We appreciate comments from Christopher Henson, Zeke Medley, Ankit Kumar, and Pete Manolios. This post was initialized by Max’s twitter thread.

Introduction

There's been a lot of chatter recently on HN and elsewhere about how formal verification is the obvious use-case for AI. While we broadly agree, we think much of the discourse is kinda wrong because it incorrectly presumes formal = slopless.^[1]Over the years, we have written our fair share of good and bad formal code. In this post, we hope to convince you that formal code can be sloppy, and that this has serious implications for anyone who hopes to bootstrap superintelligence by using formality to reinforce “good” reasoning.

A mainstay on the Lean Zulip named Gas Station Manager has written that hallucination-free program synthesis^[2]is achievable by vibing software directly in Lean, with the caveat that the agent also needs to prove the software correct. The AI safety case is basically: wouldn’t it be great if a cheap (i.e. O(laptop)) signal could protect you from sycophantic hubris and other classes of mistake, without you having to manually audit all outputs?

A fable right outta aesop

Recently a computer scientist (who we will spare from naming) was convinced he had solved a major mathematics problem. Lean was happy with it, he reasoned, given that his proof mostly worked, with just a few red squigglies. As seasoned proof engineers, we could have told him that in proof engineering, the growth in further needed edits is superlinear in number-of-red-squigglies (unlike in regular programming). The difference between mistakes in a proof and mistakes in a program is that you cannot fix a broken proof in a way that changes its formal goal (the theorem statement). In contrast, many, if not most changes to traditional software impact its formal spec, for example by adding a side-effect or changing the shape of an output. Therefore proof bugs are 1) harder to fix, and 2) more likely to imply that your goal is fundamentally unachievable (the theorem is wrong). This made up chart illustrates the principle, a rough “lore” level consensus in the field without any hard data.

It is possible he will post a finished proof, but the referee-time of bets he made has lapsed, so we can take away some lessons. Did our protagonist take to heart the promise of formal methods as slopless?

Your formal model might not be proof-id

...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments