The four eras of AI, and the one now starting

Four, counting generously, and the fifth has started without anyone finding it a name that sticks. The division is arbitrary in places, but it illuminates something useful: each era moved what we asked of the machine, never simply how much power it had.

The division, and what is wrong with it

Symbolic AI, 1950s to 1980s. Knowledge was written by hand, as rules and expert systems. The machine discovered nothing, it applied what a human had formalised, and it failed the moment the world stepped outside the rule.

Machine learning, 1980s to 2014. The reversal consists of no longer writing the rule but letting it emerge from data. Regression, decision trees, random forests: tools still in use, and often still the best fit when the problem is well posed.

Deep learning, 2015 to 2022. Deep networks and large datasets finally make vision, speech and translation affordable. This is the era where the machine stops needing you to describe the relevant features, because it finds them itself.

Generative since 2022, extended by agentic since 2024. The model no longer classifies, it produces, and then it acts. The jump comes not from a new architecture but from scale, and from the discovery that one model can serve tasks it was never shown.

The weakness of this story is that it reads as a succession when it is a stack. A company in 2026 runs logistic regression under its scoring, business rules written twelve years ago in its payroll, a vision model over its documents and an agent on top of all of it. Nothing was replaced, everything was layered, and it is usually the oldest layer that decides the outcome.

Why it feels like a plateau

The feeling has been widespread for a year and it rests on a measure that stopped working. While progress came from pre-training, watching model size was enough to follow along: more parameters, more data, better results.

That source of gains is reaching the end of its easy dividends. Scaling laws are not dead, but differentiation has moved to what is done to the model after training, namely mid-training and reinforcement learning. Since that work does not show up on a spec sheet, somebody comparing parameter counts reasonably concludes that nothing is moving.

It is the same optical illusion as with prices. Token cost collapses while bills rise, because the saving funds quality rather than an economy. Here, size plateaus while capability advances, because the work moved somewhere nobody is counting.

The three efforts that define what comes next

What follows agentic AI has no settled name yet, but the directions are identifiable and they do not address the same problem.

Continual learning. A current system retains nothing from one conversation to the next unless you feed it back in, and retraining it on new material makes it forget the old. Solving that catastrophic forgetting would change daily use more than ten points on a benchmark, because an agent that learns from its corrections stops repeating the same mistake.

World models. Rather than predicting the next word, the aim is to build an internal representation of an environment and simulate what would happen in it. It is the most serious avenue for anything requiring long-range planning, and it is also the one that directly challenges the dominance of large language models.

Hybrid architectures. The pure transformer is expensive in compute and in energy, and infrastructure is starting to hit genuine power constraints. Combinations mixing several attention mechanisms or several model families are gaining ground for that economic reason before performance justifies them.

The debt each era left behind

Here is the angle you find nowhere, and it explains why information systems are in the state they are in. Every wave left behind a particular kind of debt, and those debts now coexist inside the same company.

The symbolic era left rules nobody dares change. They are written somewhere, they work, the author left long ago, and the only documentation is the observed behaviour. Any long-established firm has some in its margin calculation or its billing rules.

Machine learning left models nobody can explain. The scoring runs, it produces reasonable results on average, and when an edge case surfaces there is no simple way to find out why that particular profile was ranked as it was.

Deep learning left data pipelines that are expensive to maintain, whose real cost only appears when a source changes format.

The generative era left instructions scattered everywhere: in shared documents, in different tools, in people’s heads. Nobody knows any more which version is authoritative or who wrote it.

The agentic era leaves a debt of a new kind, and it is the one to watch now because it is accumulating as we speak: permissions granted that nobody revokes. An access opened for a three-week trial, a token created for a test, an integration wired up by somebody who has since left. Unlike the previous four, this debt is not merely expensive to maintain: it is exploitable, and it grows with every connection added.

The lesson common to all five is the same, and it is worth more than any prediction about the next era. What costs money ten years later is never the technology chosen: it is whatever was put in place without writing down why, and that nobody dares touch afterwards.

One consolation for anyone feeling behind: none of these eras arrived on schedule, and the people who predicted their timing were wrong in both directions. Symbolic AI was supposed to deliver general intelligence by 1980. Deep learning was widely dismissed as a dead end five years before it worked. The honest position on what comes after agentic systems is that the directions are visible and the dates are not.

What it changes for a buyer rather than a builder

Three practical consequences, none of which requires following the research.

The first is that an agent which learns nothing from one week to the next will stay limited, whatever the quality of the model behind it. So the question for a vendor is not which model they use, but what their system retains from your corrections and where that memory lives. A tool that makes you repeat the same instruction every Monday does not have a memory, it has a history.

The second is that the boundary between workflow and agent will move, in the agent’s direction. Tasks too unpredictable to script and too risky to delegate today will gradually cross over, which changes nothing about the rule that separates them but shifts the line every year.

The third is that waiting for the next era is pointless. Each of the four preceding ones took between five and fifteen years to produce industrial effects, and the companies that got something out of them were the ones already using the previous one. The cost of starting is not in the technology, it is in the data to put in order and the learning of delegation, and neither of those shortens because the model improves.