This might sound trivial if you already agree, and too reductive if you don’t. But I want to make and defend the claim that every problem we face is ultimately a problem about making the right predictions. Whether you’re deciding where to invest, whom to hire, which medication to take, or what to say to a friend who’s upset; whether you’re a theoretical physicist, an engineer, a historian, or a theologian; whether you’re a child, a teen, or an adult—whatever the nature of the problem you’re dealing with, it is fundamentally a prediction problem. Consequently, the better your predictions, the better your solutions.
Why might such a straightforward, commonsensical view be contentious? In fact, throughout history, many influential thinkers have explicitly rejected it. From Plato on beliefs to Albert Einstein on quantum physics to Noam Chomsky on contemporary AI, philosophers and scientists have insisted that knowledge requires something more than true predictions. And they’ve had good reasons for saying so. Lucky guesses clearly aren’t knowledge. Shallow pattern matching breaks when challenged, as imperfect AI systems remind us.
But I want to suggest the critics, although pointing at something real, have misunderstood the nature of the problem. What they’re identifying isn’t an alternative to prediction. Instead, it’s what helps us make more successful predictions.
The familiar objection goes back to Plato. In the Theaetetus, Socrates considers whether knowledge might simply be true belief, and rejects the idea with a thought experiment about a jury.
Imagine jurors deliberating a case. They weren’t eyewitnesses to the crime. They’ve heard persuasive rhetoric. Suppose they reach a guilty verdict, and they happen to be right. The defendant did indeed carry out the crime.
But did the jurors know he was guilty? No. They had a true belief, but something important was missing. They reached the right conclusion through a process that could easily have led them astray. A more charismatic defense attorney might have swayed them the other way. They were “justly persuaded”, perhaps, but rhetorical persuasion isn’t the same as knowledge.
We feel this intuition immediately. If we had to predict whether the jurors would get the next case right, we wouldn’t trust them. Their success doesn’t seem systematic or reliable. It seems contingent on who spoke most eloquently that day.
For more than two thousand years, this has been the standard objection to identifying knowledge with correct prediction: you can get the right answer for the wrong reasons. And when that happens, you don’t really know.
What, then, is missing? The standard answer is that the jurors lacked explanation, something more than mere successful prediction.
But I would argue that what the jurors really lacked was a method that would continue to work, not something mysterious called “explanation” whose nature is never made concrete.
Thinkers like the philosopher Karl Popper argue in favor of explanation over mere prediction. But explanation matters only to the extent that it extends our predictive success. We never observe causes, as David Hume pointed out, but causal understanding still matters if it tells us what will happen when circumstances change. Knowing what’s impossible matters not because it’s real (we never observe it, since it doesn’t actually occur), but because it rules out paths in advance and thus helps make predictions about what does occur. These are not alternatives to prediction. They are methods that help us attain greater success with our predictions.
Our quest for predictive success may necessitate the use of these methods. As Ilya Sutskever, co-founder of OpenAI, notes, when you train a neural network to successfully predict the next word, it doesn’t just learn statistical correlations; it learns “some representation of the process that produced the text”. The pursuit of prediction, pushed far enough, generates what we describe as explanation.
But if we take explanations as the essence of knowledge, if we see explanations as the end, we run the risk of obscuring our ignorance and producing a satisfactory subjective state through narratives that we mistake for knowledge. Reliance on predictions helps mitigate this risk, as demonstrated by the success of modern science, which delivers predictions that pre-modern natural philosophy, despite its elaborate explanations, failed to provide.
Every problem is a prediction problem. Our challenge is to see how far our successful predictions go, and how to make them go further.
Continue Reading:
→ A Response to Noam Chomsky on Machine Learning and Knowledge
→ Michael Huemer: Nature of Knowledge, Foundations of Morality (podcast)
→ A Conversation with Martin Frické on the Epistemology of Machine Learning
Other Projects:
→ Universal Open Textbook Initiative — World’s largest repository of free, multilingual textbooks.
→ Nasada — A productivity suite for the AI era.