The AI Copy‑Paste Problem
Key Points
- The biggest emerging strategic issue in AI isn’t ethics or security, but the “copy‑paste problem”: while LLMs dramatically lower the cost of intelligence, moving the generated data and code between tools remains painfully difficult.
- Traditional software business models that relied on lock‑in (e.g., paying for a SaaS and staying stuck with it) are breaking down because AI makes switching cheap, making data interoperability essential.
- Even though LLMs can instantly produce useful artifacts (like a React component from Claude), there’s no seamless way to integrate those synthetic tokens into existing workflows, convert them to other languages, or share them without manual copy‑pasting.
- This persistent data‑silo issue limits the real value of AI, because intelligence is only as good as the data it can operate on, and bringing AI‑generated output back into production pipelines is currently a major friction point.
- Addressing this interoperability gap will be a critical competitive advantage for companies that want to capitalize on the falling cost of intelligence.
Sections
- AI Undermines Software Lock‑In - The presenter warns that the dramatic drop in AI's cost of intelligence is rendering classic software lock‑in business models ineffective, since it’s now cheap to replace tools.
- Scrapping Projects and SaaS Loyalty - The speaker argues that negligible refactoring costs justify restarting projects, and lasting SaaS loyalty is earned by delivering flawless end‑to‑end data flows and easy data ingress/egress rather than assuming customers are inherently disloyal.
Full Transcript
# The AI Copy‑Paste Problem **Source:** [https://www.youtube.com/watch?v=xdFqZpBpXTo](https://www.youtube.com/watch?v=xdFqZpBpXTo) **Duration:** 00:06:14 ## Summary - The biggest emerging strategic issue in AI isn’t ethics or security, but the “copy‑paste problem”: while LLMs dramatically lower the cost of intelligence, moving the generated data and code between tools remains painfully difficult. - Traditional software business models that relied on lock‑in (e.g., paying for a SaaS and staying stuck with it) are breaking down because AI makes switching cheap, making data interoperability essential. - Even though LLMs can instantly produce useful artifacts (like a React component from Claude), there’s no seamless way to integrate those synthetic tokens into existing workflows, convert them to other languages, or share them without manual copy‑pasting. - This persistent data‑silo issue limits the real value of AI, because intelligence is only as good as the data it can operate on, and bringing AI‑generated output back into production pipelines is currently a major friction point. - Addressing this interoperability gap will be a critical competitive advantage for companies that want to capitalize on the falling cost of intelligence. ## Sections - [00:00:00](https://www.youtube.com/watch?v=xdFqZpBpXTo&t=0s) **AI Undermines Software Lock‑In** - The presenter warns that the dramatic drop in AI's cost of intelligence is rendering classic software lock‑in business models ineffective, since it’s now cheap to replace tools. - [00:03:49](https://www.youtube.com/watch?v=xdFqZpBpXTo&t=229s) **Scrapping Projects and SaaS Loyalty** - The speaker argues that negligible refactoring costs justify restarting projects, and lasting SaaS loyalty is earned by delivering flawless end‑to‑end data flows and easy data ingress/egress rather than assuming customers are inherently disloyal. ## Full Transcript
So, the last time I did this kind of a
YouTube, it was very unpopular. So, if
you think this is terrible, okay, I'm
doing it anyway. I want to talk about
something strategic. I want to talk
about a problem I see is critical in the
AI industry that we are not talking
about very much. It's not ethics. It's
not security. It's not privacy. We talk
about all those things. No, this is copy
paste. It's very simple.
Fundamentally, AI is enabling the cost
of building anything to drop through the
floor because the cost of intelligence
is falling. That's something we talk
about all the time. Heck, most of my
YouTube channel is ways to build things
that are getting easier and easier. The
problem is the old method no longer
works if you have intelligence going
through the floor. And by the old
method, I mean the old strategy for
software. software in the 2010s was
built around the idea that you could
build a tool that people would be loyal
to and that they would pay for it. Don't
you love that my Slack is going off?
Isn't that just appropriate? Uh that
people would pay for and that when they
paid for it, they would be loyal to. And
so at the end of the day, if I bought
Salesforce, I was stuck in Salesforce,
right? And there was no way I was
leaving. Well, famously CLA in 2024
rebuilt and left Salesforce and
embarrassed Mark Beni off publicly on
stage and there was this whole thing.
