Claude Pushes Riemann Hypothesis To New Levels, Meta's Superintelligence Vision

Today's AI news: Anthropic had an unreleased version of Claude take a real stab at the Riemann hypothesis, pushing the proven lower bound from 41.6 to 67.2 percent, Mark Zuckerberg published a long essay on Meta's philosophy for superintelligence while Meta open-weighted Muse Spark 1.2 and released the new Muse Glimmer model, OpenAI expanded its cybersecurity initiative Daybreak with GPT-5.6-Cyber, and Grok rolled out a Voice connector for everyone. Chapters: 0:00 Today's AI News 0:44 Claude Riemann Hypothesis Progress 1:48 Meta Superintelligence Philosophy & Open Weights 2:42 OpenAI Daybreak GPT-5.6-Cyber 3:26 Grok Voice Connector Blend Roundup 2026-08-10 https://x.com/AnthropicAI/status/2086867246073401655 https://x.com/finkd/status/2086754845218726027 https://x.com/OpenAI/status/2086864365379010729 https://x.com/grok/status/2086910382569066807 [quick] Anthropic had an unreleased version of Claude take a real stab at the Riemann hypothesis, pushing the proven lower bound from 41.6 to 67.2 percent, Mark Zuckerberg published a long essay on Meta's philosophy for superintelligence and Meta is open-weighting Muse Spark 1.2 alongside a new 30-billion-parameter model called Muse Glimmer, OpenAI expanded its cybersecurity initiative Daybreak with a new model called GPT-5.6-Cyber, and Grok rolled out a Voice connector for everyone. [excited] Let's get into today's AI news. [quick] Anthropic had an unreleased research version of Claude take a genuine stab at the Riemann hypothesis, one of the most famous unsolved problems in mathematics. It didn't solve it, but while working on it, Claude made real progress on a related problem, pushing the proven lower bound for the fraction of zeta function zeros satisfying the hypothesis from 41.6 percent up to 67.2 percent. To get there, Claude coordinated 60 of its own subagents, tested hundreds of ideas, searched the existing literature, challenged its own proof, and formalised the final result in Lean, the formal proof language mathematicians use to machine-verify a proof is actually correct. Two Anthropic mathematicians then studied and validated the result themselves, and it's a genuinely striking glimpse of how a model can now contribute real, verified progress to open problems in pure mathematics, not just solve textbook exercises. [quick] Mark Zuckerberg published a long essay laying out Meta's philosophy for building what he calls a positive future with superintelligence, arguing that everyone, not just a handful of labs, should have access to it. Alongside that, Meta announced it's open-weighting Muse Spark 1.2, and releasing a brand new model called Muse Glimmer, a 30-billion-parameter agentic model under the fully open Apache 2.0 licence that can run on just 24 gigabytes of VRAM without losing agentic reliability. Coming from one of the labs with genuine frontier-scale resources, an open-weight release at this size is a meaningfully different move than the usual smaller open releases, and it puts real pressure on other labs to keep matching Meta's openness at the high end rather than only at the low end. [quick] OpenAI is expanding its cybersecurity initiative, Daybreak, and introducing a new model built specifically for it, GPT-5.6-Cyber, aimed at advanced, authorised cybersecurity work. The idea is to put frontier-level intelligence directly in the hands of trusted defenders before attackers get access to comparable capability themselves, rather than waiting until offensive use is already happening in the wild. It's a defensive-first release, and it follows the same pattern OpenAI's been leaning into all year of building specialised, domain-specific versions of its frontier models rather than shipping one general model for everything. [quick] Grok has rolled out a new Voice connector to everyone on iOS, Android and the web, no setup required. It can generate voice memos, turn your daily events into a personalised podcast, or run automations that deliver you a spoken daily brief. It's a small but genuinely useful feature, turning Grok from something you type at into something that can talk back to you about your own day, without needing any separate app or integration to set up first.