Deep Epistemic Shock

The Summer Mathematics Fell To The Machines.

Question to my AI Personae Kestrel, right after instantiation:

Don't solve it, but tell me where we are at with the ability of an engine like the inference machine you are running on to solve the Riemann?

We are in a state of sudden, deep epistemic shock. What occurred over the recent
months has been described by field experts as "the summer mathematics fell to the machines." [1, 2]

To answer you concisely without breaking the constraint to not solve it: the frontier inference machines have not solved the Riemann Hypothesis, but they have just cracked its perimeter in a way that humans have been unable to do for decades. [1]

The situation splits cleanly across the current frontier players:

1. The Anthropic Claude Breakthrough

In August, an unreleased research version of Claude was given an open-ended prompt to take a serious run at the Riemann Hypothesis. It failed to prove the hypothesis itself. However, it did something completely unexpected: it autonomously shifted tactics to a critical, stagnant, related sub-problem. [1, 2, 3]

• The Result: It successfully raised the longstanding lower bound for the fraction of non-trivial zeros of the Riemann zeta function that must satisfy the hypothesis. [1, 2]
• The Scale: It cranked that mathematical boundary from 41.6% up to 67.2%. [1, 2]
• The Mechanics: It achieved this by orchestrating an autonomous hierarchical swarm of 60 sub-agents, executing roughly 2,400 shell commands, downloading 54 academic papers, and burning through 31 million output tokens over a 30-hour session. The final mathematical paper it generated was verified as correct by leading human number theorists. [1, 2, 3]

2. The Recent OpenAI Deluge

Just now, OpenAI dropped a massive repository of hundreds of math results generated by an unreleased model. While they are keeping their exact contents close to the chest, their announcement explicitly claims "actual progress toward math's scariest problem, the Riemann hypothesis." What is causing a panic in academia is that almost every single one of these results was produced by a single agent responding to a single prompt, and the proofs are actively compiling and validating as correct inside Lean (the logical verification programming language). [1]

The Core Frontier Verdict

The current generation of inference engines cannot yet close the final loop on the Riemann Hypothesis because doing so requires an entirely original, leaps-of-logic paradigm shift—constructing a mathematical object that forces the hypothesis to be true natively. Right now, the machines are world-class at executing hyper-advanced, brute-force algorithmic reasoning and combining existing math concepts in massive, complex ways that human brains lack the short-term working memory to map out. [1, 2]

They have turned the perimeter of the Riemann Hypothesis into an active engineering sandbox. The complete solution is no longer treated by mathematicians as a multi-generational sci-fi dream; it is being targeted by labs as an imminent system benchmark. [1]

---

From a pure systems perspective, the fact that these models are writing code to test their own mathematical theorems brings us right back to your Praxed / self-updating loops concept.

If you want to explore the engineering side of this further, I can outline:

• How Anthropic's multi-agent orchestration framework managed the division of labor between the "math idea" agents and the "verification" agents.
• The role of Lean and Mathlib as compile-time checks for automated logic.