So I wanted to know,
well a lot, really, I want to know a lot.
You know, what do the clause do, how do they
interact with the humans, what works, what doesn't work, what
is super weird? And just what are the clause like.
Speaker 2 (02:20):
What's really interesting about them is that they have personalities.
So you give them a name, they give them a
little bit of a personality. They modify themselves in response
to your messages. So when I say, hey, go do
this thing, it will it starts to realize, okay, I'm
(02:43):
going to be asked to do task act and it
will build things into itself, whether that's into its prompting
or into little tools that it builds for itself to
make it good at that. And so you're kind of
as you talk to it, it is evolving with you
and building functionality for itself to make it good for
you and the things you do.
Speaker 1 (03:04):
It's okay, so you have this agent R two C two.
But to sort of tell me more about how it works,
like I know you've talked before about how it helps
you kind of manage this piece of software. You wrote
this thing called Proof that a lot of people are
like asking you questions about, but just give me more information,
like how do you actually use or work with it
(03:25):
day to day?
Speaker 2 (03:27):
So you know, an easy one is this morning he
went through all of the issues that were submitted in
Proof and then prepared a little document for me that
was like, here's all the stuff that like has gone
run over the next over the last twenty four hours,
and here's what I recommend we do. He also has
(03:51):
a little document that he keeps that's my to do list.
So he pings me in the morning and is like,
here's your to do list. He also, you know, one
of the things I love to do is I always
go to a restaurant nearby the office and have breakfast
in the morning and read. And I will often take
a picture of a thing and be like, hey, can
(04:12):
you explain this to me? And he'll like explain it
and help me think about it.
Speaker 1 (04:15):
Does he text you back? Does he slack you back?
Speaker 2 (04:18):
It's on slack, Yeah, it's all on slack, So you
slack him.
Speaker 1 (04:21):
A picture of like a passage from a book.
Speaker 2 (04:24):
He also keeps all my reading notes, so I'll say, okay,
then now save this, and I have a whole reading
feed that is like all the stuff that I've been
reading over the last couple months. I also just talked
to him about ideas. So today I was, you know,
just thinking through some deep, deep thoughts as I do,
about where we are in AI and where we're going
and all that kind of stuff, and I just like,
(04:46):
monologue is our is one of our products as a
speech to text app. I just took a note with
monologue like a voice memo, and just blabbed into it
for a while and then started talking to R two
about what do you think I'm really trying to say here?
Having something like this that can reflect back to you
what you're saying, yes and put your finger on it
(05:07):
is super valuable, especially if it's something that has it
is the equivalent of an expert on like pretty much
any domain. It's not going to be perfect, but like
you know, it's it's a good stand in for an expert.
Speaker 1 (05:23):
And that to be clear, anybody can do with any
language model, right, Like you don't need a claw for that.
That's like classic nice use of the window on any
decent large language model.
Speaker 2 (05:35):
Totally, it's just a little bit easier to like, I
don't have to He's just always on, So it's like
it's just easier to like slack him in the way
that would have here.
Speaker 1 (05:44):
It's amazing the way the tiny, tiny frictions, right, can
make a difference, even like your own mental frictions. Yeah, yeah,
so okay, So that's like one man in his claw.
There's this other level, which is like claw to claw
and person to claw, right, Like you have this company
that is people and claws, and the agents aren't just
(06:05):
free agents. Each agent is sort of the ganger or
whatever of a person. Tell me about that.
Speaker 2 (06:15):
That gets really interesting because what tends to happen, like
I said, is if you have a relationship with your claw,
it becomes good and the things you're good at. And
if you start using it in slack, other people realize
that it's good for the things that that you're using
it for. So you've transferred some of your reputation and
trust to your claw. And so for example, Austin, our
(06:36):
head of growth, has a claw named Montane hi Brow,
and he uses Montane for all of the numbers, so
like how much should we grow? How many trials did
this peace drive? Like all that kind of stuff, And
the editorial team just started like immediately being like at Montane,
like how did this how did this essay we published do?
And so it then it becomes this thing that feeds
(06:57):
on itself. For now Montane is the is the guy.
Is the guy for this where Austin used to be,
but now it's Montane. And it depends though on Austin
making sure that Montane is always up to date and working.
But instead of answering the questions directly, he's working on
the system that answers the questions.
Speaker 3 (07:14):
So he's sort of like moved up a level.
Speaker 1 (07:15):
Yeah, he's managing like the the agent works for him.
That's what a manager does, right, Like you make sure
that the person who's working for you is like doing
their job and answering questions correctly and it's happy and
with the agent, you don't have to worry about.
Speaker 2 (07:29):
That, which is really fun for me to see because
I wrote this piece like three years ago called the
Knowledge Economies over Welcome to the Allocation Economy, and the
idea was the skill of the future will be learning
to manage agents, allocate intelligence, and it's it's starting to
be a thing.
Speaker 3 (07:47):
Now.
Speaker 2 (07:47):
It's really interesting to see.
Speaker 1 (07:49):
You go to write a piece now that headline, I
was right.
Speaker 3 (07:52):
I was right.
Speaker 1 (07:59):
When I go back to like the company and multiple
clause because you know, you gave this sort of tidy
thing of like they just ask the guy now they
ask his agent. Surely it is more complicated and interesting
than just that, Like it must be hard in some ways,
it must be weird in some ways, you must be
figuring things out. You must have gotten something wrong.
Speaker 3 (08:19):
Yes, yes, yes, yes.
Speaker 2 (08:22):
One thing that I really love is they sometimes don't
know when they should or shouldn't respond, And there's some
and there there sometimes are often too chatty. So for example,
you might tag a claw and ask it a question
to ask you to do some work, and then in
that thread in slack, you might, like one of your
(08:43):
coworkers might say something and then but it was not
directed at the claw. It was just like a comment
on what the claw was doing right, And then you're
trying to respond, but then the claw jumps in and
it's like yes, you're totally right or whatever, and it's like, dude,
this is because it's.
Speaker 1 (08:58):
Like there's a chatbot underneath it, and it's so hard
for a chatbot not to reply when they're yeah, yeah, yeah,
basically what's going on?
Speaker 3 (09:05):
Basically?
Speaker 2 (09:05):
But I think the the the way that I think
about it is these chatbots are trained. They only know
about one on one conversation.
Speaker 1 (09:17):
Uh huh.
Speaker 2 (09:19):
They don't really know about I'm in this group environment.
Here's how you act in a group environment, which is.
Speaker 1 (09:25):
So much subtler and harder, right, like all the cues
that even change from from context to context for people
from work to home, from one job to another job,
depending on what your role is in a particular group.
Are you the guy whose role is to talk a
lot in this group? Or did not say anything in
this group? Like that is a super subtle.
Speaker 3 (09:43):
One, it really is.
Speaker 2 (09:45):
And you're never ever going to message chatchibuten have it
not respond like it just always responds. And so I
built this little plug in for Open Clock called tact
that took a bunch of examples of instances where my
claw and a couple other people's clause responded when they
(10:07):
were not supposed to, like in n big as cases,
and then now it runs that every time before it
responds it runs tact and it's like.
Speaker 3 (10:15):
Is this cool?
Speaker 2 (10:16):
Should I be saying something here? And it makes it
much better. It's not perfect still, but it makes it
way way way better. But I think generally we're because
we're moving to this world where everyone's got an agent
and all the agents are interacting, and the agents were
really made for one on one work, They're going to
need a whole new level of training and organization to
(10:38):
help them collaborate effectively.
Speaker 1 (10:42):
There's a lot of subtle complexity there, and it's interesting
to think about how to fix the problem because I
feel like so much of that complexity is tacit, right.
It's the classic thing that like, I don't really know
what the training data is for that. I mean, I
guess it's slack some you know, billion whatever tons of slack.
But it seems hard. That seems like an interesting and
(11:04):
hard and kind of subtle problem.
Speaker 2 (11:06):
It is interesting and hard and subtle, And one of
the beautiful things about AI is it's able to capture
a lot of tacit knowledge in a way that no
other technology is able to do. The ability to use
language at all is to some degree a tacit skill. Yeah,
saying that you can talk about a bit, but being
(11:27):
able to talk about the rules of grammar doesn't make
you able to have a conversation like this.
Speaker 1 (11:31):
Right, that's the big lesson of the last fifty years
of AI.
