Takes "one version for everyone" apart as a cost calculation, then hands you the architecture and verification problems of a version per user.
The Pipeline Is Dead - Iris ten Teije, Sky Valley Ambient Computing · Iris ten Teije
20 min total·Actually worth watching closely: ~5 min·2 must-watch clips
- 0:00 – 3:00Listen
What the distribution stack is actually solving
The opening strings pipelines, package registries and App Store reviews into one thing: their entire job is getting a frozen artifact safely onto the machine where it runs. What she wants to ask is why that stack exists at all.
All that release infrastructure serves the frozen-artifact premise, not software quality itself.
Just the speaker talking over a text-only slide — nothing is lost by looking away, so this stretch is fine to listen to while doing something else.▶ Jump to 0:00Speaker · Iris ten Teije - 3:00 – 7:01Listen
One version for everyone is a cost calculation, not a fact
She points out that "everyone runs the same version" never had to win an argument — it was a by-product of the constraint that a version per user meant forking the codebase and hand-maintaining both. Now that the cost of producing a correct, scoped change is collapsing towards zero, that premise no longer holds.
The pipeline didn't break; the assumption it was built for went away.
A step-by-step argument with no charts and no demo on screen — following the words is far more useful than watching.▶ Jump to 3:00Speaker · Iris ten Teije - 7:01 – 9:20Listen
The demand for personalization was always there
Decades of old evidence that the demand already existed: the professional services industry enterprise software grew around it, engineers' own dot files, millions of people building their own programs in Excel. Feature flags and A/B testing were the predecessors of divergence, but they could only sort users into buckets declared in advance.
It's not that users didn't want software of their own — it just didn't pay off for most software.
Dense with examples, but all of it spoken aloud; the screen holds a few words at most, so hearing it is enough.▶ Jump to 7:01Speaker · Iris ten Teije - 9:20 – 14:05Listen
Stem plus divergences: the agent as runtime
Here is the core architecture: deploy one canonical stem, and every user runs their own divergence of it — same origin, individually adapted. Divergences are isolated and individually reversible, so a change blows up one user's context at most. Developers keep a veto and can declare areas like auth or payments permanently off limits. She then grounds it all in a CRM example.
The danger isn't AI-written code, it's divergence with no boundaries; draw the boundaries and the blast radius is a single user.
The place you most want a diagram is exactly where there isn't one — no architecture slide, no demo, the stem-and-divergence relationship is entirely spoken. Slow down here, and replay the examples if you need to.▶ Jump to 9:20Speaker · Iris ten Teije - 14:06 – 16:20Listen
Attribution once there is no single version
Once no single artifact exists, "what is this user running" goes from reading a version number to running a graph query. Every divergence has to be immutable, inspectable and attributable, and testing has to reason about the stem plus every possible divergence of it.
Provenance stops being metadata and becomes the foundation of the system.
The most engineering-heavy stretch, delivered fast with no code or UI on screen; the visuals can't help you, so replay it if you care about feasibility.▶ Jump to 14:06Speaker · Iris ten Teije - 16:20 – 18:25Listen
Correct doesn't mean it should exist
She calls getting a model to generate code the easy 80% anyone can do. The hard part is measurement: a change can be perfectly correct and working and still be one that shouldn't exist, and whether the adaptation delivered a real uplift has to be judged against each product's own goal metrics.
Correctness is the floor; whether the change was worth having is the question you actually have to answer.
The two minutes that need the most attention — turns of argument come thick and fast with no visual support at all. Not background listening.▶ Jump to 16:20Speaker · Iris ten Teije - 18:25 – 19:48Listen
The next twenty years
The close is about how updates propagate: not everyone runs the same commit, everyone converges on the same goal through their own path — you merge intent and outcome, not code. Twenty years of practice went into shipping one version well to everyone; what comes next is shipping the right version to each person, on isolation and provenance.
The final bar is trust — the system has to be legible and reliable enough that humans choose to step back from the loop.
A standard closing stretch: the speaker wrapping up her argument for the audience. Whether you watch the screen makes no difference.▶ Jump to 18:25Speaker · Iris ten Teije