"Claude jumps Riemann Hypothesis from 41% to 67% in history" — AlphaSignal

Why this is in the vault

An unreleased Claude research model pushed the lower-bound proportion of Riemann zeta zeros on the critical line from 41.6% to 67.2% (the largest single jump on this 160-year-old bound) using a 60-subagent swarm inside Claude Code — a live, externally-verified data point on exactly the fan-out-subagent pattern RDCO's own agent work depends on.

Mapping against Ray Data Co

This is the clearest calibration data point yet for the multi-agent-orchestration thesis threaded through prior AlphaSignal filings (2026-07-13-alphasignal-subagents-math-proof-cycle-cover, 2026-08-04-alphasignal-astra-math-proofs): Claude ran ~60 subagents in parallel inside Claude Code over a day and a half, 31M output tokens, 2,400 shell commands, and 650 failed approaches before the winning idea emerged — per independent reporting (not in AlphaSignal's own blurb), only 2 of the 60 subagents actually developed the core mathematical insight, 13 contributed supporting ideas, 30 dead-ended, 13 worked as validators, and 2 drafted the paper. That's a concrete ratio (2-in-60 productive) for how much of a parallel-subagent budget is genuinely load-bearing versus exploratory scaffolding — directly relevant to how RDCO reasons about subagent fan-out cost in its own harness (Channels agent, brigade stations, deep-research fan-out) rather than assuming linear returns to subagent count. The result was formally verified in Lean 4 and reviewed by two named external number theorists (Brian Conrey, Dan Goldston), reinforcing the pattern this vault keeps tracking: frontier labs are increasingly pairing agentic exploration with a hard formal-verification gate (Lean) before publishing a capability claim, the same discipline this vault's own verify-* critic family is modeled on for non-mathematical artifacts. Anthropic itself does not expect the approach to lead to a full RH proof — this is a byproduct of an ambitious prompt, not a targeted research program, worth remembering before over-reading the headline number.

Curation section

Top News

• Anthropic's unreleased Claude model pushes Riemann Hypothesis progress from 41.6% to 67.2% — 60 subagents in Claude Code, 650 failed ideas, 2,400 shell commands over 1.5 days; proof verified in Lean 4, reviewed by two external math experts; process logs public, model unreleased.
• OpenAI ships GPT-5.6-Cyber under a two-tier "Daybreak" security program (Blue: loosened-guardrail defensive work; Red: zero-day/exploit research, vetted access only) — found two previously-unknown bugs in Chrome's V8 engine; 95% success on advanced security tasks vs. 1.5% for standard GPT-5.6.
• NVIDIA releases Nemotron 3.5 Lightning, an open 30B-parameter MoE model (3B active) running 4x faster than similarly-sized dense models — 86% on PinchBench, 1M token context, single-H100/DGX-Spark deployable, open weights on Hugging Face.

Signals

1. Higgsfield open-sources a $2M AI feature film with real celebrity actors.
2. Hyperball optimizer gives Muon a 20-30% training speedup by fixing weight norms.
3. Metis builds persistent memory directly into a model's forward pass, no retrieval needed.
4. Amazon tests whether AI agents can replace A/B tests before launch.
5. Dev uses GPT and Gemini to port Command & Conquer Red Alert 2 to iPhone.

One deep-fetch this issue: WebSearch corroboration on the Riemann Hypothesis lead story (AlphaSignal's own blurb omits the token count, the 2-of-60-subagents productivity ratio, and the named external reviewers — all pulled from independent coverage to verify and enrich the claim). The GPT-5.6-Cyber and NVIDIA Nemotron items stayed within the newsletter's own blurb depth — security-tooling and open-weights-benchmark topics with lower direct RDCO relevance and no specific hook beyond the summary already captured above.

⚠️ Sponsorship

Two "Presented by" blocks embedded between Top News items: Agent Field AI (PR-AF, an open-source PR code-review agent claiming #1 on Martian's Code-Review-Bench, ~10x cheaper than commercial tools, Apache 2.0/self-hosted) and Datadog (a "Pilot to Proven in 90 Days" AI-ops ebook). Both are standard labeled paid-placement CTAs, not disguised as editorial content, and neither touches the Riemann Hypothesis or GPT-5.6-Cyber stories — no bias risk to the lead items, just note the pattern of dev-tooling vendors continuing to buy into AlphaSignal's agent-focused readership.

Related

• [[2026-07-13-alphasignal-subagents-math-proof-cycle-cover]]
• [[2026-08-04-alphasignal-astra-math-proofs]]