Speaker 3 (11:34):
Right.
Speaker 1 (11:35):
They spent thirty years trying to get lessons to work
and it didn't work. And they were like, just go
read the internet and figure.
Speaker 3 (11:41):
It out, and it works exactly.
Speaker 1 (11:45):
You have a podcast and I heard you talking to
your editor in chief, the editor in chief at your company,
on the podcast, and she mentioned that she doesn't use
her claw, her plus one, her shadow self, And so
I'm curious. And that's at your company, where the whole
point of being at your company is to play with
(12:06):
all the things and use all the things. And so
I'm sure when you go other places people are like, no,
I'm not going to make a shadow self. I'm me
and I don't want a shadow self.
Speaker 3 (12:13):
Yeah.
Speaker 1 (12:14):
Maybe I'm afraid it's going to take my job. Maybe
I think it's an extremely poor taste, maybe it's a
pain in the ass. Whatever, Like how does it play
and how do you how do you work with that?
How do you get people who don't want to do
it to do it or show them that it's great.
Speaker 2 (12:26):
So first that has changed. Kay, our editor in chief
now has a clock. Her name is Parker, who she
uses for a lot of the a lot of questions
that she might have on the editorial front, like around
the calendar and that kind of stuff.
Speaker 1 (12:43):
I mean, this is more the consulting side of your business, right,
Like you have consciously created a company, as everybody is
supposed to play with stuff. But you go out into
the world to some other company and you're like, yeah,
this is great, it'll save you drudgery, and people are like, yeah,
no way, that's so creepy, Like what do you say.
Speaker 2 (12:57):
Well, I have some interest. I think that's the harder case.
There's a more I think there's a really interesting one too,
which is even inside of our company, there are people
who it's it's to your benefit with this stuff because
like the organizational culture rewards and they're still like, I
actually don't want to use the claw and that's really interesting.
There's there's many reasons why people don't want to, and
(13:19):
it's really depends on the person. Sometimes it's like, look,
I'll use it, but I've got I've got a job,
and I've got like a life outside of this, and
I want to do my job well, but I also
just like want to go home, and my thing is
not playing with new technology. It is playing the flute
or hanging with my kids or whatever it is. And
(13:42):
so I think for those people, just showing them, actually
showing them, here's how you can use this in a
way that is useful for you. And I think it's
most people. It's sort of you can have these like
aha moments, we're like, holy shit, I cannot believe that
that is a thing, and and you just have to
get to that aha moment, get them to that aha moment.
Speaker 1 (14:04):
And to figure out, like you want to show them
how it can be useful to them. Is the thing
when you ask them, like, what's the thing you do
all the time that you wish you didn't have to do?
Speaker 2 (14:12):
I mean, that's that's an easy one. Yeah, that's an
easy one. I think for people who are really really
against it, like really like this is it's terrible for
the world, and I will never use it, I think
like sometimes that's it's not it's just not a thing
that you can ever convince someone. And I think that's fine.
(14:34):
But I think that most of those people one thing
that is happening is they're like, it's going to replace me. Yeah,
well it's going to replace the thing I love. I
have no control over it. Yeah, And I think that
is one perspective and it's super valid. But also I've
seen a lot of people then look at some of
(14:55):
the stuff we do and they're like, oh shit, you
can actually do really cool, interesting creative stuff with this stuff.
I didn't realize that was possible. I thought it was
all slap. So even if you can't have the debate
directly with a lot of those people, I think what
we try to do is is.
Speaker 3 (15:11):
Live in a.
Speaker 2 (15:11):
Way that shows people like, your job might change, but
or the thing that you love might change. But if
you use the models to do the thing that you love,
it's actually it can actually be really cool, and that
is possible. And it's also possible to use it for
slop and you know, being a cyber criminal or whatever,
but if you want to use it for to do
(15:31):
good stuff, you can.
Speaker 1 (15:33):
So fear of losing your job is one, I would say,
not unreasonable fear. There's another related but different kind of
a version that's interesting to me, and I see it
a fair bit in like people who are journalists or writers.
That's sort of the I live in New York and
(15:54):
that's you know, those are the people I hang around with.
And there's a fair bit of AI version that feels
to me kind of identity and ego related, and you know,
I see it especially with respect to writing. And it's
so I'm curious about your experience of that. As a
(16:16):
person who writes uses AI. How do you read that?
How do you think about that?
Speaker 2 (16:24):
Read the sort of ego, the threatened ego? You give
me an example of like a hypothetical writer that you
may have encountered.
Speaker 1 (16:35):
I mean me, I'll just talk about me. Like, I
think I'm good at writing, and I think writing is
getting commodified, and like I think I am aware enough
to recognize that soon I will be able to fairly
easily get a model to create writing that is indistinguishable
(16:57):
from my own writing.
Speaker 2 (16:59):
I actually do not think that's true. And let me
explain why. Okay, because I think I think it matters.
Speaker 1 (17:10):
It's a cliffhanger. After the break, Dan explains, and he
and I debate what AI will mean for writing.
Speaker 2 (17:26):
My feeling about this stuff, first of all, it comes
from what is really good writing what's really good writing.
I think really good writing is the opposite of whatever
cliche is in the sense that it is a true
and honest response to like what you actually think or feel,
(17:47):
and that is very, very highly individual. And it's not
just what you think or feel in one moment. It
is what you think and feel over and over and
over and over again as you are writing and as
you are revising. And a really great piece is full
of all these like little tiny decisions that are yours.
And this is I'm not making this up. This is like,
(18:07):
this is a George Saunders thing, who I love and
I think is a very good it's a very good
articulation of what makes good writing good writing, which means
it's super individual. And one of the things that you
observe with language models is the first time you see
it right, you're like, holy shit, it's doing this amazing thing.
I can't believe it. Oh my god, writing is going
(18:27):
to be over. It's going to copy everything that I do.
And then once you've read like the fiftieth language model
piece of writing, you're like, oh, yeah, it has all
these little things. It's like it says X is X
not y, and it does all these like little inappropriate things.
And what that means when you start to recognize that,
(18:48):
is that you are learning faster than the model. And
the model does not learn as fast as you, because
in order to release a new model, they have to
go gather a bunch of training data and then they
have to like train it, and then they have to
test it, and then they have to put it out.
And right now that process is like months. It used
(19:09):
to be years, Now it's months. And if you are
learning faster than the model, and really great writing is
a reflection of who you actually truly are. You're always
even if you upload your stuff, you're always going to
be way ahead of where the model is because it's
not you. It can't have the experiences that you have.
You can extend this out and say, well, there's going
(19:31):
to be a point at which models are able to
learn on the fly, They're able to like basically be
trained as you talk to them. I think that will
certainly happen. It will probably take at least a few years.
I think we're at least few years away from that.
But even in that case, think about identical twins. They
(19:53):
start at the same point, but as they live even
if they live together, they accumulate different experiences and perspectives,
even if they have the same DNA that make them
different people. And even if you hypothesize an agi that
is a being that knows all about you, it will
have fundamentally a different perspective because it's different from you.
(20:18):
It's a different process, it's exposed to different things. So
while you'll have a lot of points where you're like,
oh my god, it can write exactly like me, what
you're really saying is it can write exactly like me
in a certain narrow circumstance.
Speaker 1 (20:33):
I think the identical twin analogy is fine for me.
Like that one I'm willing to stipulate. And so in
that universe, it's not that it is exactly me, but
it's that it's as good as me, like or better. Right,
And what does good mean? Fine? Blah blah blah, But
like in a professional sense, like not in a human
(20:55):
meaning sense. I've fortunately, like to some extent at least,
sort of taken my ego out of being smart. Like
I think that's a big part of people's response to it. Right,
is like the commodification of intelligens right, and people gain
status from intelligence and I think that will decrease, right,
And so that's interesting in a maybe broader way. But
(21:17):
the my identical twin can write as well as me,
And sure, my identical twin is not me, and I'm
still a person, but in a professional sense, like the
identical twin could write an equally compelling, equally funny, equally
idiosyncratic podcast script sort of, but like surely in thumb setting. Yeah,
I mean, I hope people actually like me as a
human being. Blah blah blah, Like I haven't given given up,
(21:40):
but like I feel like you're being too extreme in
your position. Like, get, I get at the limit what
you're talking about, But there's so much space between here
and there, do you know what I mean?
