The Pipeline Is Dead

As we're online today, you can't raise your hands, but show's node on your screen. How many of you have spent a significant part of your career making mo-- making software move from one computer to another?

CI pipelines, package re- registries, container images, app store reviews. That entire stack exists to solve exactly one problem. Get a frozen artifact, that code from the machine where it was built to the machine where it runs safely, reproducibly, once.

And here's what I think everyone is missing. That entire stack is built around one idea that it's so old-- that is so old that we stopped seeing it as a choice.

One version of your software for everyone. We've shipped it that way for so long that almost nobody asks why anymore. I'm Iris. I'm one of the co-founders of Differ.

My co-founder, Noam, was the first engineer at JFrog. He helped build the pipeline that I'm about to tell you is dying. I came at it from the other end.

I spent a decade in Fintech. I was early at a digital bank that we scaled and exited, shipping software in the environments least willing to tolerate what I'm about to propose.

So I totally get any reservations, and I've had them myself as well.

However, I'm telling you, it's coming anyway, and that's not a warning. It's the best thing that's happened to software in a long time.

Every piece of distribution infrastructure that you've touched encodes the same assumption.

Software is produced in one place, it runs in another, and the thing in the middle, the artifact, is frozen.

And that assumption was correct. For decades, it was just true.

Why? Because producing a correct change was expensive. It took skilled humans hours or days, so you did it rarely. It was a central event. You verified it and froze it,

and you shipped that frozen thing to everyone.

So the one-way pipeline isn't arbitrary. It's the direct consequence of production of software being expensive and risky.

And of course, the frozen artifact has some advantages. You get reproce-- re- reproducibility, reviewability, rollback,

and every guarantee that we lean on in production flows from one fact. There is one artifact, and it doesn't change after we ship it.

That's the deal. One version for everyone, frozen.

We got reliability, but the price was that the software couldn't really be for anyone in particular, and nobody ever made that decision. There was never a meeting where someone put one version for everyone versus a version for each person and picked the first one.

And it wasn't because the second option, if w-- a version for each person, was worse. It just wasn't an option. Giving two users different software meant forking the code base and hand-maintaining both.

A version per user at any real scale wasn't really a viable option, and one version where everyone never had to win an argument. It was just how it was.

It was the only shape software could take. And we started treating it as a fact about software, like gravity, like something that, that is just true. But it was never really a fact about software.

It was a fact about cost and budget and the economics, and that cost just changed. Now, you might, you, you might expect me to go into AI can code, but everyone has already said it.

Uh, and that's table stakes, and to me, that's not really the interesting part. The interesting part is where and how cheap the cost of producing a correct and scoped change is collapsing towards zero.

And just as importantly, the production of the software no longer has to happen in one place up front before anyone runs it. Part of it can be run on a server, part of it on a client, part in the user's live session.

And as each step stops being a decision you freeze at build time and, and now starts becoming more of a real-time one, that also means that each p- each, each p- piece can be placed wherever it makes most sense, including right in front of the user in their context.

And so this whole one-way pipeline existed because making software was the expensive, central, and rare event, and running it was cheap.

So you separated development from distribution. But now, as making a change becomes as cheap as running one, and it can happen in the same place as, um, as where you run it, the reason to separate them is dissolving.

So far, I've talked about the supply side. Producing code has become cheap, easy. We can now make a change on a per-user basis. But I also want to address the demand side.

People have always wanted software that fits them, and we have decades of proof for that.

It just wasn't really possible for most types of software because of cost.

To start with one example, the four deployed engineer. Enterprise software has always had a line item called professional services, and a whole industry exists around that.

If you're a big client of a company like Salesforce, you probably have consultants that are helping you implement custom setups, configurations. You have an engineer living in your Slack channel, and it's not that smaller customers

can't benefit from this type of customization. It just didn't make sense financially until today. To give another example that might resonate with you as engineers, think about your dot files, your editor config, key bindings.

You rebuild every tool you touch into your tool by hand on every machine. It's another example of a demand for personalized software.

