February 28, 2026
Proofs, vibes, and pitchforks
Deterministic Programming with LLMs
Math bots solve puzzles; devs clap back: “It’s just tests with extra steps”
TLDR: An AI-assisted pipeline used ChatGPT, Aristotle, and Lean to verify a new math proof, suggesting a path to dependable AI coding. Commenters battled over whether this is genuine progress or just “write tests first” with new branding, mixing memes, skepticism, and arguments about repeatability.
The essay pitches a calm, repeatable way to use AI code tools—think “make the machine check itself”—and points to math’s big flex: an AI-assisted proof pipeline where ChatGPT sketched an argument, Aristotle patched the logic, and Lean (a proof checker) verified a previously unsolved Erdős problem. Drama alert: the moment “deterministic programming” hit the feed, the crowd split. One camp cheered the math analogy; another rolled eyes so hard you could hear it.
Top snark came from andyfilms1, who asked if this just loops back to old-school genetic algorithms and dropped a retro Formulize/Nutonian link. dataviz1000 gave the sober take: the real move is code-checking code, especially for stuff with known outcomes (like paginated tables), while the unknown, “vibes-only” features still need humans. Then computersuck came in swinging: “long article, says nothing… probably AI,” boiling it down to “write the tests first” with extra buzzwords. The single-word meme soon landed like a prophecy, and nemo1618 pushed back on the “LLMs are random” framing, hinting you can force repeatability with careful settings.
Bottom line: the article says math bots can make trustworthy work when verified; the comments say tech has seen this movie, and the twist is whether “vibecoding” becomes “verified coding” or just a rebrand of testing.
Key Points
- •The article advocates exploring deterministic use of large language models in software development.
- •It highlights risks of accepting AI-generated mathematical proofs due to hallucinations and subtle errors.
- •Formal proof assistants like Lean can verify proofs rigorously, though they are hard for humans to use directly.
- •The article reports a January 2026 AI-assisted solution to an Erdős problem via a pipeline using ChatGPT, Aristotle, and Lean.
- •Programming agents such as Claude Code and Gemini Code Assist are referenced in the context of deterministic tooling for coding.