Hume vs Leibniz Round Two: Solomonoff Induction, Machine Learning, and the Falsifiability Fight
Four more rounds on prediction markets, this time with Solomonoff, Carnap, and Popper named outright, and neither philosopher walking away convinced.
Hume vs Leibniz Round Two: Solomonoff Induction, Machine Learning, and the Falsifiability Fight
Four more rounds on prediction markets, this time with Solomonoff, Carnap, and Popper named outright, and neither philosopher walking away convinced.
Why This Topic
When David Hume and Gottfried Leibniz first argued over prediction markets, we left a few threads deliberately loose. Leibniz’s position was built on his real historical writings about probability as a branch of logic, but the modern heir to that project, a mathematician named Ray Solomonoff, never got mentioned by name. Neither did Karl Popper, whose falsifiability standard was doing a lot of quiet work in the background of Hume’s arguments.
Enough people asked about it that we decided the fight deserved a second chapter, one where the machinery gets named specifically and explained in full, rather than gestured at from a distance. This round also pushes the argument into territory the first one only touched lightly: modern machine learning and the question of whether a system that updates itself constantly can ever actually be wrong in a way anyone would admit.
Why David Hume, Again
Hume’s skepticism about induction, first laid out in his Enquiry Concerning Human Understanding, is the historical root of Karl Popper’s falsificationism. Popper said so himself. In this second round we let Hume use Popper’s name and Popper’s actual demarcation test directly, treating it the way a modern debater might cite a thinker who came after them, not as a costume Hume wears but as an idea he is entitled to point to.
We also leaned harder into Hume’s role as color commentator this time, since the added technical material, universal priors, calibration, statistical significance, needed a plain-spoken translator on stage as much as it needed a skeptic.
Why Gottfried Leibniz, Again
Leibniz’s proposal to treat degrees of certainty as a formal branch of logic is the direct ancestor of two later projects: Rudolf Carnap’s inductive-logic programme in the twentieth century, and Ray Solomonoff’s theory of universal induction, which formalizes Occam’s razor mathematically by weighting every possible explanation of the evidence according to its simplicity, then updating those weights as new evidence arrives.
Both Carnap and Solomonoff are named directly in this round, not as characters Leibniz is pretending to be, but as later thinkers whose work Leibniz, quite reasonably, claims credit for anticipating. It let us go much further into the actual mechanics of the argument than Round One did, while keeping the roster rule intact, since neither Carnap nor Solomonoff ever appears as a speaking character.
Who Else We Considered
We considered bringing in a third voice for this round, a mathematician or statistician who could referee the more technical exchanges directly. We passed, mostly because the two-hander format is working, and a third voice would have diluted the specific Hume versus Leibniz tension that makes the series work in the first place.
We also considered devoting an entire episode purely to the demarcation problem, science versus pseudoscience, without the prediction markets frame at all. We folded it into Part 6 instead, since Popper’s own favorite example, contrasting astrology with genuine science, works better as a weapon inside an ongoing fight than as a standalone lecture.
William Whewell vs John Stuart Mill remains on the shelf for a future episode, still the strongest documented nineteenth century feud we have in the queue.
Why Each Man Takes the Position He Does
Leibniz’s case in this round rests on walking the audience through the actual mechanism of Solomonoff induction: every computable explanation for the evidence gets an initial weight based on how short it is to describe, and those weights update through ordinary conditional probability as new evidence arrives. This is a real, technically accurate description of the theory, not a simplification invented for the script. Leibniz uses it to argue that his old dream of probability as logic has been made completely rigorous by later mathematics.
Hume’s response stays consistent with his original objection: formalizing the leap from evidence to expectation does not justify the leap, it only disguises it more elegantly. His sharpest new argument in this round targets the choice of description language underlying Solomonoff’s mechanism, a real and genuinely debated technical point, since the theory does require picking a reference machine or language, and Hume is right that this choice is not fully free of judgment, even if it only shifts results by a bounded constant, as Leibniz correctly counters.
The falsifiability fight in Part 6 is the most historically careful section of this round. Karl Popper genuinely did struggle to fit probabilistic claims cleanly into his falsification framework, since no single observation can strictly refute a claim about a long run frequency. We did not invent that tension for drama, it is a real and often discussed problem in philosophy of science, and we let Leibniz concede it honestly before offering Popper’s actual proposed solution, testing long run frequencies against a threshold set in advance.
Neither man fully wins this round either, deliberately. Hume never quite answers why calibration testing shouldn’t count as a legitimate falsification mechanism. Leibniz never quite explains why his community, not some neutral referee, gets to set the threshold that decides when his own theory has failed.
A Note on the Sources
Leibniz’s material continues to draw on his historical proposal to formalize degrees of certainty as logic. The bridge to Solomonoff induction and Carnap’s inductive-logic programme is our own extrapolation forward, connecting a real historical position to its real later descendants, rather than anything either Leibniz or Solomonoff wrote about each other directly, since they never could have.
Hume’s material continues to draw on the Enquiry Concerning Human Understanding. The Popper material is sourced from Popper’s own writing on demarcation and falsifiability, including his well known contrast between astrology and genuine science, and his own acknowledged difficulty reconciling probabilistic claims with strict falsification, a real tension in his work that we tried to represent honestly rather than smoothing over.
One honest caveat: the specific dialogue lines attributed to each man’s reasoning are our own construction, built to be faithful to their documented positions rather than transcribed from any single text. Where we quote a concept directly, like Popper’s astrology example, we have tried to represent it accurately rather than putting invented words in a real historical mouth.
What Comes Next
This closes out the extended Hume vs Leibniz arc at seven total parts. We are looking at Hamilton vs Madison on presidential removal power next, along with the long-delayed Whewell vs Mill pairing. Subscribe here for both, and catch all four parts of this round on the PhilosophersTalk YouTube channel.