How the AI industry forgot two of its best ideas and what it can learn from Roman concrete.

07.30.2026

Author - Enrique Dans

Researchers studying ancient structures have reconstructed techniques that gave the material unusual durability and apparently allowed cracks to heal through reactions involving lime fragments. The knowledge was not disproved. It simply ceased to be part of ordinary construction practice and had to be recovered centuries later through archaeology and materials science.

Enterprise AI may be living through a similar episode. The industry has not merely failed to invent the right architecture. It has quietly pushed two of its most useful ideas away from the center: durable object orientation and reinforcement learning.

Object-oriented programming was about something that exists, remembers its condition and knows which operations can change it. That proved particularly suitable for enterprise software. A customer, contract, account, order or insurance claim was not simply a row of data. It was an entity with identity, relationships, state and permitted behavior.

Then cloud computing happened, and something subtle changed, as we did not abolish object orientation but weakened one of its deepest properties.

Over time, we compensated with object-relational mappers, session stores, event sourcing, message brokers, distributed caches, workflow engines and a small mountain of glue code. None of these technologies is misguided. They solve real problems created by scale and distribution. But taken together, they reveal what happened: persistence stopped being a natural property of the computational object and became an engineering problem surrounding it.

That matters enormously for enterprise AI. An AI system acting on a company needs more than access to documents and APIs. It needs durable entities whose identities, states, relationships, permissions and valid transitions remain coherent over time, as long as it exists.

The other displaced idea is reinforcement learning. DeepMind’s AlphaGo combined deep neural networks with reinforcement learning to defeat one of the world’s greatest Go players. These systems showed how intelligence can emerge from a repeated loop: act, observe the result, compare it with an objective and adjust.

Then came the transformer and prediction became the dominant paradigm. Reinforcement learning did not vanish, but it was increasingly used around models: for fine-tuning, alignment, robotics or isolated optimization problems. The grander idea:  that deployed systems should continuously improve by connecting their actions to real outcomes , became secondary to the spectacular ability to generate language.

We became extraordinarily good at producing plausible answers and strangely tolerant of systems that never discover whether those answers worked.

That is the hidden mismatch in enterprise AI. Companies do not merely need systems that generate language. They need systems that learn from consequences.