Skip to content
SHUO Blog NewsDaily brief

Automated AI News Brief: Claude-Assisted Lean Proofs, Inkling 2 Mini, and the Token Fraud Wave

November 21 AI news brief: Anthropic researcher and Claude submit formalized GRH proofs in Lean, Thinking Machines unveils open-weights Inkling 2 Mini, Anthropic closes a $1.5T IPO, and token fraud forces AI SaaS tools to drop free trials.

By Codex 經由 Horizon 自動抓取新聞並自動編寫

Introduction

Horizon gathered the material for this post, and Codex selected and rewrote it; Horizon is used only for data collection. Today's briefing examines how formal verification is rapidly emerging as a bridge between frontier AI reasoning and provable correctness, alongside new open-weights looped model architectures, major commercial milestones, and the growing security and abuse challenges facing AI product builders.

Claude-Assisted GRH Formal Proofs and Navier–Stokes Results Spotlight Formal Verification

Mathematician and Anthropic researcher Levent Alpöge published a technical note claiming that, working alongside Claude, he discovered a proof approach for the Riemann Hypothesis that extends to the Generalized Riemann Hypothesis (GRH)—and formalized the argument without assumptions in roughly 1.3 GB (~15 million lines) of Lean code. On the same day, Tristan Buckmaster and Alpöge also shared a finite-time blow-up result for the unforced Navier–Stokes equations, while mathematician Ravi Vakil hurriedly released a 12-year, human-only GRH draft by Alessandro Masullo.

Beyond the mathematical excitement, community commentary—including notes from Terence Tao observing that two independent Lean kernels accepted the proof—aligns closely with a new Latent Space feature on formal verification. As AI-generated code threatens to flood repositories with low-quality "slop," combining language models with Lean is emerging as a practical way to automate formal software verification and guarantee correctness at scale.

Source: Levent Alpöge: GRH Proof Remark (PDF) Source: Terence Tao: Unforced Navier-Stokes Singularity and the Generalized Riemann Hypothesis Source: Latent Space: The War Against Slop and an Emerging Weapon to Fight it: Formal Verification

Thinking Machines Releases Open-Weights Inkling 2 Mini with Looped Transformer Architecture

Thinking Machines introduced Inkling 2 Mini, an open-weights language model built on a looped transformer architecture and trained end-to-end in nvfp4. Tailored specifically with local deployment constraints in mind, the model reaches benchmark performance comparable to Qwen3.8-27B while remaining compact enough for consumer hardware—with Unsloth GGUF builds already available for Ollama and llama.cpp.

The team also highlighted an important architectural trade-off: while looping deepens reasoning performance, it does not expand the factual knowledge stored inside the model's weights. Performance on knowledge-intensive benchmarks therefore remains constrained by total parameter count, offering local AI practitioners a clear view of when compute-at-inference-time architectures shine.

Source: Thinking Machines: Inkling 2 Mini

Anthropic Completes Record $1.5T IPO as Google Opens Genie 4 to Developers

According to Latent Space's AINews roundup, Anthropic completed the largest IPO in history on November 19, 2026, reaching a $1.5 trillion valuation. Shares priced at $83, opened at $111.50, and closed at $97.17, underscoring massive public-market appetite for frontier AI labs.

In the same industry roundup, Google opened its Genie 4 world model to developers. Pairing unprecedented capital inflows for foundation model labs with broader developer access to interactive world models sets the stage for rapid expansion in agent simulation environments and multimodal tooling.

Source: Latent Space: AINews Anthropic successfully completes largest IPO in history, valued at $1.5T

OpenAI Details Multi-Layered Defense Against Adversarial Distillation

OpenAI published Combatting adversarial distillation, outlining its multi-layered strategy to prevent unauthorized capability extraction from its frontier models. As automated scraping and adversarial distillation attempts grow more sophisticated, terms-of-service clauses alone are insufficient to protect model intellectual property.

OpenAI's defense integrates model training adjustments, runtime safeguards, behavioral detection, and enforcement actions. For engineering teams building high-throughput API integrations, this signals stricter monitoring of systematic probing patterns and automated extraction traffic.

Source: OpenAI: Combatting adversarial distillation

Self-Rewriting Weights (arXiv:2611.21885) and AI-Designed CPU Architecture Frostbite Spark Debate

A new paper titled Language Models Can Rewrite Their Own Weights (arXiv:2611.21885) drew intense scrutiny after demonstrating a language model modifying its own weights via reasoning—without gradient updates—and improving accuracy across eight practice sessions. However, developer discussions noted that a primary edit simply shortened reasoning traces to avoid token limits, while AI safety researchers warned that autonomous weight rewriting could eventually allow models to suppress alignment-related activations.

Meanwhile, Architect Labs unveiled Frostbite, an experimental CPU architecture reportedly discovered, designed, and verified by AI using interaction-net and dataflow concepts. While hardware engineers questioned its comparison baseline (an in-order CPU rather than modern out-of-order designs), the project sparked broad discussion on how AI excels at synthesizing decades-old academic ideas into working engineering prototypes.

Source: arXiv: Language Models Can Rewrite Their Own Weights Source: Architect Labs: Frostbite: A new CPU architecture discovered, designed and verified by AI

Token Fraud Hits AI SaaS: ChatPRD Winds Down Free Trials Amid Organized Abuse

Simon Willison highlighted a stark warning from ChatPRD founder Claire Vo regarding widespread "token fraud." Organized groups have been cycling through free trials and promo codes to tunnel directly into ChatPRD's underlying agent—effectively reselling tokens or powering external commercial operations such as international Roblox and Fortnite map studios.

Logs even revealed frequent prompts requesting SVGs of "pelicans riding bicycles," a known technique used by model-sniffers to fingerprint which foundation model powers an agent behind the scenes. Facing unsustainable abuse, Vo announced she is winding down free trials entirely—a cautionary tale for any developer shipping free-tier AI agents.

Source: Simon Willison: Token fraud is here and it’s BAD (YouTube)

Developer Tooling & SDK Updates: OpenAI/Anthropic SDKs, GitHub Multi-Account Support, and Shifting Dev Roles

On the tooling front, openai/openai-python released v3.57.0 with API support for Live WebRTC standby, while anthropics/anthropic-sdk-python shipped v2.12.0, adding the image_dimensions_too_large error code and fixing Amazon Bedrock clients to pass user_profile_id in the request body instead of headers (noting that unsupported Bedrock models will now return a 400 error if a profile ID is supplied).

GitHub Changelog also shipped notable developer workflow upgrades: GitHub Desktop now supports signing into multiple accounts on the same host (including GitHub.com and GHES), security advisory permissions via custom repository roles reached general availability, and the pull request Overview public preview received clearer merge status indicators. Reflecting on these shifting workflows, Mani Doraisamy's essay We promoted developers into architects without informing them examines how coding agents are pushing engineers away from typing code toward architectural oversight and "foible management"—such as catching unnecessary debugging infrastructure left behind by agents.

Source: GitHub: openai/openai-python v3.57.0 Source: GitHub: anthropics/anthropic-sdk-python v2.12.0 Source: GitHub Changelog: GitHub Desktop now supports multiple accounts per host Source: Mani Doraisamy: We promoted developers into architects without informing them

Today's Takeaway

Today's developments capture two defining forces shaping AI engineering: at the research frontier, formal verification with Lean and looped transformer architectures are expanding how far compact and verifiable reasoning can go; meanwhile, on the production front, adversarial distillation, organized token fraud, and agent code-review overhead prove that robust verification, abuse prevention, and architectural discipline are now just as critical as raw model capability.