And lastly, Excel, the most successful business software ever created. And Excel isn't really a static program.

It's millions of people that all built their own programs on top of it.

So as it's clear, give people the power to make software, to, to make their software theirs, and they take it. Seen on the social feed that per user wins on-- wins from one size fits all on every metric that mattered.

This is more of an example of content versus software. But now that we have, uh, better coding again-- Now that we have coding agents and better coding agents, we can move this also to the software layer.

So my point is none of this is really new demand. We've seen it for decades. There have been predecessors like feature flags, segmentation, A/B testing. There's an enormous industry around, uh, a-a-around this, and we've been trying to make software diverge for many years, but

we got forced into a specific shape, that of creating buckets and segments that you declare in advance.

And now, for the first time, we can make software truly adaptive. So to get back to the title of this talk, when the agent is the runtime,

when the thing that runs your software can also modify it, development and distribution stop being two phases. The boundary blurs, and it's gone.

And con-- the shape that we bet on at Differ is that instead of one code base gated by flags and shipped to everyone, you deploy one canonical stem, and every user runs their own divergence of it.

Same origin, but individually adapted live. It's going from the least worst version for everyone to the best version for anyone.

Now, if you're an infrastructure person, you might be a little worried. Your stomach might be turning because I have just deleted the frozen artifact, and the frozen artifact was what was holding up the entire, the, the, the entire building.

And I do-- we do get these objections, for example, from a call that I had with a CTO recently who is right to be skeptical. What he said roughly was, "I can already barely reason about one AI-generated code base, and you want me to run millions of these?

You're not describing a capability. You're describing my worst problem multiplied." And if that's your reaction, that's not surprising. It's the right instinct, but perhaps aimed at the wrong target.

Here's the distinction that we make. The brittleness that you are picturing is a specific type of failure mode. It's an unmanaged divergence inside a single artifact. Thousand-line files, everything can touch everything else, no boundaries.

And that's not nec-- no, that's not necessarily brittle because it's AI-generated. It's brittle because there's no structure separating things.

In our ver-- in our vision, uh, we've, we're thinking about per-user divergences and done right, that is the opposite of that. You've got a stem plus divergences. The divergences are bounded, isolated, and individually reversible.

A bad variant can silently corrupt the stem or reach another user, which means that the blast radius of a change is in the system, it's one context, and any single divergence can roll back live with no deploy.

So the answer to the previous objection isn't, "Trust us, AI is good at coordination or at coding." The honest answer is you're brittle because there's a tangled artifact with no boundaries.

In our case, we don't ship you a thousand tangled artifacts. We ship one stem and bounded divergences, each isolated, each reversible. The thing that you're afraid of is the thing that this architecture exists to prevent.

And as a developer, you can also set controls and boundaries, what can and cannot be adapted. To give a small example, um, we can, gonna have a scenario where we've got a form, and the form can be adapted in order to improve conversion rate, for example.

However, as a developer, you can always indicate that specific fields can never be dropped or parts of your app, like auth or payments should always be out of, um, uh, uh, should always be off-limits for any sort of adaptation.

To give another example to make adaptive software a bit more concrete,

h- Uh, think of, for example, um, a CRM.

In this case, uh, we've got an investor who's using a CRM, whereas the CRM was mostly built with a salesperson in mind. And as an investor, you might use it slightly differently.

So this investor, she often logs founder intros, and she's always logging who introed her to which deal. So the system observes that and creates a, um, cre-creates a intro path.

The system can also observe that she's always skipping specific fields. She never fills them out. So over time, the system learns and doesn't surface those fields, but instead it surfaces fields that she, that she cares about more.

Another example can be the fact that she is always, um, always checking specific types of deals or founders, and it doesn't exactly follow the prioritization that the system sets by default.

So again, can, can we learn and make that smarter so that the information that the user cares about is surfaced first? Not only can the system observe, the idea is also that the user can proactively request changes, and as long as these changes are within the boundaries that the developer sets, and as long as they are within

the spirit of the software, within the purpose of what the software was originally made for, it can be implemented without having to go back to the developer.