Speaker 2 (21:56):
Yeah, I think that Here's what I'll give you.
Speaker 1 (22:00):
Yeah.
Speaker 2 (22:01):
What tends to happen with art forms, formats and form
of expression that are older that get as new technology
paradigms come around and make new forms of work possible,
it tends to happen to the older ones. Is they
become higher status, more expensive, more artisanal, and smaller market.
Speaker 1 (22:25):
Yeah. Right, there used to be a portrait painter in
every town.
Speaker 2 (22:29):
Exactly, Yes, And books and magazines newspaper used to be
mass media and now they are. Now they're aspirational. You know,
It's like I want to read a book, but I
read TikTok.
Speaker 1 (22:41):
I read TikTok is. I love reading TikTok. I wish
I wish we read TikTok. I would take it.
Speaker 2 (22:50):
So, so what I will grant what I will grant you,
and I appreciate you pushing on this. What I will
grant you is that there may be a smaller market
for the exact thing that you're doing. I think it
will still exist, and I think it will be higher status,
and I think it will be a more affluent kind
of aspirational thing.
Speaker 1 (23:10):
And you could imagine it being more relationship based, right,
you could imagine like whatever the live show, the Discord channel, whatever,
like Yeah, presumably one hopes being a human being will matter,
like fundamentally, even if the machine can write the same
words I can write, the machine is a machine and
(23:31):
I am a body.
Speaker 2 (23:32):
I think it really will. And also I suspect that
there will be new mediums that are AI enabled that
may not feel like you like them, They may feel
kind of like yucky, but that are that are actually
made by humans.
Speaker 1 (23:53):
I'm open to it being great by the way, Like
I'm not a hater. You know, there's certain things that
make me sad. Was that like a reading at a
bookstore last night? I was like this is this is
this is an anachronism, but it was there. It was
standing m you know.
Speaker 2 (24:11):
Yeah, I love. I love when I go on the
subway and I see people reading.
Speaker 3 (24:15):
I'm like, this is inazing.
Speaker 2 (24:17):
It's the best reading a paper book.
Speaker 1 (24:19):
You were a writer, you were an AI guy, like
just on a process level, like those old Paris Review
interviews where they did interview the writer. They'd be like,
pencil or plas, what time of day? How do you
use AI to write?
Speaker 2 (24:31):
Writing? Any good writing starts with reading. So I use
it all the time to read. So whether that is
taking a picture of something and being like can you
save this for me? Or taking a picture of something
and asking it to like explain it, which is I
think really important for I really I like reading, you know,
(24:52):
like Russian literature, and so I can like be like, hey,
there's this line in check off, what was the actual
original Russian? And how did he really write that?
Speaker 3 (25:01):
You know? So I use it for that.
Speaker 2 (25:05):
I use it a lot for the kind of thing
I just told you, which is I have a lot
of thoughts swimming in my head and I just like
blab to it and then have it like lay it out,
and then I have it helped me construct the argument
so I can like think something through better.
Speaker 3 (25:20):
So's that's a lot.
Speaker 2 (25:20):
Of the initial creative process of figuring out what it
is that I want to say, and then when I'm
actually in draft mode.
Speaker 3 (25:32):
Depends on the piece.
Speaker 2 (25:33):
But like it's really good for Okay, give me like
a take how would you open this piece? And I
don't take it, but it like gives me a little
thing where I'm like, ah, here's like it's not that.
Speaker 4 (25:44):
But it's this as the a I would say, yeah, exactly,
And or it's really good for I'm trying to find
a metaphor similar for this thing.
Speaker 2 (25:59):
Yeah, here's what I'm trying to say, but that's not
quite right, you know, going through a lot of different
options like that. It's really good for that. It's also
great for a lot of the time that you spend writing.
If you're writing about ideas, you spend a lot of
time summarizing ideas, and they're summarizing ideas you already know
about because they're like famous ideas and it's hard to
do that summarization. But a lot of it's pretty wrote.
(26:22):
So anything that feels rote like that, I use it
for that.
Speaker 1 (26:25):
So you use it to like write a first draft
of the like here I'm gonna summarize.
Speaker 2 (26:30):
Yeah, here's what utilitarianism is. Three sentences. Yeah. We also
I use it a lot for editing, so some before
I send it to an editor. We have a few
human human editors. We still employ human editors. I think
we will continue to employ human editors. I often have
it like, okay, read this, what do you think kind
of thing? Help me understand me? Like when am I
(26:53):
missing this argument? And it allows me to think about
things and do work that otherwise I would have had
to go, like get like four different master's degrees or whatever.
And I really like that. And then and then our
editors use it. We have a style guide, so everything
from the developmental editing to we can't do all the
copy with some of the copy, to finding aiisms and
(27:15):
taking them out, like there's a whole list of things
that they all do. And yeah, so it's basically the
whole loop. There's like places where it slots.
Speaker 1 (27:24):
In, and how is it, How has it changed the output,
how has it changed your writing, how you feel about writing,
your relationship to writing.
Speaker 2 (27:31):
I can do much more ambitious pieces. I can do
them more frequently and with less time.
Speaker 1 (27:41):
Everybody is haterish, and I realize it's your job to
not be haterish, especially because like you're in New York
kind of media or media adjacent, and so I appreciate
that role. It does seem like you're overplaying it a
little bit, like I don't know, I don't know how
you feel. Are you as optimistic as I'm reading you
(28:02):
to be? Are you optimistic?
Speaker 2 (28:10):
It's really interesting because I would not have read what
I've been saying in this interview as optimistic in a
like utopian sense. I didn't say utopian, but like because
that's the that's the alternative, right, like the sort of
like the more silicon valley. Well, let's set aside, that's
a straw man alternative. But like when losing jobs came up,
(28:30):
you're like, well, there'll be new jobs, which like eventually,
but we don't know how long that'll take. And there
are real shocks, like you didn't want to be making
furniture in North Carolina when in two thousand and three
you didn't want to be a shearer in Nottingham in
eighteen oh five, like there are times when technological changes
really bad for people.
Speaker 1 (28:51):
I made it. And similarly, when I talked about writing,
like you made a case, but like it's you know,
So that's why I read what you were saying as
relatively optimistic.
Speaker 3 (29:05):
I see.
Speaker 2 (29:07):
I do think that I, as a per person, feel
like the discourse is so bifurcated between people who hate
AI and people who think it's going to be the
best fucking thing in the world and nothing could ever
go wrong with it.
Speaker 1 (29:18):
Yes, the bimodal discourse.
Speaker 2 (29:19):
Yeah, yeah. And so what I love about what we
do at every is we're just like we're just going
to discover what it would mean to like do good
work and live a good life with this stuff, and
we're not going to fake it, like I'm going to
tell you that the claw doesn't work for this specific thing.
So I think I always try to base what I
(29:43):
say in the stuff that I'm encountering, which is always
like a it's a.
Speaker 1 (29:46):
Bubble, it's a bubble, right, Yeah, if you got to
work at a media company right now, A media company
that's building AI tools and doing a consulting is a
good media company to work at.
Speaker 2 (29:56):
It is, it is, it is, And so I think
I generally react to this sort of default negative sentiment
and trying to like go a different way forward. I
am sure that jobs will change a lot, and I'm
sure that there will be some people that lose their
(30:18):
jobs and that we will need to deal with that.
I don't think that that's a something to take lightly,
and I don't I think it should be part of
the conversation. And I also I've been through a lot
of the cycles in the last three and a half
years of Oh my god, I saw GBD three and
it was going to like totally take away X y
(30:39):
Z all these different jobs like immediately.
Speaker 3 (30:41):
And I think if it as.
Speaker 2 (30:44):
What did people in the Middle Ages think would happen
when you reached the horizon? They were like, you know,
it falls off into nothingness, there's dragons, it's like going
to be terrible, right, And I think what happens when
you reach the horizon is you find another one, and
there's always going to be good things and bad things
over the horizon, and probably in similar proportions to the
(31:07):
ones that we have now the problems change, but I
also think that there's the potential to make that much better.
And that's what I'm trying to what I'm trying to do,
and I'm trying not to fall into the trap of
oh my god, over the horizon is like dragons and
it falls off into the void.
