What's the difference between retrieving a fact and truly reasoning? Prof. Kambhampati Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) begins by noting that human reasoning is tricky to define. Yet, since the Greeks, we’ve relied on formal logic (like syllogisms) to guide sound reasoning. A thread 🧵👇

2/7 He points out that we lack a neat definition of “human reasoning,” but that hasn’t stopped our civilization from moving forward. We built entire disciplines—from Aristotle’s logic to modern computer science—on the foundation of formal, structured reasoning.

3/7 Kambhampati contrasts retrieval (pulling stored info) with reasoning (connecting ideas with logical rigor). Just because we string ideas together doesn’t guarantee we’re actually reasoning—there must be standards for correctness and validity.

4/7 He humorously references Monty Python’s witch trial scene: the argument is “If she floats like wood, she must be a witch.” It looks like reasoning—there’s a chain of statements—but it’s clearly not sound. It’s a playful reminder that not all stepwise arguments are valid.

5/7 Between raw retrieval (“She’s a witch!") and truly logical inferences, there’s a vast middle ground of fallacies and “Monty Python logic” that mimics reasoning but fails basic tests of correctness.

6/7 Hence, in AI (and broader AGI pursuits), we can’t just replicate the surface features of human reasoning. We need rigorous definitions and methods—logic, probability, evidence—to separate mere associations from true inference.

7/7 By preserving “sound reasoning” standards, Kambhampati argues we uphold the legacy of centuries of philosophical and mathematical thought, aiming for AI that doesn’t just retrieve but reasons in a formally robust way.

What is the nature of an existence that is experienced entirely outside of time itself? Can a single decision that is made in a state of timelessness simultaneously affect EVERY point in time and space? Groundbreaking reconciliation of creationism with natural science.

@rao2z I touched on the same concept on my last substack. LLMs are statement generators with populist vote based on training data and hallucinations are not bug but feature. We need formal systems like logic for verifying truth among these statements.

@rao2z Come on who needs logic? 🤪