Anthropic's unreleased Claude pushed a 160-year-old Riemann bound from 41.6% to 67.2%
Anthropic says an unreleased research version of Claude raised the known lower bound for the proportion of Riemann zeta function zeros satisfying the Riemann hypothesis from 41.6% to 67.2%, a 25.6 percentage point jump. The process involved two Claude Code sessions, about 60 subagents working over a day and a half, and 31 million output tokens, after Claude first checked 54 existing papers to confirm the result was genuinely new. It produced a Lean proof that passed standard machine verification, and two Anthropic mathematicians independently checked the result. Anthropic is explicit that this does not get any closer to proving the Riemann hypothesis itself.
Why it matters: This is a case of an AI model extending an actual open problem in pure mathematics with a result mathematicians independently verified, not a benchmark score, which is a different category of claim than most "AI does math" stories.
OpenAI launches a cybersecurity-only model that stopped refusing security researchers' requests, and it's already found real vulnerabilities
OpenAI's new GPT-5.6-Cyber, available to vetted researchers through the higher-access "Daybreak Red" tier, completed 95% of requests on OpenAI's internal cybersecurity evaluation, versus 1.5-2% for the standard model. Researchers using it have already found two previously unknown Chrome V8 engine vulnerabilities (Google assigned CVE-2026-15903 and fixed them), at least five vulnerabilities in a major mobile OS, three critical database vulnerabilities, and more than 400 privilege-escalation issues in a widely used OS kernel.
Why it matters: A model tuned specifically to stop refusing legitimate security work, with real CVEs to show for it within days, is a concrete data point in the argument that AI can shrink the window between a bug existing and someone finding it, for defenders and attackers alike.
Nvidia lines up six Wall Street firms to raise $500 billion in financing for AI data centers
Nvidia signed a preliminary agreement with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion in third-party capital so hyperscalers, AI labs and enterprises can build data centers and buy Nvidia hardware. Nvidia itself isn't providing the money, it's the anchor customer whose chip demand the financing structure is built around; CEO Jensen Huang called AI compute an "investable asset class." Nvidia shares fell nearly 3% on the day amid concerns the structure resembles circular financing.
Why it matters: Moving AI infrastructure from "companies buy chips project by project" to "AI factories get financed like productive infrastructure" is a structural shift in how the buildout gets funded, and the stock reaction shows the market isn't fully sold on it.
Zuckerberg's new AI manifesto argues concentrated control of superintelligence is the real risk, backed by a $1B fund for communities near Meta's data centers
Meta published a 6,500-word manifesto from Zuckerberg titled "The Future is for Everyone," arguing that a small number of companies or governments controlling superintelligence is a bigger risk than any single rogue-AI scenario, and proposing "personal superintelligence" built for individuals rather than institutions. Alongside it, Meta announced the "Future Is For Everyone Fund," $1 billion aimed at schools, public services and first responders in US communities near its data centers, on top of a projected $145 billion 2026 capital expenditure budget for AI infrastructure.
Why it matters: This is Meta trying to reframe the AI-safety conversation around who controls the technology rather than what the technology can do, while using a nine-figure local-community fund to blunt the backlash its own data center buildout is generating.
AMD is buying a startup that etches AI models directly into silicon, claiming 48x Nvidia's inference speed
AMD agreed on 6 August to acquire Taalas, a Toronto startup whose chips skip loading model weights entirely by etching them permanently into the transistors, trading general-purpose flexibility for specialised inference speed. Taalas's HC1 chip runs Meta's Llama 3.1 8B at 16,960 tokens per second, a figure AMD says is 48 times faster than Nvidia's GPUs and 8.5 times faster than inference specialist Cerebras. AMD plans to fold the technology into its Helios rack-scale systems alongside its existing Instinct GPUs, with the deal expected to close in Q4 2026. It comes seven months after Nvidia's own $20 billion acquisition of inference chip maker Groq.
Why it matters: Both of the industry's biggest GPU makers now own a company built around giving up general-purpose flexibility for raw inference speed, which says the industry expects inference cost, not training capability, to be the next real competitive battleground.
AI Impact Hub