In an era where Artificial Intelligence (AI) is increasingly embedded in our daily lives, the notion of knowledge production and transmission has become a subject of concern. At the beginning of the year I highlighted a key fundamental challenge around skills transmission (and its potential breakdown, if we remain in the logic of considering humans or labor as mere factors of production, or, even worse, as farmland). The rise of AI-powered systems that can retrieve information with remarkable accuracy raises fundamental questions about the role of human inference in shaping our understanding of the world. I have repeatedly written or mentioned on my social media that “Inference Will Meet Retrieval”. The advent of “reasoning” models highlights this distinction:
The most recent iteration of reasoning models allows for retrieval based on reasoning traces. This gets us close to, well, holding the keys to ones brain. To gain, root level access. What does this mean? Well, it quite literally means knowing how to hack people and deeply reprogram them in ways that they may not even consider strange. That is why I keep on banging about the very real fact that the US is quite aggressively rolling out its form of digital ID, in the process fiscalising the world on its very own terms. My own vision of the world here is one of decentrally stewarded identity, that enables decentral reprogramming that may be done by friends and family in healthy interpersonal relational setup. The only barrier, as usual, is distrust and coordination failures, possibly supported by a baked in demand for authoritarianism baked in paternalistic programming of most societies relational and care structures.
The Distinction between Inference and Retrieval
At its core, inference is a statistical process that enables us to learn patterns and make predictions based on data. This is achieved through econometric exercises, which allow us to extract insights from complex datasets. It is often explicitly high dimensional in nature to ensure there is some form of convergence on a narrative. A summary of a high dimensional observation may then well be turned into an isolated individual case study, or, lets call it a story. But of course, the base information layer in societies is presently in control of, well, often times, special interest groups.
What is retrieval? In contrast, retrieval refers to the act of asking questions and seeking answers, often in the form of binary statements (true or false, yes or no). While inference is concerned with adding nuance to our understanding, retrieval provides a more simplistic, yes-or-no response. If a query lacks context, a simple yes- and no- answer may induce individual actions that can be attributed to the interaction of a latent belief set or, lets call it the base layer of programming of individuals (call it trauma, experiences etc., pre cognition), with the answer to the question. Of course, you are not in control of that interaction because often times you neither know what the base layer is, nor do you know what the question was. Low dimensional retrieval is thus — incredibly dangerous. It does not help that politics, media etc and increasingly science favors a selection and award of “success” or recognition that may encourage selection on narcissistic personality traits.
The Friction between Inference and Retrieval
So in other words, my framework of thinking suggests that there exists a friction or relationship between these two approaches to knowledge. When we retrieve information, we often receive a binary statement that can be either true or false. However, inference is a more complex process that involves adding noise and nuance to our understanding, making it inherently more difficult to pin down a definitive answer. This friction arises from the fact that inference requires us to consider multiple factors and perspectives, whereas retrieval provides a more straightforward response.
A key dimension of this friction is the high dimensionality of reasoning itself. Inference is not just about processing information; it’s also about making connections between seemingly unrelated concepts and ideas. This process can be incredibly complex, involving multiple variables, relationships, and contexts. It does require high dimensional considerations. In contrast, retrieval often relies on a more one-dimensional approach, where answers are filtered through pre-existing categories and labels.
The big selection or sorting that I alluded to (“die Spreu wird sich vom Weizen trennen” in the 2023 SHAPER Keynote that I had to interrupt as I was overwhelmed by emotion) could then well be among individuals that search for homophily in reasoning space; while others, search for homophily in values space. Or put different: logic clashes with values. This is a far more sophisticated way of thinking about it then the simplistic narrative of values versus the economy that is offered around explanations of populism. If you zoom out, you will see Economics taking on some narratives from Political Science, and presently, Political Science taking on some narratives from Economics.
But the above is very suggestive that we will not see alignment but some form of congruence or sorting which will, maybe, appear like alignment in a different latent or quite physical space (this is the population swap comment).
The ease with which AI systems can retrieve information can have unintended consequences for human behavior. Even if an answer is not necessarily accurate or nuanced, it can still induce humans to act based on the information provided. This highlights the need for a more balanced approach that acknowledges both the benefits of retrieval and the limitations of inference. In any case, Putin has weaponised many of these dimensons in his hybrid war.
Broader Implications
The implications of “Inference Will Meet Retrieval” extend far beyond the realm of AI-powered systems. As we increasingly rely on AI-augmented knowledge systems, we must consider the potential consequences for human reasoning and knowledge transmission as I highlighted at the beginning of the year. If we can retrieve answers to questions with ease, but are no longer required to engage in critical thinking and inference, what does this mean for our understanding of the world? The fundamental human challenge is that we enter the world with an empty brain and ultimately, it is each generations job to re-learn the body of knowledge that mankind has produced. This is increasingly difficult and may generate societal structures that can be represented through this matrix of firm-to-firm or sector-to-sector trade relationships within an economy. I will write some more about this in the context of the US/EU trade dispute which seems to be escalating.
How will education and knowledge transmission evolve in an era where AI is increasingly prevalent?
The Future of Human Reasoning
As we navigate this complex landscape, it is essential that we consider the role of human inference in shaping our understanding of the world. By acknowledging the friction between inference and retrieval, we can begin to design knowledge systems that complement rather than replace human reasoning. We must also recognize the limitations of AI-powered as well as human powered retrieval and ensure that these systems are used in ways that promote nuanced thinking and informed decision-making. In the end, it will all boil down to the answer of questions around what it means to be human and human as a relational being.