Arvind Narayanan at ICML 2026: AI adaptation is the slow work of decades

Arvind Narayanan gave a keynote at the International Conference on Machine Learning in Seoul last week that was widely discussed for its timing and its thesis. Titled “What will be left for us to work on?” and published on his Substack on July 13, the talk directly addressed the anxiety spreading through the AI research community as models grow more capable. Narayanan, a Princeton professor who leads a team working on AI agent evaluation, made three arguments. The first: the “AI as Normal Technology” framework he co-developed with Sayash Kapoor is correct and useful for thinking about AI’s medium-term impacts. The second: recursive self-improvement, while worth taking seriously, does not imply a sudden, total displacement of human work. The third: jobs of the future will be radically different, but adaptation will take decades.

The talk is worth reading closely, not because its claims are novel — the “AI as Normal Technology” essay has been circulating since 2024 — but because of where and when it landed. ICML 2026 is the premier machine learning research conference. The audience was the people building the models. And Narayanan used the platform to argue that the most important work ahead is not another scaling law or training run, but organizational redesign.

Narayanan walked through a four-stage framework borrowed from the history of electricity. Invention is followed by innovation (building the appliances), then diffusion (adoption), then adaptation or structural transformation. He argued that adaptation is the slowest phase and that in most fields, including software engineering, it has not really started. The evidence he cited: the gap between what AI can do in benchmarks and what it actually does in deployment. His team looked at 10 to 12 reliability metrics and clustered them into four dimensions — consistency, robustness, calibration, operational safety. They found that agent benchmarks measure accuracy but not whether failures are predictable, recoverable, or catastrophic.

This is the kind of detail that makes the talk more than a pep talk. Narayanan is not saying “don’t worry.” He is saying that the bottleneck is not capability but reliability, and that reliability is an organizational problem, not just a model problem. A 70% accuracy number tells you nothing about whether the agent can be deployed on a fixed subset of tasks or whether it fails unpredictably 30% of the time. That distinction is everything for production systems.

The electricity analogy is the core of the argument. Factory owners initially tried to replace steam boilers with electric generators as a drop-in replacement. It did not work. What took 40 years was the recognition that electricity is portable and allows reorganization around the assembly line — new layouts, new labor laws, new training pipelines. Narayanan’s claim is that AI will follow the same pattern. A couple of decades from now, work will be fundamentally reorganized. But that reorganization is not the job of AI companies. It is the job of every organization that uses AI.

This is where the talk gets interesting for the AI community itself. Narayanan noted that the question of adaptation is hitting software engineering and AI research first. “If we simply roll over and accept that a lot of our work will be done by AI in the future, instead of setting clear boundaries, I think it will lead to an even stronger political backlash against AI than what we are seeing today,” he said. That is a direct challenge to the “permanent underclass” framing that has circulated in Silicon Valley.

Narayanan offered a speculative vision of what adaptation might look like in software engineering. If coding agents can create ten-million-line code bases without bugs and security vulnerabilities, he argued, it no longer makes sense to build one piece of software for billions of people. The logical endpoint is extreme personalization — software tailored to each individual or team. That would shift software development in-house, away from software companies and toward the teams that actually use the software. “Do we even need software companies anymore?” he asked.

The speculation is useful precisely because it is concrete. It gives the audience something to argue with. But the deeper point is about time horizon. Narayanan is saying that the adaptation phase takes decades. The AI industry operates on a quarterly cycle. The two timeframes are mismatched.

There is a tension in the talk that Narayanan acknowledged but did not resolve. He argued that the “AI as Normal Technology” framework is correct unless there is a discontinuity from recursive self-improvement. He then argued that recursive self-improvement is not worth losing sleep over. That is a bet, not a proof. The history of technology is full of discontinuities that looked unlikely until they happened. But Narayanan’s bet is at least grounded in a specific mechanism: reliability bottlenecks that are not solved by scaling alone.

The most important line in the talk may be this: “When we look at past technologies, this kind of change tends to be very slow.” That is the observation that should stay with the reader. The AI industry has spent five years racing to build more capable models. The next five years will be about figuring out what to do with them. That work is not glamorous. It is organizational, regulatory, and cultural. And it has barely begun.