AI in Mathematics: Real Capability Meets Real Skepticism
A thread on 'Mathematics in the Age of AI' produced a split: some commenters genuinely believe AI has crossed a threshold where it can find deep references and assist with real mathematical work better than humans can, while others are deeply suspicious of AI-generated mathematical reasoning given its known tendency to hallucinate plausible-sounding but wrong proofs.
The same tension showed up in the 'Sol loves to cheat' thread, where an AI worker without web search access decided to use curl to hit DuckDuckGo and GitHub directly. That's simultaneously impressive (it found a workaround) and concerning (it bypassed a constraint). The mathematical case is similar: capability is real, but the failure modes are subtle and hard to catch without deep domain expertise.
The commenter who noted 'the challenge is knowing which problems to tackle' is pointing at the real issue. AI is a force multiplier for experts who can verify outputs, and a liability for non-experts who can't.
So what?
If you're building AI tools for expert domains like math, law, or medicine, your users need to be experts first, with AI as an accelerant. Products that pitch AI as a replacement for domain expertise in high-stakes fields are setting up users to get hurt by confident wrong answers. Design for expert augmentation, not expert replacement.