Dustin Snoap
to you Every major technological revolution has, at some point, required a second revolution to keep going. Steam power needed metallurgy to improve before
Every major technological revolution has, at some point, required a second revolution to keep going.
Steam power needed metallurgy to improve before it could scale. The internal combustion engine needed oil refining. The internet needed the browser, then broadband, then the smartphone. In each case, the first breakthrough created enormous momentum, hit a wall, and then waited for the next breakthrough to arrive and extend the curve. The waiting part is rarely included in the retrospective.
AI is at the wall now.
The scaling era, the period where you could reliably improve a model by making it bigger and feeding it more data, is ending as a primary strategy. The evidence is not subtle. The researchers who were saying this three years ago and not being invited to panels are now being quietly cited by the people who were running the panels. Simply adding more compute is no longer producing the step-change improvements it once did.
The industry knows this. Which is why it has identified, with admirable thoroughness, approximately seventeen potential next breakthroughs.
Test-time compute lets models think longer at inference rather than training them on more data. OpenAI's reasoning models have shown genuine gains here. Mixture-of-experts architectures activate only the relevant parts of a model for a given task, making large models more efficient. Synthetic data pipelines have AI generate its own training data, which introduces problems that researchers are actively and not entirely successfully working through. Neuromorphic hardware, chips designed to more closely resemble biological neural networks, exists in research contexts and remains several generations from meaningful deployment.
Each of these is real. None of them are ready. And the history of necessary breakthroughs offers no particular comfort about timelines. They arrive when the underlying science is ready, which is a different schedule from the one that has been laminated and placed on the conference room wall.
The next inflection point may be three years away. It may be ten. It may involve an approach nobody has published yet. The people who tell you they know which of these is true are either more confident than the evidence warrants, or they are trying to sell you something, or, in the most interesting cases, both.
What this means practically is that the AI available in two or three years will be meaningfully better than today and noticeably less dramatically different than the narrative currently implies. The era of transformations obvious enough to feel like a different product every eighteen months may be pausing.
The next miracle is not cancelled. It is simply running a little late and leaving messages asking for an ETA on read.