Speaker 1 (31:25):
Yeah, no, that's it is interesting. I was thinking about
my own thinking about this, and just yesterday was realizing
that I've been thinking in this kind of binary way,
like we're at the end of the before and we're
about to get to the after. And I was like, oh,
maybe it's not going to be like that. Maybe it's
(31:46):
going to be like with semiconductors, right, maybe it's just
going to be fifty years of now it can do this, Now,
it can do this, and like there is no after,
and it's already now. By the way, this is already after.
That's the other weird thing, right.
Speaker 2 (32:03):
You're reminding me there's this scene in Spaceballs where they like,
we got to get out of here.
Speaker 1 (32:07):
Well that we'll be back in a minute with the
lightning round. We're going to finish with the lightning round.
(32:30):
What was the last argument you had with Ai. Not
about AI, but with AI with the language model or
your claw or whatever.
Speaker 2 (32:38):
I mean, I get frustrated at it all the time.
My girlfriend always says that I'm like, really mean to
my claw.
Speaker 1 (32:43):
I've heard you should be mean to your language model.
I've heard that it makes it work better. I'm not
inclined to do that, but.
Speaker 2 (32:49):
There's conflicting research. I think, just as a moral thing,
you should not be mean to that mean to it.
I'm not always the most moral person. I just but
I do think as a moral thing, you should not
be mean to it.
Speaker 1 (32:58):
And just in case it's my boss, Like, just in
case i'm training it how to talk to me, I
say please and thank you.
Speaker 2 (33:04):
Yes, yes, well that would be more of a pragmatic reason.
Speaker 1 (33:07):
You're right, you're right instrumental. Yes, it's true.
Speaker 2 (33:10):
Yeah, there is some interesting research that if you yell
at claude, it gets anxious and then it gives you
worse results.
Speaker 1 (33:17):
So oh interesting, just the word to the wise, it
gets anxious. Yeah, keep going.
Speaker 2 (33:28):
So, I I mean, I was arguing with it this morning.
I brain dumped a big thing that I've been thinking about,
and then I asked it to help me kind of
figure out what I was trying to say. And it
was obviously just like skipping like many steps in the
argument and filling things in that I hadn't said, and
I was like stop, like we need to go back,
this is not right. Yeah, bringing in ideas that like
(33:52):
I didn't really agree with. So there's a lot of
that going on all the time.
Speaker 1 (33:56):
What's your spiciest day? I take.
Speaker 2 (34:02):
Automation is a lie great go on when you have agents.
The thing that makes an agent work is a living
a relationship with a human, And if the agent does
not have a relationship with a human, it gets stale
very very quickly. There are some corner cases where this
(34:22):
is not true, but in general it gets stale very quickly,
which means that keeping an agent running is a job,
and it's not even just a job for one person.
It is like it's you know, there's one person per
agent usually, but also like there's a whole machinery of
people who are gathering training data and training models and
trying to make these models better that are making your
(34:43):
agent work too, and if they didn't exist that it
would not get any better over time. And what that
says to me is the view of automation that we have,
which is like, oh, you set it up and you
don't need anybody else, and it's like you never have
to hire anyone and you can just go off and
go to the beach or you know, also society breaks
(35:07):
down because no one has jobs. My experience with that
is it's actually work to make sure your agent is working,
and it's leveraged work, but it is work.
Speaker 1 (35:15):
Yes, I mean you need a productivity gain for it
to have a point, right, you need the output to
be there. Well, you need the output to be more
valuable for the same amount of work by a human.
Speaker 2 (35:28):
You, yes, And once you can automate a workflow and
make the old output cheaper to do, there's always like
a larger frame or a larger way of thinking about
things that the human ends up doing. And finding that
language models are currently not very good at.
Speaker 1 (35:47):
Yeah, you fed your journals to a language model. You've
done your research though, right, well, you put it all
on the internet. It's easy research. It was GPT three,
which seems so quaint now, But what you learn uploading
your journals to GPT three a lot?
Speaker 2 (36:10):
I think even back then, I was like, these things
are really good at mirroring back to you who you
are and articulating things about yourself that you probably could
feel but had never said. And we talked about how
powerful articulating things is, and these things are really good
at our at articulating stuff that you're you're kind of
dancing around or it is kind of latent in there,
(36:31):
but you haven't fully said, Like I remember one example
in the GBT three Days, I was at that point
I was recording my therapy sessions and putting them into
GBT three.
Speaker 1 (36:42):
Oh wow, your actual human to human therapy sessions.
Speaker 3 (36:46):
Yeah, yeah, yeah.
Speaker 2 (36:47):
And I will also say I was like I was
not in a good place that at that point, I
have OCD, and and OCD is like it's it's a bitch.
I highly recommend not having OCD. Now it's much better,
and there are very good treatment options available if you
get into the right treatment. Okay, just to say, but
at that point I was, it was like it was
(37:08):
pretty bad. And I'm one of those people that tends
to if something's wrong. I can keep a pretty good
like straight face, and it can be hard to tell
how bad things are. And I was trying to articulate
it was going through something. I was going through a
personal thing, and I was trying to articulate to my
therapist like what I was feeling, and I put it
(37:31):
into GPT three after that and it was like, it
sounds like you feel really overwhelmed. And that's a very
simple observation, but one that I had not made, because
I wouldn't have never thought of myself as being overwhelmed,
and my therapist had not made because I'm kind of
like pretty good at seeming like I've got it under control.
(37:52):
And it just like fucking broke me. It like totally
broke me. And it was it was true, like it
put his finger on something real. And then I went
and talked about it in therapy and it was like
really helpful, and I think it's so good. It's still
so good for that kind of stuff, and that extends
beyond the therapy conversation. It goes into writing and any
(38:14):
kind of creative work. Like the articulation of what you
like is really important, and language models are really good
at that.
Speaker 1 (38:21):
Have you uploaded personal stuff like therapy sessions to AI
more recently?
Speaker 2 (38:27):
Yeah, I think I do it less, probably because I'm
just in a better place. But I definitely still do it.
I think especially for interpersonal situations which happened at work,
they happen at home, all that kind of stuff where.
Speaker 3 (38:46):
You want to.
Speaker 2 (38:48):
There's something that you're feeling that again you can't quite say,
and you need to have a tough conversation, but you
don't really know how to have it in a way
that feels sensitive to the other person. But also its
like really expresses how you actually feel for someone like
me who you know sort of like people pleaser type person,
(39:08):
which you may not have realized from this interview, but
I can, I can get into that mode.
Speaker 1 (39:13):
I found it quite cordial. Did you find itchonistic? I
hope not.
Speaker 3 (39:17):
No, no, no, no, no no, I just yeah, yeah.
Speaker 2 (39:20):
I feel like if you're if you're sort of a
people pleaser type person, that can really help remind It
helps remind me like here's here's what I actually feel
in this situation, and also here's how I can articulate
it in a way that is probably going to be productive,
especially because it knows some of my foibles. And yeah,
I really like it for that. It's not good for
(39:41):
like I wouldn't copy paste the chatchip Ta output and
put it into a text. But it is good for
helping you understand how you feel. Thank you for being
so generous with your time. I really appreciate it. Thank
you for having me. It was a super fun conversation.
Speaker 1 (40:02):
Dan Schipper is the co founder and CEO of every
Today's show was produced by Gabriel Hundreds and edited by
Lydia Jane Kott. Our engineer this week was Hansdale Sheet.
We're always looking for ideas for who to talk to
and what to cover on the show. You can email
us at problem at Pushkin dot fm. You can find
(40:24):
me on x at Jacob Goldstein. You can find me
on LinkedIn. I'm Jacob Goldstein. Thank you very much for
listening to the show, and we'll be back next week
with another episode.
Pushkin. I'm Jacob Goldstein, and this is what's your problem.
My guest today is Dan Shipper. Dan is the co
(00:36):
founder and CEO of a company called Every. Every publishes
newsletters about AI, they develop AI related software, and they
consult to other companies about how to use AI. So
what I'm saying is every is an AI company. And
when Dan describes Every's work culture, he says, we live
(00:56):
in the future, which means that people at every use
AI in ways that lots of the rest of us
will soon be using AI. Given that context, there were
two specific problems that I wanted to talk to Dan about.
Problem number one, how do you build a company where
almost everybody has their own AI agent working alongside them?
