Assume It Works: Why the Steam Engine Only Paid Off in New Factories
Gavin Lew has spent three decades in this field and taught some of the people sitting in the room twenty years ago. His Nivan Live session moved from the past to the present through a single historical observation about the industrial revolution, then applied it to autonomous vehicles, fleet operations and the enormous invisible infrastructure that decides whether an innovation survives its own success.
Gavin Lew at Nivan 2026
The lesson from a mentor
Gavin credited Arnie Lund, his mentor at Ameritech, with the framing that organised the talk. Technology innovations have happened before, and the pattern of what makes them succeed repeats. Ask what happened last time and you get a usable prediction about this time. That is a different discipline from forecasting, and it is available to anyone willing to look backwards before looking forwards.
Steam engines in old factories
When steam engines were installed in existing factories during the industrial revolution, productivity did not improve. The gains only arrived when new factories were built assuming the engine, designed around it rather than retrofitted to accommodate it. The innovation was never the bottleneck. The design of everything around it was.
The instruction that follows
Gavin's translation for the present is deliberately provocative. Assume it works. Do not sit back and evaluate whether the technology will mature, because it is maturing faster than the evaluation cycle. Anticipate the experience and design for it now. The alternative is to spend the window watching, and to inherit a retrofit.
A fraud algorithm and thirty years of interface
In 1992 his college roommate wrote a credit card fraud detection algorithm. It ingested data and produced a coefficient, and above a threshold the card was cut. That year, in Barcelona for the Olympics, Gavin had one credit card and the system cut it. When he interviewed his roommate for a book decades later, the underlying algorithm was largely the same, with more data and more speed. What changed entirely was the designed interaction: a notification on a phone and a single tap to confirm it was you.
The pillars underneath any AI experience
That story is the argument in miniature. Understand the context and what data the system is given, define the designed interaction, and only then does trust and a humanised experience become possible. Gavin noted those pillars hold whether the technology is a fraud model, generative AI, or whatever arrives next. If you do not design it and simply let it happen, the productivity will not materialise.
People will walk away and not come back
Expectations are now high enough that a poor first encounter is often the last one. Gavin was blunt that if you build it they may come, and they may also leave permanently. That raises the cost of shipping an undesigned experience in a way that was not true when a category was new and users were forgiving.
Cameras, lidar and stopping the car
On autonomous vehicles he was even handed and specific. The vision of dropping a camera equipped car anywhere and letting it drive is admirable, but it was promised a decade ago and remains at level two, with a small Austin fleet, incidents and data suggesting significantly worse performance than a human driver. Waymo, using lidar with maps and a geofence, is already at level four. To the objection that cameras and lidar might disagree, his answer was that the design decision is obvious: stop the car.
The end to end experience nobody sees
His teams researched how a blind rider confirms which vehicle is theirs, how a deaf rider is communicated with, how someone with mobility needs gets a wheelchair into the vehicle, and what happens when a rider is dropped five blocks from where they should have been. None of that is the innovation. All of it determines whether the innovation is usable by everyone.
What growth does to an interface
Waymo went from ten thousand rides a week in 2023 to two hundred fifty thousand a week last year, more than a million a month, in under two years. Gavin used an air traffic control analogy. An interface showing four vehicles and highlighting the one needing assistance is a good interface. The same interface at fleet scale makes finding that one vehicle and routing it to the right depot nearly impossible without memorised keystrokes.
The ecosystem is the design problem
Behind the ride sits a system almost nobody designs: work orders, depot procedures consistent across different mechanics, camera recalibration after a minor collision, remote operation interfaces, call centre tools for supporting someone in a stopped vehicle, a scan for a phone left behind, and a process for capturing information for law enforcement when a person attacks a vehicle with a bat or an axe. That last example was not hypothetical.
Designing for the users you are not
Gavin situated himself deliberately as a cassette tape user who reached college before email. Today's children assume touch and swipe. The challenge he set was to make the innovation accessible across all of those starting points, which requires research and, more than anything, requires not waiting. The ecosystem is what supports an innovation, and the ecosystem is where the present work is.
