{"slug": "how-the-ai-industry-forgot-two-of-its-best-ideas", "title": "How the AI industry forgot two of its best ideas", "summary": "The AI industry has abandoned two of its most valuable ideas—durable object orientation and reinforcement learning—not because they failed but because architectural shifts and incentive changes pushed them aside, according to a Fast Company analysis. The article compares this loss to Roman concrete techniques that were forgotten for centuries before being rediscovered by MIT researchers. Enterprise AI now risks repeating this pattern by neglecting proven approaches that could improve system robustness and learning efficiency.", "body_md": "[Technological progress](https://www.fastcompany.com/91577984/the-great-scramble-to-build-ai-compute-you-can-actually-own) is usually told as a story of accumulation. We imagine each generation inheriting everything the previous one learned and adding something new.\n\nReality is less tidy.\n\nSometimes technologies do not disappear because they failed. They disappear because the surrounding architecture changes, incentives shift, or a newer paradigm captures everyone’s attention. Years later, [the industry](https://www.fastcompany.com/91581867/the-ai-industry-is-rallying-around-open-models-is-it-more-than-talk) discovers that something valuable was left behind.\n\nRoman concrete offers a useful example. [Researchers studying ancient structures have reconstructed techniques that gave the material unusual durability and apparently allowed cracks to heal through reactions involving lime fragments](https://urldefense.proofpoint.com/v2/url?u=https-3A__news.mit.edu_2025_pompeii-2Doffers-2Dinsights-2Dancient-2Droman-2Dbuilding-2Dtechnology-2D1209&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=nEHT0D7VBBH5JNXQYSCIi6GKFfvqCj2asqBcYl3xUWM&e=). 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](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nature.com_articles_s41467-2D025-2D66634-2D7&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=Mq-JY2LUcELUcirfa9UTxhnXY0rCcFwWSVgLqISEziY&e=). MIT researchers have described [how “hot mixing” may explain these properties](https://urldefense.proofpoint.com/v2/url?u=https-3A__news.mit.edu_2023_roman-2Dconcrete-2Ddurability-2Dlime-2Dcasts-2D0106-3Futm-5Fsource-3Dchatgpt.com&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=DNifDWWf49Oepr7u0a6w4-kvUEqBGMCWTOlVBMz7cAw&e=).\n\n[Enterprise AI](https://www.fastcompany.com/91555415/real-reason-enterprise-ai-stuck) may be living through a similar episode. [The industry](https://www.fastcompany.com/section/artificial-intelligence) 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.\n\nObject-oriented programming was never just about classes and inheritance. Its fundamental intuition was more concrete: an object combines identity, state and behavior. It represents something that exists, remembers its condition and knows which operations can change it.\n\nThat 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.\n\nThen cloud computing happened, and something subtle changed.\n\nCloud-native systems are commonly designed to be stateless. Computation should not depend on the particular machine handling a request because that machine may disappear, be replaced or be joined by hundreds of identical instances. [AWS explicitly recommends removing state from individual components](https://urldefense.proofpoint.com/v2/url?u=https-3A__docs.aws.amazon.com_wellarchitected_2024-2D06-2D27_framework_rel-5Fmitigate-5Finteraction-5Ffailure-5Fstateless.html&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=dWl0qzzfYZ1Gn1WwNSaK_zCNXWMvGo5HV2Jrw0kq_u4&e=) so workloads can scale horizontally and recover from failure. [Microsoft similarly describes web-session data as ephemeral information that should be stored in an external cache or database rather than trusted to the application process itself](https://urldefense.proofpoint.com/v2/url?u=https-3A__learn.microsoft.com_en-2Dus_aspnet_core_fundamentals_app-2Dstate-3Fview-3Daspnetcore-2D10.0&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=Glzfhnpf256YYhzJvBAi5MZrnC0keChTNsMTZzvVenU&e=).\n\nThat is a perfectly sensible engineering trade-off. It is also a philosophical retreat.\n\nThe object may still exist in the source code, but its durable state no longer lives with its behavior. It is distributed across databases, caches, object stores, queues, event streams and orchestration systems. Every request has to reconstruct enough of the object’s reality to do something useful and then return its state to external infrastructure before the computation disappears.\n\nWe did not abolish object orientation. We weakened one of its deepest properties.\n\nOver 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.\n\nThat matters enormously for enterprise [AI](https://www.fastcompany.com/section/artificial-intelligence).\n\nAn 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. A customer must remain the same customer across interactions. A process must survive interruptions. A contract must carry its constraints. An agent must know not merely what happened in a conversation, but what changed in the operational world.\n\nToday’s systems often call this “memory.” But memory is not an object model. Memory can recover fragments of the past. An object model defines what exists and how it is allowed to change. Anthropic, for example, [presented Claude’s ability to draw on conversations stretching across weeks or months as an important advance](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.axios.com_2025_10_23_anthropic-2Dclaude-2Dmemory-2Dsubscribers&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=6ZsK03eSB0CFXhCvOEf9vKnjKpvq5kOOT7licZRUgY4&e=). And for a chatbot, it was. But from the perspective of enterprise software, the milestone sounds strangely modest.