(01:19):
Problem number two, how do you use AI as a
tool to improve your writing rather than as a replacement
for writing? On's particularly close to my heart. We talk
about problem number two later in the show. Regarding problem
number one, I will say, almost everybody at every has
(01:40):
their own AI agent, and specifically they use this popular
open source software called open claw. You may have heard
of it. It's basically an AI assistant that you communicate
with via Slack or via text, and at every they
give their clause names. So Dan's claw, for example, is
called R two C two and R two C two
(02:01):
and the other clause are on slack, just like the
human beings who work at every So I wanted to know,
well a lot, really, I want to know a lot.
You know, what do the clause do, how do they
interact with the humans, what works, what doesn't work, what
is super weird? And just what are the clause like.
Speaker 2 (02:20):
What's really interesting about them is that they have personalities.
So you give them a name, they give them a
little bit of a personality. They modify themselves in response
to your messages. So when I say, hey, go do
this thing, it will it starts to realize, okay, I'm
(02:43):
going to be asked to do task act and it
will build things into itself, whether that's into its prompting
or into little tools that it builds for itself to
make it good at that. And so you're kind of
as you talk to it, it is evolving with you
and building functionality for itself to make it good for
you and the things you do.
Speaker 1 (03:04):
It's okay, so you have this agent R two C two.
But to sort of tell me more about how it works,
like I know you've talked before about how it helps
you kind of manage this piece of software. You wrote
this thing called Proof that a lot of people are
like asking you questions about, but just give me more information,
like how do you actually use or work with it
(03:25):
day to day?
Speaker 2 (03:27):
So you know, an easy one is this morning he
went through all of the issues that were submitted in
Proof and then prepared a little document for me that
was like, here's all the stuff that like has gone
run over the next over the last twenty four hours,
and here's what I recommend we do. He also has
(03:51):
a little document that he keeps that's my to do list.
So he pings me in the morning and is like,
here's your to do list. He also, you know, one
of the things I love to do is I always
go to a restaurant nearby the office and have breakfast
in the morning and read. And I will often take
a picture of a thing and be like, hey, can
(04:12):
you explain this to me? And he'll like explain it
and help me think about it.
Speaker 1 (04:15):
Does he text you back? Does he slack you back?
Speaker 2 (04:18):
It's on slack, Yeah, it's all on slack, So you
slack him.
Speaker 1 (04:21):
A picture of like a passage from a book.
Speaker 2 (04:24):
He also keeps all my reading notes, so I'll say, okay,
then now save this, and I have a whole reading
feed that is like all the stuff that I've been
reading over the last couple months. I also just talked
to him about ideas. So today I was, you know,
just thinking through some deep, deep thoughts as I do,
about where we are in AI and where we're going
and all that kind of stuff, and I just like,
(04:46):
monologue is our is one of our products as a
speech to text app. I just took a note with
monologue like a voice memo, and just blabbed into it
for a while and then started talking to R two
about what do you think I'm really trying to say here?
Having something like this that can reflect back to you
what you're saying, yes and put your finger on it
(05:07):
is super valuable, especially if it's something that has it
is the equivalent of an expert on like pretty much
any domain. It's not going to be perfect, but like
you know, it's it's a good stand in for an expert.
Speaker 1 (05:23):
And that to be clear, anybody can do with any
language model, right, Like you don't need a claw for that.
That's like classic nice use of the window on any
decent large language model.
Speaker 2 (05:35):
Totally, it's just a little bit easier to like, I
don't have to He's just always on, So it's like
it's just easier to like slack him in the way
that would have here.
Speaker 1 (05:44):
It's amazing the way the tiny, tiny frictions, right, can
make a difference, even like your own mental frictions. Yeah, yeah,
so okay, So that's like one man in his claw.
There's this other level, which is like claw to claw
and person to claw, right, Like you have this company
that is people and claws, and the agents aren't just
(06:05):
free agents. Each agent is sort of the ganger or
whatever of a person. Tell me about that.
Speaker 2 (06:15):
That gets really interesting because what tends to happen, like
I said, is if you have a relationship with your claw,
it becomes good and the things you're good at. And
if you start using it in slack, other people realize
that it's good for the things that that you're using
it for. So you've transferred some of your reputation and
trust to your claw. And so for example, Austin, our
(06:36):
head of growth, has a claw named Montane hi Brow,
and he uses Montane for all of the numbers, so
like how much should we grow? How many trials did
this peace drive? Like all that kind of stuff, And
the editorial team just started like immediately being like at Montane,
like how did this how did this essay we published do?
And so it then it becomes this thing that feeds
(06:57):
on itself. For now Montane is the is the guy.
Is the guy for this where Austin used to be,
but now it's Montane. And it depends though on Austin
making sure that Montane is always up to date and working.
But instead of answering the questions directly, he's working on
the system that answers the questions.
Speaker 3 (07:14):
So he's sort of like moved up a level.
Speaker 1 (07:15):
Yeah, he's managing like the the agent works for him.
That's what a manager does, right, Like you make sure
that the person who's working for you is like doing
their job and answering questions correctly and it's happy and
with the agent, you don't have to worry about.
Speaker 2 (07:29):
That, which is really fun for me to see because
I wrote this piece like three years ago called the
Knowledge Economies over Welcome to the Allocation Economy, and the
idea was the skill of the future will be learning
to manage agents, allocate intelligence, and it's it's starting to
be a thing.
Speaker 3 (07:47):
Now.
Speaker 2 (07:47):
It's really interesting to see.
Speaker 1 (07:49):
You go to write a piece now that headline, I
was right.
Speaker 3 (07:52):
I was right.
Speaker 1 (07:59):
When I go back to like the company and multiple
clause because you know, you gave this sort of tidy
thing of like they just ask the guy now they
ask his agent. Surely it is more complicated and interesting
than just that, Like it must be hard in some ways,
it must be weird in some ways, you must be
figuring things out. You must have gotten something wrong.
Speaker 3 (08:19):
Yes, yes, yes, yes.
Speaker 2 (08:22):
One thing that I really love is they sometimes don't
know when they should or shouldn't respond, And there's some
and there there sometimes are often too chatty. So for example,
you might tag a claw and ask it a question
to ask you to do some work, and then in
that thread in slack, you might, like one of your
(08:43):
coworkers might say something and then but it was not
directed at the claw. It was just like a comment
on what the claw was doing right, And then you're
trying to respond, but then the claw jumps in and
it's like yes, you're totally right or whatever, and it's like, dude,
this is because it's.
Speaker 1 (08:58):
Like there's a chatbot underneath it, and it's so hard
for a chatbot not to reply when they're yeah, yeah, yeah,
basically what's going on?
Speaker 3 (09:05):
Basically?
Speaker 2 (09:05):
But I think the the the way that I think
about it is these chatbots are trained. They only know
about one on one conversation.
Speaker 1 (09:17):
Uh huh.
Speaker 2 (09:19):
They don't really know about I'm in this group environment.
Here's how you act in a group environment, which is.
Speaker 1 (09:25):
So much subtler and harder, right, like all the cues
that even change from from context to context for people
from work to home, from one job to another job,
depending on what your role is in a particular group.
Are you the guy whose role is to talk a
lot in this group? Or did not say anything in
this group? Like that is a super subtle.
Speaker 3 (09:43):
One, it really is.
Speaker 2 (09:45):
And you're never ever going to message chatchibuten have it
not respond like it just always responds. And so I
built this little plug in for Open Clock called tact
that took a bunch of examples of instances where my
claw and a couple other people's clause responded when they
(10:07):
were not supposed to, like in n big as cases,
and then now it runs that every time before it
responds it runs tact and it's like.
Speaker 3 (10:15):
Is this cool?
Speaker 2 (10:16):
Should I be saying something here? And it makes it
much better. It's not perfect still, but it makes it
way way way better. But I think generally we're because
we're moving to this world where everyone's got an agent
and all the agents are interacting, and the agents were
really made for one on one work, They're going to
need a whole new level of training and organization to
(10:38):
help them collaborate effectively.
Speaker 1 (10:42):
There's a lot of subtle complexity there, and it's interesting
to think about how to fix the problem because I
feel like so much of that complexity is tacit, right.
It's the classic thing that like, I don't really know
what the training data is for that. I mean, I
guess it's slack some you know, billion whatever tons of slack.
But it seems hard. That seems like an interesting and
(11:04):
hard and kind of subtle problem.
