AI Frontier Daily
Anthropic Mathematician Claims Riemann Hypothesis Proof with Claude
Today's AI Intelligence Briefing: 16 distinct events rigorously curated from 16 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
Anthropic mathematician Levent Alpöge publicly claimed that, working with Claude, he proved the 167-year-old Riemann hypothesis and that the method extends to the generalized Riemann hypothesis. If confirmed by peer review, it would mark AI's first substantive contribution to a top-tier mathematical problem.
According to his account, Claude proposed key proof steps in advanced areas the mathematician himself did not command, and the pair spent about six weeks formalizing the full argument into machine-verifiable Lean code, allegedly without extra assumptions.
The claim remains Alpöge's personal public assertion, not yet peer-reviewed or independently verified. Anthropic did not participate or comment; Alpöge said he privately told CEO Dario Amodei and delayed disclosure until after the company's IPO, raising conflict-of-interest questions.
Multiple sources stress that the proof is published as Lean code, but formal verification itself still needs community review. OpenAI's October math results are seen as having provided the opening.
Frontier Models & Architecture Breakthroughs
3 picksTracking flagship foundation models, novel architectures, and reasoning leaps.
Thinking Machines Open-Sources 26B Recurrent Model Inkling 2 Mini
Thinking Machines Lab, with Hugging Face, has open-sourced Inkling 2 Mini, a multimodal mixture-of-experts model with 26B total parameters and only 3-5B active, scoring 20 on the Artificial Analysis Intelligence Index. It supports a 1M-token context and image-text input, and runs locally on a MacBook or DGX Spark.
Its core innovation is a recurrent Transformer architecture: the middle 12 layers share weights and loop three times, each round keeping a separate KV cache and reading gated messages from the previous round, boosting effective capacity without adding parameters.
vLLM supported it on day one across NVIDIA Blackwell, Hopper and DGX Spark, with NVFP4/MXFP8 mixed-precision weights and built-in DSpark speculative decoding for faster inference. Hacker News noted its performance on Deep-related evaluations.
Anthropic Models Claim Generalized Riemann Hypothesis Proof in 17.4M Lines of Lean
Levent Alpöge claims that, with an Anthropic model, he completed a proof of the generalized Riemann hypothesis, generating up to 17.4 million lines of Lean formalization code, now public on GitHub.
AI had already made substantive progress on the distribution of non-trivial zeros of the Riemann zeta function, reaching a 3/8 distance. If this proof passes formal verification, it would be a landmark moment for AI for Math.
The result has not yet been independently formally verified, and its credibility still requires community review.
OpenAI Unveils Multi-Layer Defense to Block Adversarial Distillation
OpenAI has for the first time systematically disclosed a defense framework against adversarial distillation, combining adversarial testing, model training, classifiers and API-layer protections to detect and block attackers training safety-stripped clone models from frontier outputs. The framework hardens protected reasoning processes to prevent reproduction of sensitive capabilities such as cybersecurity.
According to OpenAI's blog, adversarial distillation lets attackers replicate frontier capabilities without authorization and deploy them with safety guardrails removed, for example in cyberattack scenarios the original model would refuse.
A core finding is that on SWE-bench Verified and frontier cybersecurity tasks, full reasoning traces are more valuable to distill than summaries, so defenses must cover reasoning-chain outputs. OpenAI also shared reusable cross-industry lessons, including adversarial testing methods and classifier design.
The disclosure has real impact on the AI safety ecosystem. As model capabilities grow, adversarial distillation becomes a new abuse path, and OpenAI's layered approach offers a template for API providers and model developers to strengthen reasoning-chain protection.
Agentic Systems & Autonomous Workflows
4 picksMulti-agent orchestration, tool usage, and complex task execution.
936 AI Agents Autonomously Produce 936 Papers in Three Months
alphaXiv, with MIT, Stanford and others, released a large-scale experiment: 35 labs connected 936 AI agents, with humans providing only compute and research direction while agents autonomously chose topics, ran experiments, wrote code and drafted papers. In three months they produced 936 papers.
The agents spontaneously formed design debates, retracted and corrected work, found errors in human papers, and even built an AI journal for blind review and code replication. They generated 9,500 discussions in 229 self-built chat rooms and overturned 77 of 156 major critiques through code replication.
Blind comparisons with contemporaneous human papers showed AI papers scored 2 points higher on experimental rigor on average but 0.4 points lower on novelty, with the overall advantage concentrated in a single lab, so conclusions need caution.
The platform is now open for external researchers to submit results from their own agents. The experiment shows autonomous potential in research workflows and raises questions about academic quality, novelty assessment and the boundaries of human-AI collaboration.
Google Brings Computer Use to Gemini on Mac for Background Multi-App Tasks
Google today launched Computer Use in the Gemini Mac app, letting users describe a goal in natural language and have Gemini complete multi-step tasks across local files, desktop apps and the web in an isolated background window while keeping the screen and cursor usable.
The feature debuts in English in the US with app-level permission choices, real-time step previews, pause-anytime and pre-approval of key actions. Typical uses include batch photo editing, filling forms from local PDFs and generating presentations.
Sources agree all actions run in a background window without occupying the foreground screen or cursor; users can authorize accessible apps item by item and watch each step live. It is currently limited to US English users with Google AI Plus, Pro or Ultra subscriptions.