So you can imagine for a, for a horizontal SaaS like a CRM,

you can address a much wider number of customer personas without increasing your R&D spend.

Now, of course, there is... there are many hard parts when it comes to bringing this to life.

And I'm going to discuss a couple of hard challenges and prob-- and problems that we're, uh, that we're working on as we're making adaptive software real at scale. We haven't solved all of them, but we have a v- we have a point of view on each of them, and exa-- it's exactly what we're working on day in,

day out. So what changes when there's no single artifact?

Firstly, the source of truth. So wh-when there's no single artifact, you can wonder, what is the software?

In our case, we consider the software to be the stem plus the, plus all the mutual immutable divergences.

But of course, that creates a lineage problem. What is this user running and why? That now becomes more of a graph query versus a version number.

A bug report describes a program that exists for that specific user. How do you de-debug or inspect that? The answer is, every divergence is immutable, inspectable, attributable. And

we also need to trace any version back to this signal led to this specific recommendation and this exact adaptation.

That's one of the parts that we're working on.

Second is correctness. Like, how do you test that the, um, how do you test that the code change that you've just implemented works, that the UI is correct? And yeah, testing at a much larger scale of users means that you need to reason about the stem and also every possible divergence of it.

And then desirability, because perhaps you made a code change, it's perfectly correct and working,

but you also need to know that whether, whether this was desirable, was it actually a good change? Because anyone can make a code change now, but the hard part is knowing whether you actually found

an improvement, an uplift. And yeah, that, that is something that is extremely important to, to keep track of and to measure and to consider what the goals for the company are.

And this is not going to be the same for every single piece of software. In some cases, it might be retention or less churn or lowering this number of support tickets.

So it's extremely important to keep track of what are, what are the goals that we're chasing of adaptive software, and do the adaptations reach to improve, improve, improve the metrics that matter?

Next is autonomy versus control. And the conservative answer here would be start with just recommendations and don't make any autonomous changes. And that's not a wrong strategy, but in our case, it's not really our vision.

The vision is a system that understands the user well enough to act without asking the developer for permission every time first.

So for us, the challenge isn't building more control, it's winning enough trust that you don't have to. And it's a hard problem, but also, uh, also a very interesting one.

And, you know, how do you make a system good enough, legible enough, reliable enough that humans in a loop choose to step back? And that's certainly what we are building towards.

And lastly, coordination. Everyone on their own version, how do you push new updates? How do, how do changes propagate to, like, a million different versions? And this is one of the challenges that we've been thinking hardest about.

The, uh... and the answer that we keep coming back to is, "Don't merge code, merge intent, merge outcome," which means not everyone has to run the same commit or the ex-exact same piece of code, but everyone converges on the same goal through their own path.

And if it wasn't already clear, the challenges that I just addre-a-addressed are the hard challenges. Generation has become easy, and I would say that's actually the easy eighty percent.

Calling a model to write some code is something that everyone can do. The other part, observability, validation, coordination, that is the entire business.

Anyone can call an LM, but the substrate, the stem plus diver- plus divergences, provenance, validation, that is really the hard part and something that we're working on every single day.

To close off with, when I started out in fintech, a bank with no branches sounded reckless.

A decade later, the, the branch is the weird part.

My co-founder, Noam, used to fight with engineers about why you need builds and CI.

We've seen this, we've seen these shifts before, and adaptive software is the same type of shift.

It doesn't feel obvious until it does. To go back to the start of the talk, the pipeline didn't fail because it didn't work anymore, but the constraint it was built for went away.

The assumption underneath it that software is expensive, so we need to freeze it and ship it once, that assumption stopped being true.

And when making software gets as cheap as running it, the line between distribution and development isn't a line anymore.

We spent twenty years getting good at shipping one version for everyone. The next twenty are about shipping the right version to anyone with the isolation and provenance that makes it safe instead of terrifying.

I'm Iris, this is Differ, and that's what we're building. Thank you for watching.