Astra Publishes Ten Math Proofs For $2,000.

Mazen Al Sabbagh

— 4 min read
Share

TL;DR

  • OpenAI's Astra: The next-gen model debuts by solving ten long-standing mathematical problems for an estimated $2,000 in compute.
  • EU Act enforcement: The European AI Act's general-purpose model provisions become active today, enabling fines up to 3% of global revenue.
  • US framework stalled: Washington's self-imposed deadline for its frontier-model framework lapsed without action, contrasting sharply with European developments.
  • Big Tech capex scrutiny: Recent earnings reports rewarded Microsoft and Amazon for demonstrating ROI on AI spend, while Meta and Alphabet faced market skepticism.
  • DeepSeek's new agent: DeepSeek released V4-Flash, a swift, open-weight agentic model, further cementing China's competitive position in cost-efficient inference.

Lead Story: Astra Publishes Ten Math Proofs For $2,000

OpenAI's formal introduction of Astra, its next-generation model, arrived not via product launch or pricing announcement, but through a 249-page manuscript and accompanying Lean 4 proof certificates on GitHub. The company asserts this unreleased system independently solved ten open mathematical problems, each having eluded human resolution for over a decade, at an estimated compute cost of $2,000 using GPT-5.6 Sol API rates.

The salient achievement is the explicit construction of a non-sofic group, addressing a problem first posited by Mikhail Gromov in 1999. Astra additionally disproved Connes's rigidity conjecture, proved Ehrhart's volume conjecture, and resolved three problems from Paul Erdős’s collection. The remaining solutions span sphere packing, coding theory, quantum complexity, and lattice cryptography. Critically, the Lean repository's "sorry" count is zero, indicating every formalized step is machine-verified, removing any dispute regarding correctness.

The broader significance, however, remains under evaluation. While OpenAI's math-research lead, Sébastien Bubeck, described the proofs as "beautiful," and Thomas Bloom of erdosproblems.website considered them superior to OpenAI's previous unit-distance counterexample, Simon Willison and others observe that machine-checkability confirms accuracy, not necessarily conceptual novelty. That determination rests with specialized mathematical communities, and their verdict is still pending. Astra's architecture emphasizes long-duration, multi-agent coordination, a capability profile that previously posed internal containment challenges for OpenAI. No release timeline, commercial terms, or branding have been disclosed. Notably, this model also lacks a federal safety review, as the framework intended to govern such frontier AI missed its operational deadline yesterday.

In Other News

EU AI Act enforcement begins. Effective today, the European Commission's AI Office is empowered to investigate and enforce compliance for general-purpose model providers. This authority includes demands for documentation, technical evaluations, mandated risk mitigation, market restrictions or withdrawals, and fines up to €15 million or 3% of global turnover, with non-cooperation incurring its own penalties. Article 50 transparency requirements also activate today, necessitating user disclosure for machine interaction and machine-readable labels for synthetic media. The AI Office intends to commence with "technical compliance dialogues," which they indicate "may intensify."

US frontier framework stalls. Executive Order 14409 set an August 1 deadline for federal agencies to establish a voluntary "early access" framework for frontier models. This deadline lapsed without any public deliverables—no Federal Register notice, NIST or CISA guidance, nor an OSTP statement. While a draft was reportedly circulated to major AI developers, suggesting NSA and federal AI safety institute oversight, it remains unsigned. The contrast with Brussels' swift regulatory action this weekend is stark: Washington's alternative framework has yet to materialize, leaving advanced models like Astra without a clear U.S. review process.

AI capex now tied to demand. The recent Magnificent Seven earnings calls, concluding July 31, revealed a clear market bifurcation based on return on AI investment. Microsoft's stock rose 8% and Amazon's approximately 10%, driven by AWS's 37% growth to $42.2 billion, its fastest in 18 quarters. Conversely, Meta and Alphabet saw share price declines, failing to adequately justify their extensive infrastructure buildouts, while Apple noted softness in services and supply. The four hyperscalers are collectively projected to invest approximately $725 billion in 2026 AI capex, a 77% year-over-year increase, which the market will now critically re-evaluate quarterly.

DeepSeek drives cost-efficiency in agents. DeepSeek launched V4-Flash on July 31, a rapid, open-weight model optimized for agentic workflows. This release reinforces China's consistent delivery of high-capability, low-cost AI, already resulting in a significant shift of U.S. firms' inference workloads to Chinese models. The launch coincides with ongoing, yet unaddressed, U.S. policy debates regarding restrictions on domestic use of Chinese open-weight models, even as Kimi K3's weights circulate freely.

X / Social Pulse

The immediate reaction to Astra was led by Sébastien Bubeck’s concise "beautiful" and Thomas Bloom's endorsement of "big news," lending initial credibility from the Erdős-problems community. Gary Marcus, however, quickly introduced skepticism, highlighting that the reported $2,000 compute cost omits the significant human formalization and curation required for such proofs. Terence Tao offered a broader perspective, characterizing the event as "big mathematics", emphasizing a decentralized human-machine collaborative future over direct machine replacement. The overarching sentiment across discussions was a collective "Deep Blue moment" for mathematics, though consistently tempered by the reminder that peer review, not mere machine verification, remains the ultimate arbiter of conceptual novelty.

One to Watch

The governance divide. The current regulatory void defines the immediate landscape. Europe's AI Office now possesses the authority to impose fines and restrict models; the U.S. presents an expired deadline and an unfinalized draft framework. The next significant model release—whether Astra, Grok 4.6, or a Chinese frontier system—will determine whose regulatory mechanisms, if any, prove effective first. Focus on the EU AI Office's initial "compliance dialogue" targets and any delayed, quiet publication of the U.S. frontier framework by OSTP.

Quick Hits

  • OpenAI pricing: GPT-5.6 Luna inference prices were reduced by approximately 80% on July 30, sharpening OpenAI's competitive edge in the escalating inference cost landscape.
  • Nvidia-OpenAI data center deal: Nvidia is reportedly negotiating to underwrite up to $250 billion for OpenAI's lease of SoftBank's 10-gigawatt Piketon, Ohio campus.
  • Hugging Face's demands: Clément Delangue of Hugging Face awaits OpenAI's response to his demand for $100 million in compute credits and full agent-trace disclosure following last month's security breach.
  • California AI Transparency Act: California's SB 942, the AI Transparency Act, takes effect today, mandating large model providers offer complimentary AI-detection tools.
  • Google Earth feature rollback: Google withdrew a new Google Earth image-generation feature within 24 hours of its July 31 launch, citing user feedback.

This weekend crystallized the current paradox of AI governance: a machine generating verified, novel mathematics juxtaposed with two major governmental powers unable to align on regulatory oversight. The relentless progression of AI capabilities demonstrates no deference to bureaucratic timelines.

Sources

Lock in. M. mazen@thorterminal.com

Read more