Speaker 2 (11:06):
It is interesting and hard and subtle, And one of
the beautiful things about AI is it's able to capture
a lot of tacit knowledge in a way that no
other technology is able to do. The ability to use
language at all is to some degree a tacit skill. Yeah,
saying that you can talk about a bit, but being
(11:27):
able to talk about the rules of grammar doesn't make
you able to have a conversation like this.
Speaker 1 (11:31):
Right, that's the big lesson of the last fifty years
of AI.
Speaker 3 (11:34):
Right.
Speaker 1 (11:35):
They spent thirty years trying to get lessons to work
and it didn't work. And they were like, just go
read the internet and figure.
Speaker 3 (11:41):
It out, and it works exactly.
Speaker 1 (11:45):
You have a podcast and I heard you talking to
your editor in chief, the editor in chief at your company,
on the podcast, and she mentioned that she doesn't use
her claw, her plus one, her shadow self, And so
I'm curious. And that's at your company, where the whole
point of being at your company is to play with
(12:06):
all the things and use all the things. And so
I'm sure when you go other places people are like, no,
I'm not going to make a shadow self. I'm me
and I don't want a shadow self.
Speaker 3 (12:13):
Yeah.
Speaker 1 (12:14):
Maybe I'm afraid it's going to take my job. Maybe
I think it's an extremely poor taste, maybe it's a
pain in the ass. Whatever, Like how does it play
and how do you how do you work with that?
How do you get people who don't want to do
it to do it or show them that it's great.
Speaker 2 (12:26):
So first that has changed. Kay, our editor in chief
now has a clock. Her name is Parker, who she
uses for a lot of the a lot of questions
that she might have on the editorial front, like around
the calendar and that kind of stuff.
Speaker 1 (12:43):
I mean, this is more the consulting side of your business, right,
Like you have consciously created a company, as everybody is
supposed to play with stuff. But you go out into
the world to some other company and you're like, yeah,
this is great, it'll save you drudgery, and people are like, yeah,
no way, that's so creepy, Like what do you say.
Speaker 2 (12:57):
Well, I have some interest. I think that's the harder case.
There's a more I think there's a really interesting one too,
which is even inside of our company, there are people
who it's it's to your benefit with this stuff because
like the organizational culture rewards and they're still like, I
actually don't want to use the claw and that's really interesting.
There's there's many reasons why people don't want to, and
(13:19):
it's really depends on the person. Sometimes it's like, look,
I'll use it, but I've got I've got a job,
and I've got like a life outside of this, and
I want to do my job well, but I also
just like want to go home, and my thing is
not playing with new technology. It is playing the flute
or hanging with my kids or whatever it is. And
(13:42):
so I think for those people, just showing them, actually
showing them, here's how you can use this in a
way that is useful for you. And I think it's
most people. It's sort of you can have these like
aha moments, we're like, holy shit, I cannot believe that
that is a thing, and and you just have to
get to that aha moment, get them to that aha moment.
Speaker 1 (14:04):
And to figure out, like you want to show them
how it can be useful to them. Is the thing
when you ask them, like, what's the thing you do
all the time that you wish you didn't have to do?
Speaker 2 (14:12):
I mean, that's that's an easy one. Yeah, that's an
easy one. I think for people who are really really
against it, like really like this is it's terrible for
the world, and I will never use it, I think
like sometimes that's it's not it's just not a thing
that you can ever convince someone. And I think that's fine.
(14:34):
But I think that most of those people one thing
that is happening is they're like, it's going to replace me. Yeah,
well it's going to replace the thing I love. I
have no control over it. Yeah, And I think that
is one perspective and it's super valid. But also I've
seen a lot of people then look at some of
(14:55):
the stuff we do and they're like, oh shit, you
can actually do really cool, interesting creative stuff with this stuff.
I didn't realize that was possible. I thought it was
all slap. So even if you can't have the debate
directly with a lot of those people, I think what
we try to do is is.
Speaker 3 (15:11):
Live in a.
Speaker 2 (15:11):
Way that shows people like, your job might change, but
or the thing that you love might change. But if
you use the models to do the thing that you love,
it's actually it can actually be really cool, and that
is possible. And it's also possible to use it for
slop and you know, being a cyber criminal or whatever,
but if you want to use it for to do
(15:31):
good stuff, you can.
Speaker 1 (15:33):
So fear of losing your job is one, I would say,
not unreasonable fear. There's another related but different kind of
a version that's interesting to me, and I see it
a fair bit in like people who are journalists or writers.
That's sort of the I live in New York and
(15:54):
that's you know, those are the people I hang around with.
And there's a fair bit of AI version that feels
to me kind of identity and ego related, and you know,
I see it especially with respect to writing. And it's
so I'm curious about your experience of that. As a
(16:16):
person who writes uses AI. How do you read that?
How do you think about that?
Speaker 2 (16:24):
Read the sort of ego, the threatened ego? You give
me an example of like a hypothetical writer that you
may have encountered.
Speaker 1 (16:35):
I mean me, I'll just talk about me. Like, I
think I'm good at writing, and I think writing is
getting commodified, and like I think I am aware enough
to recognize that soon I will be able to fairly
easily get a model to create writing that is indistinguishable
(16:57):
from my own writing.
Speaker 2 (16:59):
I actually do not think that's true. And let me
explain why. Okay, because I think I think it matters.
Speaker 1 (17:10):
It's a cliffhanger. After the break, Dan explains, and he
and I debate what AI will mean for writing.
Speaker 2 (17:26):
My feeling about this stuff, first of all, it comes
from what is really good writing what's really good writing.
I think really good writing is the opposite of whatever
cliche is in the sense that it is a true
and honest response to like what you actually think or feel,
(17:47):
and that is very, very highly individual. And it's not
just what you think or feel in one moment. It
is what you think and feel over and over and
over and over again as you are writing and as
you are revising. And a really great piece is full
of all these like little tiny decisions that are yours.
And this is I'm not making this up. This is like,
(18:07):
this is a George Saunders thing, who I love and
I think is a very good it's a very good
articulation of what makes good writing good writing, which means
it's super individual. And one of the things that you
observe with language models is the first time you see
it right, you're like, holy shit, it's doing this amazing thing.
I can't believe it. Oh my god, writing is going
(18:27):
to be over. It's going to copy everything that I do.
And then once you've read like the fiftieth language model
piece of writing, you're like, oh, yeah, it has all
these little things. It's like it says X is X
not y, and it does all these like little inappropriate things.
And what that means when you start to recognize that,
(18:48):
is that you are learning faster than the model. And
the model does not learn as fast as you, because
in order to release a new model, they have to
go gather a bunch of training data and then they
have to like train it, and then they have to
test it, and then they have to put it out.
And right now that process is like months. It used
(19:09):
to be years, Now it's months. And if you are
learning faster than the model, and really great writing is
a reflection of who you actually truly are. You're always
even if you upload your stuff, you're always going to
be way ahead of where the model is because it's
not you. It can't have the experiences that you have.
You can extend this out and say, well, there's going
(19:31):
to be a point at which models are able to
learn on the fly, They're able to like basically be
trained as you talk to them. I think that will
certainly happen. It will probably take at least a few years.
I think we're at least few years away from that.
But even in that case, think about identical twins. They
(19:53):
start at the same point, but as they live even
if they live together, they accumulate different experiences and perspectives,
even if they have the same DNA that make them
different people. And even if you hypothesize an agi that
is a being that knows all about you, it will
have fundamentally a different perspective because it's different from you.
(20:18):
It's a different process, it's exposed to different things. So
while you'll have a lot of points where you're like,
oh my god, it can write exactly like me, what
you're really saying is it can write exactly like me
in a certain narrow circumstance.
Speaker 1 (20:33):
I think the identical twin analogy is fine for me.
Like that one I'm willing to stipulate. And so in
that universe, it's not that it is exactly me, but
it's that it's as good as me, like or better. Right,
And what does good mean? Fine? Blah blah blah, But
like in a professional sense, like not in a human
(20:55):
meaning sense. I've fortunately, like to some extent at least,
sort of taken my ego out of being smart. Like
I think that's a big part of people's response to it. Right,
is like the commodification of intelligens right, and people gain
status from intelligence and I think that will decrease, right,
And so that's interesting in a maybe broader way. But
(21:17):
the my identical twin can write as well as me,
And sure, my identical twin is not me, and I'm
still a person, but in a professional sense, like the
identical twin could write an equally compelling, equally funny, equally
idiosyncratic podcast script sort of, but like surely in thumb setting. Yeah,
I mean, I hope people actually like me as a
human being. Blah blah blah, Like I haven't given given up,
(21:40):
but like I feel like you're being too extreme in
your position. Like, get, I get at the limit what
you're talking about, But there's so much space between here
and there, do you know what I mean?
