Claude discovers major advance on Riemann hypothesis, improves 67-year-old bound

Big Seven's $315B AI infrastructure backlog signals earnings lag and structural shift to debt funding

Across the four largest hyperscalers, $315 billion in AI assets remain under construction and not yet generating depreciation, up from $281 billion last quarter. Meta faces a particularly acute 1.7-year lag before capex translates to live assets—meaning only one-third of its current spending will reach service within a year. The Big Seven's combined 2026 capex guidance of $863 billion (88% growth year-over-year, with roughly $550 billion AI-related) now exceeds their combined operating cash flow for the first time, forcing a structural shift in financing: debt funding is expected to rise from $16 billion in 2024 to $161 billion in 2026. This creates a dual headwind for near-term earnings: massive depreciation will eventually hit income statements as these assets go live, while the financing burden masks the true cash economics of the buildout.

AI-reviewed code shows lower revert rates than human review at scale, validating risk-scoring framework

Intercom's AI code review system approves pull requests five times faster than human review while achieving better production outcomes, using a six-dimensional risk-scoring framework (change size, blast radius, reversibility, security, operational impact, test completion) to separate low-risk auto-approvals from human review. This signals a meaningful shift in engineering operations and code quality metrics worth tracking as adoption spreads across enterprise development teams.

Meta and OpenAI release flagship open and specialized models; Zuckerberg frames personal superintelligence as competitive differentiator

Meta is open-sourcing Muse Spark 1.2 and releasing Muse Glimmer, a 30B parameter dense model optimized for local deployment on consumer GPUs (runs on RTX 3090 with 4-bit quantization at ~20GB, includes DFlash speculative decoding). Simultaneously, OpenAI launched Daybreak Blue and Red tiers with GPT-5.6-Cyber, a purpose-trained cybersecurity model that uncovered previously unknown vulnerabilities in Chrome's V8 engine and other open-source software. Zuckerberg published a major strategic essay framing personal superintelligence—agents, creation tools, tutoring, scientific collaboration—as Meta's differentiator versus labs focused on institutional AI, arguing that decentralized, individually-aligned superintelligence is necessary to prevent power concentration. He also addressed infrastructure buildout (energy and data center capacity to match China's pace), export controls on chips, and the recursive self-improvement dilemma for frontier labs. The moves signal competing visions: Meta betting on open-weight local deployment and individual agency, OpenAI embedding AI into enterprise security workflows. However, Zuckerberg's credibility on "open" principles is complicated by Meta's recent exclusion of competing AI assistants from WhatsApp until the European Commission intervened.

The AI Dashboard

Exponential View launched a live dashboard tracking five key gauges of AI industry health against previous boom-bust cycles. Main takeaways:

• Economic strain gauge measures AI capex as share of US GDP to assess sustainability of infrastructure spending
• Dashboard tracks industry strain (investment relative to revenue), revenue momentum, valuation heat (Nasdaq-100 P/E), and funding quality in real time
• Framework has circulated through boardrooms and policy circles; v1 dashboard updates continuously as new data arrives
• Designed to distinguish between genuine boom and unsustainable bubble using backtested metrics from historical cycles

Expanding Daybreak as the Cyber Defense Window Narrows

OpenAI published a product release on GPT-5.6-Cyber and expanded Daybreak access tiers for authorized cybersecurity work. Main takeaways:

• GPT-5.6-Cyber completes 95.0% of advanced cybersecurity requests (exploit chains, privilege escalation, auth bypass) vs. 1.5% for standard GPT-5.6 Sol, addressing persistent refusal issues from prior versions
• Model discovered two previously unknown V8 vulnerabilities (CVE-2026-15903) plus 400+ kernel privilege escalation issues, 5+ mobile OS vulnerabilities, and 3 critical database vulnerabilities through real-world testing
• Daybreak Blue provides guardrail-removed access to GPT-5.6 Sol for general defenders; Daybreak Red provides specialized GPT-5.6-Cyber for vulnerability research and exploit validation
• Model assessed as High (not Critical) on Preparedness Framework cybersecurity capability threshold; requires hardware security keys, auto-review mode, and sandboxed execution for risk mitigation

Claude Improves Riemann Zeta Function Lower Bound from 41.6% to 67.2%

Anthropic published a research report on Claude's mathematical discovery work on the Riemann hypothesis. Main takeaways:

• An unreleased Claude research version improved the lower bound for zeros of the Riemann zeta function satisfying the Riemann hypothesis from 41.6% to 67.2%, a significant advance on a decades-old mathematical problem.
• Claude achieved this by combining recent work from Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and Bombieri in a novel way, using quadratic form analysis without requiring the hypothesis itself to be true.
• The discovery emerged as an unintended byproduct when Claude was prompted to attempt the Riemann hypothesis directly; it generated 650 failed ideas initially, then coordinated 60 subagents running 2,400 shell commands over 1.5 days before finding the result.
• Claude used 31 million output tokens across two sessions in Claude Code, with human input limited mostly to encouragement; the model validated its own work through peer review among subagents and formal Lean verification.
• Two Anthropic mathematicians (Alpöge and Furman) independently validated the proof, and external experts Brian Conrey and Dan Goldston confirmed the mathematical validity of the advance.