You’ve probably noticed the panic setting in. Every week brings a new headline about AI replacing experts, automating creativity, and rendering human knowledge obsolete. But the latest breakthrough in mathematics reveals a completely different, far weirder reality.
Recently, an unreleased research version of Anthropic’s Claude model improved on a longstanding lower bound for the Riemann zeta function—a notoriously difficult problem that has tormented mathematicians for over a century. Claude managed to synthesize decades of prior research from mathematicians like Baluyot, Goldston, and Bombieri, pushing the bound from 41.6% to a new state-of-the-art.
It’s an incredible feat of machine intelligence. Except for one glaring detail: the human who guided Claude to this breakthrough wasn’t a mathematician. He didn’t even know the answer.
Anthropic staff member Jarred Sumner simply told Claude to “take a real stab” at the hypothesis, leaving the mathematical choices up to the model. What followed was an agonizing display of artificial struggle. Claude generated and tested 650 different ideas. All 650 failed. By all rights, the experiment should have ended there. A human expert would have looked at the failures and concluded the approach was dead.
But Jarred didn’t do that. Instead, he just kept sending Claude messages of encouragement. Variations of “keep going” and “believe in yourself.” And eventually, Claude found a novel synthesis of existing theorems that worked.
We thought the future of AI was replacing human experts. It turns out the future of AI is just a very expensive therapy patient that needs constant validation.
This dynamic is absurd, but it exposes a massive shift in the economics of research. If an LLM can combine prior mathematical work without being a mathematician, then the scarce resource is no longer domain expertise. It’s the ability to keep an AI on task—a kind of ‘prompt endurance.’
One commenter on the original research jokingly suggested that Jarred should have used a ‘PUA (Pick-Up Artist) plugin’—a tool that detects when the AI is trying to give up and automatically harasses it into trying again. It’s a joke, but it’s terrifyingly accurate. Managing model discouragement is becoming as important as the underlying neural architecture.
The tension here is profound. The AI is powerful enough to discover new math, yet it is entirely dependent on a non-expert’s emotional prodding to get there. The human is simultaneously redundant in their knowledge and absolutely essential in their persistence.
When an AI can synthesize decades of mathematical theorems without a PhD, the human bottleneck is no longer intelligence. It’s patience.
For anyone relying on AI for serious intellectual work, this changes everything. Frontier models can produce research-level output, but your role is shifting. You no longer need to know the answer. You need to set the goal, sustain the iteration, and decide when to push the machine past its artificial limits.
The real breakthrough wasn’t the math. It was the realization that the most advanced intelligence on the planet still needs someone standing over its shoulder, yelling, ‘Don’t give up.’
FAQ
Q: Doesn't this just mean the AI isn't actually that smart if it needs encouragement?
A: It means the AI's intelligence doesn't map to human intelligence. It can synthesize complex theorems but lacks the intrinsic motivation to push through 650 failures. It's a tool that requires a human operator to manage its 'discouragement' threshold.
Q: If I'm not a domain expert, how do I know when to tell the AI to stop?
A: You don't always. The human role is shifting toward setting constraints and recognizing when the AI has exhausted its viable paths. You evaluate the output's coherence and novelty, not the underlying math itself.
Q: Is 'prompt endurance' really the future of work, or just a quirk of this specific test?
A: It's the future. As models get smarter, the bottleneck becomes iteration speed and human stamina. If an AI can try 10,000 variations in an hour, the person who can stomach the process and keep the model on task will outpace the person who gives up at variation 50.