Speaker 2 (21:56):
Yeah, I think that Here's what I'll give you.
Speaker 1 (22:00):
Yeah.
Speaker 2 (22:01):
What tends to happen with art forms, formats and form
of expression that are older that get as new technology
paradigms come around and make new forms of work possible,
it tends to happen to the older ones. Is they
become higher status, more expensive, more artisanal, and smaller market.
Speaker 1 (22:25):
Yeah. Right, there used to be a portrait painter in
every town.
Speaker 2 (22:29):
Exactly, Yes, And books and magazines newspaper used to be
mass media and now they are. Now they're aspirational. You know,
It's like I want to read a book, but I
read TikTok.
Speaker 1 (22:41):
I read TikTok is. I love reading TikTok. I wish
I wish we read TikTok. I would take it.
Speaker 2 (22:50):
So, so what I will grant what I will grant you,
and I appreciate you pushing on this. What I will
grant you is that there may be a smaller market
for the exact thing that you're doing. I think it
will still exist, and I think it will be higher status,
and I think it will be a more affluent kind
of aspirational thing.
Speaker 1 (23:10):
And you could imagine it being more relationship based, right,
you could imagine like whatever the live show, the Discord channel, whatever,
like Yeah, presumably one hopes being a human being will matter,
like fundamentally, even if the machine can write the same
words I can write, the machine is a machine and
(23:31):
I am a body.
Speaker 2 (23:32):
I think it really will. And also I suspect that
there will be new mediums that are AI enabled that
may not feel like you like them, They may feel
kind of like yucky, but that are that are actually
made by humans.
Speaker 1 (23:53):
I'm open to it being great by the way, Like
I'm not a hater. You know, there's certain things that
make me sad. Was that like a reading at a
bookstore last night? I was like this is this is
this is an anachronism, but it was there. It was
standing m you know.
Speaker 2 (24:11):
Yeah, I love. I love when I go on the
subway and I see people reading.
Speaker 3 (24:15):
I'm like, this is inazing.
Speaker 2 (24:17):
It's the best reading a paper book.
Speaker 1 (24:19):
You were a writer, you were an AI guy, like
just on a process level, like those old Paris Review
interviews where they did interview the writer. They'd be like,
pencil or plas, what time of day? How do you
use AI to write?
Speaker 2 (24:31):
Writing? Any good writing starts with reading. So I use
it all the time to read. So whether that is
taking a picture of something and being like can you
save this for me? Or taking a picture of something
and asking it to like explain it, which is I
think really important for I really I like reading, you know,
(24:52):
like Russian literature, and so I can like be like, hey,
there's this line in check off, what was the actual
original Russian? And how did he really write that?
Speaker 3 (25:01):
You know? So I use it for that.
Speaker 2 (25:05):
I use it a lot for the kind of thing
I just told you, which is I have a lot
of thoughts swimming in my head and I just like
blab to it and then have it like lay it out,
and then I have it helped me construct the argument
so I can like think something through better.
Speaker 3 (25:20):
So's that's a lot.
Speaker 2 (25:20):
Of the initial creative process of figuring out what it
is that I want to say, and then when I'm
actually in draft mode.
Speaker 3 (25:32):
Depends on the piece.
Speaker 2 (25:33):
But like it's really good for Okay, give me like
a take how would you open this piece? And I
don't take it, but it like gives me a little
thing where I'm like, ah, here's like it's not that.
Speaker 4 (25:44):
But it's this as the a I would say, yeah, exactly,
And or it's really good for I'm trying to find
a metaphor similar for this thing.
Speaker 2 (25:59):
Yeah, here's what I'm trying to say, but that's not
quite right, you know, going through a lot of different
options like that. It's really good for that. It's also
great for a lot of the time that you spend writing.
If you're writing about ideas, you spend a lot of
time summarizing ideas, and they're summarizing ideas you already know
about because they're like famous ideas and it's hard to
do that summarization. But a lot of it's pretty wrote.
(26:22):
So anything that feels rote like that, I use it
for that.
Speaker 1 (26:25):
So you use it to like write a first draft
of the like here I'm gonna summarize.
Speaker 2 (26:30):
Yeah, here's what utilitarianism is. Three sentences. Yeah. We also
I use it a lot for editing, so some before
I send it to an editor. We have a few
human human editors. We still employ human editors. I think
we will continue to employ human editors. I often have
it like, okay, read this, what do you think kind
of thing? Help me understand me? Like when am I
(26:53):
missing this argument? And it allows me to think about
things and do work that otherwise I would have had
to go, like get like four different master's degrees or whatever.
And I really like that. And then and then our
editors use it. We have a style guide, so everything
from the developmental editing to we can't do all the
copy with some of the copy, to finding aiisms and
(27:15):
taking them out, like there's a whole list of things
that they all do. And yeah, so it's basically the
whole loop. There's like places where it slots.
Speaker 1 (27:24):
In, and how is it, How has it changed the output,
how has it changed your writing, how you feel about writing,
your relationship to writing.
Speaker 2 (27:31):
I can do much more ambitious pieces. I can do
them more frequently and with less time.
Speaker 1 (27:41):
Everybody is haterish, and I realize it's your job to
not be haterish, especially because like you're in New York
kind of media or media adjacent, and so I appreciate
that role. It does seem like you're overplaying it a
little bit, like I don't know, I don't know how
you feel. Are you as optimistic as I'm reading you
(28:02):
to be? Are you optimistic?
Speaker 2 (28:10):
It's really interesting because I would not have read what
I've been saying in this interview as optimistic in a
like utopian sense. I didn't say utopian, but like because
that's the that's the alternative, right, like the sort of
like the more silicon valley. Well, let's set aside, that's
a straw man alternative. But like when losing jobs came up,
(28:30):
you're like, well, there'll be new jobs, which like eventually,
but we don't know how long that'll take. And there
are real shocks, like you didn't want to be making
furniture in North Carolina when in two thousand and three
you didn't want to be a shearer in Nottingham in
eighteen oh five, like there are times when technological changes
really bad for people.
Speaker 1 (28:51):
I made it. And similarly, when I talked about writing,
like you made a case, but like it's you know,
So that's why I read what you were saying as
relatively optimistic.
Speaker 3 (29:05):
I see.
Speaker 2 (29:07):
I do think that I, as a per person, feel
like the discourse is so bifurcated between people who hate
AI and people who think it's going to be the
best fucking thing in the world and nothing could ever
go wrong with it.
Speaker 1 (29:18):
Yes, the bimodal discourse.
Speaker 2 (29:19):
Yeah, yeah. And so what I love about what we
do at every is we're just like we're just going
to discover what it would mean to like do good
work and live a good life with this stuff, and
we're not going to fake it, like I'm going to
tell you that the claw doesn't work for this specific thing.
So I think I always try to base what I
(29:43):
say in the stuff that I'm encountering, which is always
like a it's a.
Speaker 1 (29:46):
Bubble, it's a bubble, right, Yeah, if you got to
work at a media company right now, A media company
that's building AI tools and doing a consulting is a
good media company to work at.
Speaker 2 (29:56):
It is, it is, it is, And so I think
I generally react to this sort of default negative sentiment
and trying to like go a different way forward. I
am sure that jobs will change a lot, and I'm
sure that there will be some people that lose their
(30:18):
jobs and that we will need to deal with that.
I don't think that that's a something to take lightly,
and I don't I think it should be part of
the conversation. And I also I've been through a lot
of the cycles in the last three and a half
years of Oh my god, I saw GBD three and
it was going to like totally take away X y
(30:39):
Z all these different jobs like immediately.
Speaker 3 (30:41):
And I think if it as.
Speaker 2 (30:44):
What did people in the Middle Ages think would happen
when you reached the horizon? They were like, you know,
it falls off into nothingness, there's dragons, it's like going
to be terrible, right, And I think what happens when
you reach the horizon is you find another one, and
there's always going to be good things and bad things
over the horizon, and probably in similar proportions to the
(31:07):
ones that we have now the problems change, but I
also think that there's the potential to make that much better.