\n\n[A customer, a contract, a claim or a nine-month sales process should not remain coherent for a month](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.theverge.com_news_757743_anthropic-2Dclaude-2Dai-2Dsearch-2Dpast-2Dchats&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=4PbrprNKCrghAthMpeIAYPUiUxDo76lEajyG2fpEFug&e=). It should remain coherent for as long as it exists. The model does not need every previous token permanently inside its active context. The system underneath it needs durable identity and state, with the relevant context reconstructed whenever intelligence acts. A month of memory is impressive for a chatbot. But for enterprise software, it is an expiration date.\n\nEnterprise AI has extraordinary intelligence at the top and a fragmented operational reality underneath.\n\nThe other displaced idea is reinforcement learning.\n\nDeepMind’s [AlphaGo](https://urldefense.proofpoint.com/v2/url?u=https-3A__deepmind.google_research_alphago_&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=iC_-775aTbfQOBVK75lZqe1ileScgDeKMtnNSDbjIPY&e=) combined deep neural networks with reinforcement learning to defeat one of the world’s greatest Go players. Then, [AlphaZero went further, learning chess, shogi and Go through self-play rather than by copying human games. MuZero learned to plan without being given the underlying rules of the environment in advance](https://urldefense.proofpoint.com/v2/url?u=https-3A__deepmind.google_research_alphazero-2Dand-2Dmuzero_&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=Tm-UNQ1qu9WXG2rqibBP0pVKLkBISVM67sczebrXDJo&e=).\n\nThese systems demonstrated something larger than mastery of games. They showed how intelligence can emerge from a repeated loop: act, observe the result, compare it with an objective and adjust.\n\nThat is not merely pattern recognition. It is learning through consequence.\n\nThen came the transformer. The 2017 paper “[Attention Is All You Need](https://urldefense.proofpoint.com/v2/url?u=https-3A__arxiv.org_abs_1706.03762-3Futm-5Fsource-3Dchatgpt.com&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=xjvKd_Asd-N5YoiIzu97XbFKwjm_DtwTe2tSem0NTso2TsVX2V4buYsId0DEJeYI&s=UZh1CJ0tkNfsSGTjwihFNhqnnWL2csrkLffDHkRk2PU&e=)” introduced an architecture that was highly parallelizable and extraordinarily effective for sequence processing. It became the foundation of the generative AI wave and transformed natural-language processing.\n\nThe problem is not that transformers were a mistake. They were one of the most consequential advances in computing.\n\nThe problem is what happened to the industry’s center of gravity.\n\nPrediction 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.\n\nWe became extraordinarily good at producing plausible answers and strangely tolerant of systems that never discover whether those answers worked.\n\nThat is the hidden mismatch in enterprise AI. Companies do not merely need systems that generate language. They need systems that learn from consequences.\n\nA company is not a prompt or a chat session. It is a changing system of customers, contracts, products, employees, permissions, workflows, constraints and outcomes.\n\nTo improve such a system, AI needs two things.\n\nIt needs a durable world in which to act: entities that persist, processes that preserve state and relationships that remain coherent across time.\n\nAnd it needs a mechanism for learning what its actions accomplish: objectives, observations, feedback and the ability to adjust future behavior.\n\nRemove the first and the company becomes difficult to represent.\n\nRemove the second and the company becomes impossible to optimize continuously.\n\nThis helps explain why so much enterprise AI still feels like a clever interface sitting on top of a missing runtime. It can talk beautifully about the organization, but it cannot fully inhabit it in a durable, stateful and outcome-driven way.\n\nIt also explains why deployments become artisanal. Humans must reconstruct context, connect systems, define permissions, explain business objects, measure outcomes and redesign the loop for each use case. The model supplies intelligence, but the architecture needed to turn that intelligence into compounding organizational capability is assembled manually.\n\nThe result is an industry brilliant at demonstration and strangely poor at accumulation.\n\nThe next enterprise AI architecture will have to recover both ideas simultaneously.\n\nIt will need objects whose identity, state, relationships, permissions and behavior persist naturally, even when execution moves across machines and scales from ten interactions to ten million.\n\nAnd it will need reinforcement learning not merely as a predeployment training method, but as an operational mechanism: actions generating structured evidence, evidence connected to business outcomes, and outcomes improving subsequent actions.\n\nPersistent objects would give AI a stable enterprise world.\n\nReinforcement learning would give it a way to improve its behavior inside that world.\n\nTogether, they would turn software from a system that records what the company did into a system that helps the company learn what works.\n\nPerhaps that is the next enterprise AI breakthrough. Not another spectacular model capability, but the recovery of two principles the industry once understood and then allowed to drift apart: objects that live, and systems that learn from consequences.", "url": "https://wpnews.pro/news/how-the-ai-industry-forgot-two-of-its-best-ideas", "canonical_source": "https://www.fastcompany.com/91577826/how-ai-industry-forgot-two-best-ideas", "published_at": "2026-07-30 16:33:21+00:00", "updated_at": "2026-07-30 17:53:59.431693+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-ethics"], "entities": ["Fast Company", "MIT", "AWS", "Microsoft"], "alternates": {"html": "https://wpnews.pro/news/how-the-ai-industry-forgot-two-of-its-best-ideas", "markdown": "https://wpnews.pro/news/how-the-ai-industry-forgot-two-of-its-best-ideas.md", "text": "https://wpnews.pro/news/how-the-ai-industry-forgot-two-of-its-best-ideas.txt", "jsonld": "https://wpnews.pro/news/how-the-ai-industry-forgot-two-of-its-best-ideas.jsonld"}}