But the point is not that CLA
individually left Salesforce. The point
is is it is cheaper now to leave and
that makes data interoperability more
important. And if that bores you, I'm
sorry. You can move on. But the point is
really critical and you will live with
it whether it bores you or not. We all
live on applications that use data.
Essentially, LLMs are taking the cost of
intelligence and driving it through the
floor, but data is still stuck in silos.
Data is still not easy to get out. Data
is still really, really hard to get back
and forth. It's why I call it the copy
paste problem. Data is tough to move
around. I'm not talking about ETLs or
pipelines if you're a data engineer.
What I'm saying is that
fundamentally intelligence is only as
good as the data it can operate against.
And intelligence is enabling us to
produce synthetic tokens. LLM produce
tokens all the time. They're correctly
categorized as synthetic tokens. Getting
those synthetic tokens back into our
work streams is miserably hard right
now. Let me give you a few examples.
Let's look at Claude. Claude produces a
great little React component that I
think is a nice design for a PM
dashboard. What do I do right now? I can
publish it and I can send it to my
designer and and the designer can say I
want to work on it but then how do they
use it right? Like do they copy and
paste it? Like that's really pluggy. Uh
or my engineer can say I don't want it
to be in React. I want it to be in
something else. It should be in
Typescript. It should be in something
else. Uh great. Why is it hard? Why is
it hard to go from one tool to another?
I know the reason as a PM from the 2010s
and the 2000s. It's hard on purpose
because you want to lock people into
software. I know that reason and I know
that people still think that reason in
boardrooms today. People still think
that reason in product organizations
today. But the problem is the loyalty
ROI calculus has shifted. And again, if
this bores you, I don't care. It really
matters. The loyalty ROI calculus is
such now that no one is loyal to tools
the way they were. I am in a world as an
AI builder where I will happily run two
or three instances of lovable. I'll run
two or three instances of Bolt and I'll
run an instance or two of wind surf and
then something in cursor just because
I'm trying to work at the problem from
different angles. I'm not particularly
loyal to any given one of them. I just
want to see something come out at the
other side. If it works, great. I'm
loyal to the product and the outcome.
I'm I care about the outcome. That's
it. And the cost of refactoring and
restarting is essentially zero. So if I
want to scrap a project and start a new
instance in another tool cost me
nothing, almost nothing, right? like
it's just not that much. It's less than
my cable bill. Um, and so at that point,
why not restart? And we would never
dream of that level of of loyalty or
lack of loyalty in the 2010s. Like, we
wouldn't touch that lack of loyalty with
a 10-ft pole. SAS as a business model
was not built on that. And I am not the
kind of person who thinks that SAS is
dead. That's been sort of much bihood
and VCs talk about it all the time. I
actually don't think that's what this
means. Good products with good
distribution will build loyalty by
making it easy to either execute entire
data flows end to end so you don't need
to leave the tool or by making it really
really good at a particular piece of the
data flow you care about and then by
nailing the highways for data in and the
highways for data out. That's how you
win loyalty and win in SAS.
And so when you think about it, the the
long-term perspective here is that tools
that get good at data in and data out
are going to be like Amazon in the 2010s
that decided to care about returns.
Nobody in retail cared about returns
because everyone was like, "If you make
returns easier, you lose money." Why on
earth would we make returns easier?
That's freaking stupid. Well, it turns
out if you care about returns, you breed
long-term customer loyalty. It's goods
in and goods out, right? If you make it
easy for people to get their goods back
to you and get their money back out of
your system, they they feel more loyal.
The last time I had a good Amazon return
experience was this week. The last time
I had a good return experience anywhere
but Amazon was I can't
remember. And so the long-term loyalty
that getting it right breeds is
something that companies that looking
looking at data need to think about. If
you are building an AI powered
application, you need to think about
copy paste as a categorical fundamental
problem set. It is something that
matters differently now because of the
cost of
intelligence. I hope that made sense.
Friday afternoon. We're thinking outside
the box here and I hope you enjoy it.
Sharers.