Evan Lee Salim
Technological progress is traditionally framed as a story of linear accumulation. We tend to view history as a cumulative ledger, where each generation inherits the hard-won wisdom of its predecessors, polishing those tools and stacking new innovations atop the foundations of the past. We imagine a staircase to the future; however, the reality of software development—and specifically the current trajectory of Enterprise AI—is looking increasingly like a cycle of forgetting.
As the industry grapples with the limitations of large language models (LLMs) and the frustrations of enterprise integration, a quiet, uncomfortable realization is settling in: the industry has not merely failed to invent the right architecture. It has, through a mix of hype-driven myopia and architectural convenience, pushed two of its most potent historical lessons to the periphery: durable object orientation and reinforcement learning.
The Myth of the Tabula Rasa
The narrative of "AI disruption" often treats the history of computing as a pre-modern era—a dark age that ended the moment the Transformer architecture was introduced in 2017. This perspective posits that before the era of generative AI, software was static, brittle, and incapable of nuance. By framing the past as obsolete, the industry has granted itself permission to abandon fundamental principles of software engineering that once made complex systems manageable.
In our scramble to build bigger, faster, and more expensive compute environments, we have discarded the very structures designed to manage state, identity, and behavior. We are currently witnessing an "AI gold rush" where the focus is entirely on the acquisition of GPU clusters and the training of massive, opaque models, while the structural integrity of the software that must actually deploy these models is left to wither.
The First Thing We Forgot: Objects are Supposed to Live
Object-oriented programming (OOP) was never merely a syntactic exercise involving classes and inheritance. Its fundamental, revolutionary intuition was rooted in the tangible: an object is a union of identity, state, and behavior. An object represents something that exists, something that remembers its condition, and something that possesses an intrinsic understanding of which operations are permitted to alter it.
In the current enterprise AI paradigm, we have moved toward a "stateless" architecture that treats intelligence as a transient service. When an LLM processes a prompt, it does so in a vacuum. It possesses no persistent identity and no long-term memory of its own internal state. It is a brilliant, ephemeral calculation engine.
The Erasure of Statefulness
By stripping away the "object" from our AI architecture, we have made enterprise integration infinitely harder. In a traditional, well-engineered software system, an entity—say, a customer record or an inventory item—is an object. It carries its history and its rules. In modern AI, the "intelligence" is decoupled from the data. We spend an exorbitant amount of time and resources on RAG (Retrieval-Augmented Generation) to bridge this gap, essentially trying to paste external memory onto an entity that fundamentally refuses to remember anything.
The Second Forgotten Pillar: Reinforcement Learning
If object-oriented design provided the structure for enterprise computing, reinforcement learning (RL) was intended to provide the feedback loop for intelligence. Yet, in the race to deploy chat interfaces and summarization tools, the industry has sidelined the rigorous application of RL in favor of simpler, supervised fine-tuning.
Reinforcement learning is the process of teaching a system through consequence. It is how we imbue a machine with the ability to navigate a landscape of choices, receiving a "reward" for desirable behavior and a "penalty" for failure. This is not merely a training method; it is the fundamental mechanism of evolution. By shifting focus toward massive pre-training and away from iterative, environment-based reinforcement, the industry has created systems that are essentially "post-trained"—they are static snapshots of the internet, unable to learn from the specific, nuanced feedback of the enterprise environment they inhabit.
Chronology of a Disconnect
- 1960s–1980s (The Era of Objects): The rise of Smalltalk and C++ established the primacy of encapsulation. Software became manageable because it was organized into self-contained, stateful entities.
- 1990s–2010s (The Rise of Data): Enterprise computing shifted toward relational databases and web services. While the state was moved to the database, the logic remained tethered to the object.
- 2017 (The Transformer Moment): The publication of "Attention Is All You Need" shifted the focus from symbolic, object-oriented logic to massive statistical patterns.
- 2020–2022 (The Generative Explosion): The industry pivoted entirely toward Large Language Models. The architectural lessons of the previous 40 years were treated as legacy overhead rather than foundational wisdom.
- 2023–Present (The Reality Check): Enterprises realize that LLMs, while impressive, are "black boxes" that lack persistence, governance, and verifiable state. The current "scramble" to build "AI compute you can actually own" is a reactionary effort to regain control over the infrastructure that we accidentally ceded to the cloud-scale giants.
Supporting Data: The Cost of Amnesia
The cost of this architectural shift is not just theoretical; it is reflected in the balance sheets of every enterprise attempting to move AI from a prototype to a production environment.
Industry data suggests that over 80% of enterprise AI projects fail to reach deployment. The primary culprit is not the model’s "intelligence," but the "integration gap." Without the durable, stateful objects that once allowed systems to communicate with precision, developers are forced to build increasingly complex "middleware" layers.
According to recent surveys from cloud infrastructure providers, the average enterprise is spending nearly 40% of its AI budget on data preparation and pipeline maintenance—tasks that were historically mitigated by object-oriented design patterns. We are burning capital to solve problems that we had already solved decades ago through better software engineering.
Official Responses: The Defensive Pivot
Leading AI research labs and infrastructure firms are beginning to acknowledge the "integration friction." In recent technical whitepapers, companies like OpenAI and Anthropic have begun discussing "agentic" workflows—which is, effectively, the industry’s way of rediscovering the need for stateful, long-lived, and task-oriented systems.
However, the official response from the "AI-first" movement remains defensive. They argue that the sheer scale of the new compute allows for a "brute force" approach that renders old engineering paradigms unnecessary. "Why build a complex state machine," one researcher noted in a recent forum, "when you can simply prompt the model to simulate one?"
This sentiment ignores the inherent fragility of such a system. Simulating a state is not the same as possessing one. When the simulation fails—when the LLM hallucinates or loses its context—the enterprise is left with no recovery path, no encapsulated logic, and no way to hold the system accountable.
Implications for the Future: Rebuilding the Foundation
The implications of this trajectory are clear: if enterprise AI is to become a stable, long-term foundation for business, we must undergo an architectural correction. We need to stop viewing AI as a monolithic "brain" and start viewing it as a component within a broader, object-oriented ecosystem.
1. The Return of Stateful Agents
Future architectures will likely look less like a single, massive model and more like a constellation of stateful agents. Each agent will own its data, hold its own state, and operate within strict behavioral boundaries—effectively resurrecting the best of the object-oriented era within a neural architecture.
2. Reinforcement as a Core Competency
Enterprises must move beyond generic models to environment-specific RL. A company’s internal AI should be trained on the specific consequences of its own business decisions. This requires a shift from "pre-training" to "continuous, feedback-driven evolution."
3. Owning the Compute
The scramble to build "compute you can actually own" is a manifestation of this desire for control. Enterprises are realizing that if they do not own the infrastructure—and the architectural principles that govern it—they are merely renting intelligence that they cannot govern or fully integrate.
Conclusion
The story of technological progress should not be one of forgetting. While the Transformers and LLMs of the modern era represent a quantum leap in pattern recognition and linguistic fluency, they are not a replacement for the fundamental tenets of software engineering. By embracing the lessons of the past—specifically the durable, stateful nature of objects and the iterative power of reinforcement learning—the enterprise can build an AI architecture that is not just impressive, but reliable, governable, and truly, fundamentally intelligent.
The current scramble is not just about compute; it is about reclaiming the architecture of our future. We must stop pretending that we are starting from zero and begin, once again, to build with the wisdom of everything we have learned.
Reynand Wu
Basiran
Asro
Laily UPN
Iffa Jayyana
Raul Delapena Setiawan
Siti Muinah
Layla Zulfa