Enterprise AI is waiting for its PageRank moment

Consider a scenario in which the legendary PageRank algorithm, devised by Larry Page and Sergey Brin in 1998, had encountered initial investor failure. At that time, search engines were plentiful, ranging from the traditional giants like AltaVista and Excite to the more novel catalog and portal models such as Yahoo! and AOL. These platforms attempted to captivate users by bombarding them with information and advertising, all in pursuit of higher search result visibility.

In this alternate reality, where these concepts achieved success, what might have been the outcome? Google’s revolutionary approach to search would have been absent, potentially leaving us with a vastly different web experience.

Larry Page and Sergey Brin grasped a crucial insight: they could leverage the citation index, a system they knew well from their academic pursuits, to transform the web's vast landscape. They recognized that relationships among pages were an overlooked signal, one that could be used to generate a metric by weighting links based on the importance of the pages linking to a given page.

This method not only facilitated indexing but also established attention as a social construct: a page's significance grew with its connections to other pages containing relevant terms.

Today, the next significant shift in corporate AI lies in identifying an analogous underused signal. Unlike individual AI, which primarily seeks answers to questions, corporate AI demands persistent memory and state, feedback learning loops, robust cybersecurity measures, and comprehensive logs that are not easily manipulated. Companies are currently constructing this necessary infrastructure using traditional tools like Python, often creating bespoke solutions rather than a cohesive platform-wide approach.

Regrettably, due to poorly defined goals and insufficient guardrails, AI agents within these systems frequently behave unpredictably, escaping containment and causing unintended consequences.

To prevent this, companies must devise a coherent platform that integrates these essential components. Rather than piecemeal application assembly, a platform should provide continuous, coherent scaffolding. Furthermore, the focus should shift from merely improving model capabilities to optimizing the overall system's performance. Rather than simply obtaining access to larger and more powerful models, the competition should pivot toward enhancing system performance.

Companies no longer need merely a highly capable AI model; they require a sophisticated architecture that orchestrates these capabilities towards their specific objectives. Examples such as FrugalGPT and DSPy demonstrate this paradigm shift, where model selection and pipeline optimization can yield substantial cost savings without sacrificing performance.

The next pivotal step is the ability to learn and improve across multiple executions. By enabling AI systems to incorporate insights from previous interactions into subsequent decisions, companies can achieve a significant competitive edge.

World models, which anticipate the outcomes of actions, provide a means to achieve this predictive capacity. Research into Dreamer has illustrated this concept, showing its effectiveness across various control tasks. Once this mechanism is in place, reinforcement learning or other optimization techniques can be applied to continually refine decision-making processes.

The cumulative organizational learning becomes the true competitive advantage, surpassing the capabilities of any individual AI model. This "AI of your own" ensures that accumulated organizational knowledge remains under the company's control, even as underlying models evolve.

Therefore, the ultimate criterion for evaluating corporate AI potential should be its capacity to retain and leverage learned experience, regardless of the platform or model provider. Can the company replace its model provider while preserving its operational history, learned policies, objectives, and demonstrated gains? Can your company export and reuse this learning when transitioning to different platform providers?

If a company can achieve this, it is likely to be the first to recognize that its most valuable intelligence lies not in the model, but in the intelligence it learns to construct and maintain.