And that's what I'm trying to what I'm trying to do,
and I'm trying not to fall into the trap of
oh my god, over the horizon is like dragons and
it falls off into the void.
Speaker 1 (31:25):
Yeah, no, that's it is interesting. I was thinking about
my own thinking about this, and just yesterday was realizing
that I've been thinking in this kind of binary way,
like we're at the end of the before and we're
about to get to the after. And I was like, oh,
maybe it's not going to be like that. Maybe it's
(31:46):
going to be like with semiconductors, right, maybe it's just
going to be fifty years of now it can do this, Now,
it can do this, and like there is no after,
and it's already now. By the way, this is already after.
That's the other weird thing, right.
Speaker 2 (32:03):
You're reminding me there's this scene in Spaceballs where they like,
we got to get out of here.
Speaker 1 (32:07):
Well that we'll be back in a minute with the
lightning round. We're going to finish with the lightning round.
(32:30):
What was the last argument you had with Ai. Not
about AI, but with AI with the language model or
your claw or whatever.
Speaker 2 (32:38):
I mean, I get frustrated at it all the time.
My girlfriend always says that I'm like, really mean to
my claw.
Speaker 1 (32:43):
I've heard you should be mean to your language model.
I've heard that it makes it work better. I'm not
inclined to do that, but.
Speaker 2 (32:49):
There's conflicting research. I think, just as a moral thing,
you should not be mean to that mean to it.
I'm not always the most moral person. I just but
I do think as a moral thing, you should not
be mean to it.
Speaker 1 (32:58):
And just in case it's my boss, Like, just in
case i'm training it how to talk to me, I
say please and thank you.
Speaker 2 (33:04):
Yes, yes, well that would be more of a pragmatic reason.
Speaker 1 (33:07):
You're right, you're right instrumental. Yes, it's true.
Speaker 2 (33:10):
Yeah, there is some interesting research that if you yell
at claude, it gets anxious and then it gives you
worse results.
Speaker 1 (33:17):
So oh interesting, just the word to the wise, it
gets anxious. Yeah, keep going.
Speaker 2 (33:28):
So, I I mean, I was arguing with it this morning.
I brain dumped a big thing that I've been thinking about,
and then I asked it to help me kind of
figure out what I was trying to say. And it
was obviously just like skipping like many steps in the
argument and filling things in that I hadn't said, and
I was like stop, like we need to go back,
this is not right. Yeah, bringing in ideas that like
(33:52):
I didn't really agree with. So there's a lot of
that going on all the time.
Speaker 1 (33:56):
What's your spiciest day? I take.
Speaker 2 (34:02):
Automation is a lie great go on when you have agents.
The thing that makes an agent work is a living
a relationship with a human, And if the agent does
not have a relationship with a human, it gets stale
very very quickly. There are some corner cases where this
(34:22):
is not true, but in general it gets stale very quickly,
which means that keeping an agent running is a job,
and it's not even just a job for one person.
It is like it's you know, there's one person per
agent usually, but also like there's a whole machinery of
people who are gathering training data and training models and
trying to make these models better that are making your
(34:43):
agent work too, and if they didn't exist that it
would not get any better over time. And what that
says to me is the view of automation that we have,
which is like, oh, you set it up and you
don't need anybody else, and it's like you never have
to hire anyone and you can just go off and
go to the beach or you know, also society breaks
(35:07):
down because no one has jobs. My experience with that
is it's actually work to make sure your agent is working,
and it's leveraged work, but it is work.
Speaker 1 (35:15):
Yes, I mean you need a productivity gain for it
to have a point, right, you need the output to
be there. Well, you need the output to be more
valuable for the same amount of work by a human.
Speaker 2 (35:28):
You, yes, And once you can automate a workflow and
make the old output cheaper to do, there's always like
a larger frame or a larger way of thinking about
things that the human ends up doing. And finding that
language models are currently not very good at.
Speaker 1 (35:47):
Yeah, you fed your journals to a language model. You've
done your research though, right, well, you put it all
on the internet. It's easy research. It was GPT three,
which seems so quaint now, But what you learn uploading
your journals to GPT three a lot?
Speaker 2 (36:10):
I think even back then, I was like, these things
are really good at mirroring back to you who you
are and articulating things about yourself that you probably could
feel but had never said. And we talked about how
powerful articulating things is, and these things are really good
at our at articulating stuff that you're you're kind of
dancing around or it is kind of latent in there,
(36:31):
but you haven't fully said, Like I remember one example
in the GBT three Days, I was at that point
I was recording my therapy sessions and putting them into
GBT three.
Speaker 1 (36:42):
Oh wow, your actual human to human therapy sessions.
Speaker 3 (36:46):
Yeah, yeah, yeah.
Speaker 2 (36:47):
And I will also say I was like I was
not in a good place that at that point, I
have OCD, and and OCD is like it's it's a bitch.
I highly recommend not having OCD. Now it's much better,
and there are very good treatment options available if you
get into the right treatment. Okay, just to say, but
at that point I was, it was like it was
(37:08):
pretty bad. And I'm one of those people that tends
to if something's wrong. I can keep a pretty good
like straight face, and it can be hard to tell
how bad things are. And I was trying to articulate
it was going through something. I was going through a
personal thing, and I was trying to articulate to my
therapist like what I was feeling, and I put it
(37:31):
into GPT three after that and it was like, it
sounds like you feel really overwhelmed. And that's a very
simple observation, but one that I had not made, because
I wouldn't have never thought of myself as being overwhelmed,
and my therapist had not made because I'm kind of
like pretty good at seeming like I've got it under control.
(37:52):
And it just like fucking broke me. It like totally
broke me. And it was it was true, like it
put his finger on something real. And then I went
and talked about it in therapy and it was like
really helpful, and I think it's so good. It's still
so good for that kind of stuff, and that extends
beyond the therapy conversation. It goes into writing and any
(38:14):
kind of creative work. Like the articulation of what you
like is really important, and language models are really good
at that.
Speaker 1 (38:21):
Have you uploaded personal stuff like therapy sessions to AI
more recently?
Speaker 2 (38:27):
Yeah, I think I do it less, probably because I'm
just in a better place. But I definitely still do it.
I think especially for interpersonal situations which happened at work,
they happen at home, all that kind of stuff where.
Speaker 3 (38:46):
You want to.
Speaker 2 (38:48):
There's something that you're feeling that again you can't quite say,
and you need to have a tough conversation, but you
don't really know how to have it in a way
that feels sensitive to the other person. But also its
like really expresses how you actually feel for someone like
me who you know sort of like people pleaser type person,
(39:08):
which you may not have realized from this interview, but
I can, I can get into that mode.
Speaker 1 (39:13):
I found it quite cordial. Did you find itchonistic? I
hope not.
Speaker 3 (39:17):
No, no, no, no, no no, I just yeah, yeah.
Speaker 2 (39:20):
I feel like if you're if you're sort of a
people pleaser type person, that can really help remind It
helps remind me like here's here's what I actually feel
in this situation, and also here's how I can articulate
it in a way that is probably going to be productive,
especially because it knows some of my foibles. And yeah,
I really like it for that. It's not good for
(39:41):
like I wouldn't copy paste the chatchip Ta output and
put it into a text. But it is good for
helping you understand how you feel. Thank you for being
so generous with your time. I really appreciate it. Thank
you for having me. It was a super fun conversation.
Speaker 1 (40:02):
Dan Schipper is the co founder and CEO of every
Today's show was produced by Gabriel Hundreds and edited by
Lydia Jane Kott. Our engineer this week was Hansdale Sheet.
We're always looking for ideas for who to talk to
and what to cover on the show. You can email
us at problem at Pushkin dot fm. You can find
(40:24):
me on x at Jacob Goldstein. You can find me
on LinkedIn. I'm Jacob Goldstein. Thank you very much for
listening to the show, and we'll be back next week
with another episode.
Popular Podcasts
Hey Jonas!
Hey Jonas! The official Jonas Brothers podcast. Hosted by Kevin, Joe, and Nick Jonas. It’s the Jonas Brothers you know... musicians, actors, and well, yes, brothers. Now, they’re sharing another side of themselves in the playful, intimate, and irreverent way only they can. Spend time with the Jonas Brothers here and stay a little bit longer for deep conversations like never before.
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.
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.