Jon Markman

OpenAI's Astra model has reportedly solved ten complex, decades-old mathematical problems for just $2,000, marking a significant advancement. These solutions, spanning diverse fields like group theory and high-dimensional geometry, are notable for their accompanying machine-checkable proofs. Published in Lean on GitHub, these certificates allow instant, trustless verification, fundamentally altering the traditional peer review process. While critics question problem selection and model access, the verifiable nature of Astra's output addresses a key challenge in AI deployment. This breakthrough highlights a shift where the cost of discovery is minimal, and the ability to guarantee correctness becomes paramount, enabling AI's broader application in critical industries.

Microsoft's new in-house cyber AI handles 90% of security tasks, cuts costs by half, and reduces frontier models like GPT to an escalation role.

China has finished the world's largest fusion magnet, bringing its artificial sun closer to reality and raising the stakes in the race to power AI.

Microsoft shifts most security scanning to an in‑house AI model, cutting costs and routing only the hardest work to frontier systems like OpenAI and Anthropic.

Alibaba has launched Qwen3.8-Max, a 2.4 trillion parameter AI model, at a significantly lower price point than competitors like OpenAI and Anthropic. Priced at $2 per million input tokens and $6 per million output, it undercuts even mid-tier offerings, initiating a fierce margin battle in the AI market. This aggressive pricing strategy aims to make AI more accessible, driving demand for compute resources. While Alibaba's internal benchmarks claim near-frontier performance, independent evaluations are still pending. The company's goal is to use the model as a funnel to boost its profitable cloud services, where AI-related revenue is surging. This trend of falling AI prices, though accelerating, is seen as an expansion of the market, benefiting infrastructure providers like Nvidia and hyperscalers as AI becomes embedded everywhere. The industry awaits Western labs' response to this pricing challenge.

Nvidia has open-sourced its cuFile APIs and storage software stack, with Google, Intel, and Meta joining as founding maintainers in a new GitHub organization. This initiative enables GPUs to directly manage and access storage, bypassing the CPU and building their own data requests, dramatically accelerating data transfer. Micron demonstrated 230 million random reads per second, showcasing significant efficiency gains. While citing security and Linux interoperability, Nvidia's strategic goal is to establish an industry standard that competitors must adopt. This "Keyser Söze" strategy, reminiscent of CUDA, sees Nvidia giving away software components to solidify its hardware's central role by shaping the entire ecosystem.

Nvidia open‑sources its cuFile storage stack as Google, Intel and Meta join a new industry group shaping GPU‑controlled data access and accelerated storage.

Meta has launched Muse Code, an AI coding agent in beta, designed to handle entire engineering tasks from planning to checking code. It features a unique pricing model with two tiers: a standard rate and a "contributor" tier. The contributor tier offers significantly lower costs in exchange for Meta using user prompts and completions to train its AI models. This strategy allows Meta to acquire valuable, real-world coding data, which is scarce and superior to scraped or synthetic data due to its checkable nature. The move reflects Meta's historical approach of leveraging user-generated data for its services, now applied to developers. This launch comes as Meta faces substantial capital spending, highlighting its commitment to acquiring high-quality training data for AI development, particularly in a domain like coding where automated validation makes agents highly effective.

Anthropic is establishing an in-house team to design custom chips for its Claude AI, joining other major labs in pursuing hardware co-design for efficiency. This strategic move, driven by Anthropic's exponential growth and $30 billion revenue, aims to optimize performance and reduce the cost per token by tailoring silicon to specific workloads. However, this initiative doesn't signify an "escape from Nvidia" or the broader chip industry. Anthropic explicitly maintains a multi-chip strategy, partnering with AWS, Google, Nvidia, and AMD, and has already secured massive future TPU capacity from Google and Broadcom. Custom chip programs primarily redistribute spending within the existing chip layer, still relying on established companies for manufacturing and components, rather than exiting the supply chain.

Apple has the iPhone 18 silicon ready, but a DRAM shortage is delaying production as AI-driven memory demand tightens global supply.