Artificial intelligence has had many moments. This one feels different, and it is worth being precise about why, because the reason explains what comes next.
AI is not new. For decades it meant narrow systems built for one job, a spam filter here, a recommendation engine there, each trained from scratch on its own data. They were useful, but brittle. Step outside the task they were built for and they fell apart. What changed in the last few years is not that AI got a little better. It changed shape.
The shift is from building a separate model for every task to training one large model on a vast sweep of text, code, and images, then pointing it at many tasks. These are called foundation models. The same system can draft an email, summarize a contract, write working code, and answer a question about biology, without being purpose-built for any of them. That generality is the break from the past. It is why one tool can suddenly help across a whole company instead of one corner of it.
Most of today's progress traces to an idea introduced in 2017 called the transformer. Its key trick is attention, a way for a model to weigh which parts of its input matter most for what it is generating. That sounds modest. In practice it let models learn from enormous amounts of data efficiently, and it scaled gracefully as researchers added more. Nearly every system you have heard of since is built on this foundation. One good idea, applied at scale, reset the field.
The surprising lesson of the last few years is that bigger models trained on more data did not just get incrementally better. They picked up abilities their makers did not design in, from translating languages they were barely trained on to solving multi-step problems. Capability emerged from scale. That is unusual in engineering, where you normally get exactly what you build, and it is why the pace has felt so fast. It also means the trajectory is not finished.
The final difference is the one that matters most for business. Earlier AI breakthroughs lived in research papers. This one shipped. Teams now build real products on top of these models, and the value shows up quickly enough that adoption is compounding rather than stalling in pilots. McKinsey and others have tracked the same pattern: use is widespread, and the organizations that commit see real returns. The question for leaders has shifted from whether this is real to where it pays off for them.
It is worth staying grounded. These systems are not conscious, and they make confident mistakes. They are tools, extraordinary ones, that need judgment around them. But the combination of generality, scale, and production readiness is genuinely new, and it is durable. Treating this moment as a passing trend is the one mistake that ages badly. The smarter read is that a general-purpose technology has arrived, and the work now is learning to use it well.
Jason Kumpf has had a front-row seat to AI moving from promise to production. He is Head of US Revenue at Razorpay, a board advisor, angel investor, and speaker. More about Jason.