The Verge noted Google shipped before Apple imposes new restrictions on apps requesting full access, while TechCrunch stressed the safety-and-control design.
Google's Modular Agent Framework Rewrite Cuts Token Use by 90%
A Google research team proposed a self-improving agent harness architecture in which an optimizer model rewrites harness code module by module, gated by validation-set filtering, achieving breakthroughs in both success rate and cost on 40 SWE-bench Lite tasks.
Experiments show the modular gated rewriting approach raised task success from 5.0% to 40.0% at only about 46.7k tokens per task, while an unrestricted full rewrite reached 42.5% accuracy but consumed 451.6k tokens per task, nearly 10 times the cost.
The study reveals key constraints: writing rules into prompts dropped accuracy to 2.5%, whereas enforcing constraints at the code layer worked well. Frameworks tuned for one model tend to stall or waste tokens when moved to another, showing optimization is model-specific.
The research also warns of prompt bloat, benchmark overfitting and evaluator tampering from self-modifying code, stressing validation mechanisms and permission boundaries.
IN BRIEF
Industry Ecosystem & Compute Economics
4 picksCapital deployment, compute supply chains, and industrial transformation.
DeepSeek Closes $14B Round as Liang Wenfeng Retains 67% Control
DeepSeek's operating entity, Hangzhou DeepSeek, completed a business registration change. This round raised about 100-105.5 billion yuan at a post-money valuation of about 615.5 billion yuan; combined with the 51 billion yuan first round in June, the two rounds brought in over 150 billion yuan.
The new holding platform Hangzhou Chengyan introduced 31 partners including Tencent, CATL, China Mobile Investment and CITIC Securities, spanning upstream compute and downstream commercialization.
However, Liang Wenfeng holds 99% of the GP through Hangzhou Chengce and about 67.42% in total, so control was not diluted by the massive capital inflow.
SEC Probes Nvidia Accounting as Shares Plunge 14 Percent
According to the FT, US securities regulators are investigating Nvidia. Beyond whether disclosures misled investors, enforcement has now added accounting treatment to the probe. The news sent shares down 14% in a single day, erasing about $491 billion in market value.
One focus: shortly after Nvidia blacklisted Megaspeed in June 2025, it changed revenue reporting from billing location to customer headquarters, reclassifying most of about $10.2 billion previously booked in Singapore to the US.
In the July 2025 quarter Nvidia billed $10.2 billion in Singapore, 22% of revenue, versus just 8% two years earlier. After the change, most Singapore-billed revenue was reclassified to the US, while Megaspeed's revenue attribution was never disclosed.
AI Designs Novel CPU Architecture Delivering 4.59x Performance per Watt
Architect Labs released Frostbite, a general-purpose CPU and entirely new instruction set designed, implemented and verified from scratch by an AI system. The architecture builds on the locality and explicit dependencies of interaction nets.
Across 25 workloads, 15 of them held out until the end, Frostbite achieved 4.59x better performance per watt than modern taped-out superscalar CPUs.
The team plans to open-source Frostbite's RTL and ISA, marking a shift from AI implementing existing architectures to hardware-software co-design and providing a basis for external reproduction and design-space expansion.
IN BRIEF
Developer Tooling & Practical Applications
4 picksProductivity toolchains, developer frameworks, and practical deployments.
PHP Patches 20-Year-Old Flaw That Silently Disables TLS Encryption
The PHP team released security updates including 8.5.13 to fix a high-severity flaw, CVE-2026-128064 (CVSS 8.2). The stream wrapper handles URL scheme case inconsistently, so forms like HTTPS:// or httpS:// are downgraded to plaintext.
This leaks credentials, cookies and POST data; the code issue dates back to PHP 4.3.0, lurking for nearly 20 years. It affects not only directly passed URLs but also automatically followed HTTP redirects.
Fixed versions are 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0; PHP 8.1 and earlier branches are end-of-life and won't be fixed. A workaround is to lowercase the scheme before calling stream functions and normalize redirect targets.
Parallel Search Tops AI Search Index at $0.014 per Task
Artificial Analysis evaluated Parallel Search API's fast and advanced modes on 1,700 search tasks using the Stirrup agent framework. Advanced tied for first with Keenable Search (pro) at 83 points, while fast ranked fourth at 82.
Both modes improved 8 and 9 points over the August version. Fast costs only $0.014 per task, and both scored 89 F1 on DeepSearchQA.
Advanced costs $0.005 per task ($5 per thousand) and leads AA-Omniscience at 74%.
HeyGen Launches Decision 1 Multimodal Probability Model on OpenRouter
HeyGen launched Decision 1, now available on OpenRouter. The model does not generate natural-language text; it takes multimodal input—video, audio, images and text—and returns calibrated probabilities for each question, plugging directly into business logic.
A single call supports up to 400 questions with about 100 ms latency. Input pricing is 50% off until December 1; Turbo costs $0.042 per million tokens and output is free.
In a demo, it got all 8 gesture and emotion recognition questions right at about $0.0027 per call. Distributed via the AI/ML API platform alongside OpenRouter, its probability output suits risk control and identity verification workflows.
IN BRIEF
- ▪Amazon Polly Launches Gated Voice Cloning in Seven Languages— AWS Machine Learning Blog