AI 日报 — 2026-07-13

A recent PDF, "GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture," has been shared online. The document, accessible via an OpenAI URL, claims to present a proof generated by an AI model. This development has garnered some attention, with 8 points and 1 comment on a Hacker News thread.

The author questions the practice of fine-tuning models on summarized or censored SOTA CoT traces. They observe a belief that distillation can magically improve output quality beyond the base model's capabilities. Specifically, the author finds Fable fine-tunes perplexing, as they seem to overlook that Anthropic's reasoning traces differ significantly from the model's actual chain of thought, likely leading to degraded results.

A developer successfully ran Qwen3.5-122B on a Mac Studio by fixing three bugs in their qMLX serving stack. These bugs, including prompt instability, interrupt path issues, and checkpoint poisoning, caused significant delays in long-context inference. After the fixes, prefill times dropped from minutes to sub-seconds. The developer open-sourced their qMLX fork, optimized specifically for Qwen, and a benchmark script to help others with similar hybrid attention caching issues.

Apple has sued OpenAI, alleging trade secret theft by former Apple employees now working at OpenAI. The lawsuit claims individuals like Tang Tan and Chang Liu stole confidential information, including unreleased technologies and product designs. Apple states Tan used insider knowledge to interview candidates, directing them to bring Apple hardware and revealing project codenames. Liu allegedly downloaded thousands of pages of technical files. Apple seeks injunctive relief and damages, asserting OpenAI ignored initial concerns.

Prismata is a defense mechanism designed to secure autonomous web agents against cross-site prompt injection attacks. These attacks exploit agents' interpretation of natural language, allowing malicious content to hijack tasks. Prismata enforces "contextual least privilege" by dynamically labeling page content and restricting agent capabilities, inspired by integrity models. It redacts content and limits agent actions without requiring developer annotations. Prismata significantly reduces attack success in various web agent attacks while maintaining utility for legitimate tasks.

Apple has sued OpenAI, alleging trade secret theft and breach of contract. The lawsuit claims OpenAI's senior leadership, including Chief Hardware Officer Tang Tan, directed former Apple employees to steal confidential information. Accusations include Tan using Apple's project code names, asking candidates for hardware components, and coaching employees to evade security. Apple also alleges another former employee, Chang Liu, downloaded confidential documents. Apple seeks to bar OpenAI from using its trade secrets and demands the return of confidential materials.

Tang Jie, founder of Z.ai, the Chinese lab responsible for the GLM models, advocates for keeping frontier artificial intelligence capabilities "as open and widely accessible as possible." He believes advanced AI should remain broadly accessible.

OpenAI reportedly remains confident in its timeline to unveil its first device in 2026 and release it in 2027. However, a lawsuit filed by Apple, alleging systematic intellectual property theft, could complicate OpenAI's plans. This legal challenge may specifically impact the company's ability to hire new talent and manage its supply chains, potentially hindering the device's development and launch.

Anthropic announced on X that it is extending access to Claude Fable 5 for all paid plans. Additionally, the weekly rate limits for Claude Code will remain 50% higher. These changes are effective through July 19, according to the company's statement.

Cursor is reportedly developing a general-purpose AI agent, codenamed Sand, to compete with tools like Anthropic's Claude Cowork. This new AI is designed for non-developers and will manage various tasks, including handling emails, texts, and documents. The initiative suggests Cursor's expansion into broader AI applications beyond its current offerings.

Waze is integrating Google's Gemini AI to enhance user experience with new features. Drivers can now use conversational voice commands to report traffic incidents and suggest map updates. A new Destination Search allows users to find locations like coffee shops or gas stations using voice. Additionally, Waze is introducing a "less chatty" voice prompt mode, a Motorcycle Mode with specialized routing, and personalized route suggestions based on past trips and traffic patterns.

BuzzRadr users are discussing recent changes to AI services. The "Chad Sol Launch" is praised for its user-friendly approach, allowing resets and returning with 50% of a user's quota, and quickly reaching six million users. In contrast, the "Virgin Fable Rollout" is criticized for being overly cautious and affecting only a small number of users. Additionally, "Claude Code" is noted for consuming "Chat's tokens," while ChatGPT and Codex now have separate usage. Plus users are reportedly benefiting from the removal of a 5-hour limit, despite some paying $200 for Max and still hitting maximums.

A developer successfully ran the Gemma 4 LLM directly within Godot 4.7 using only GDScript and Vulkan compute shaders. This experimental project loads the gemma-4-E2B-it-Q4_K_M.gguf model locally, with Vulkan handling calculations and GDScript managing other LLM functions and the UI. While 10x slower than llama.cpp with CUDA and limited to one model, it demonstrates the feasibility of running an LLM entirely within Godot without external dependencies.

A new article aims to simplify the understanding of world models within the ML community. The author proposes a classification framework for different world model approaches and identifies emerging trends. Feedback is requested on the framework's completeness, clarity, and technical accuracy.

A developer is building an "agentic Alexa" with voice and touchscreen control, featuring low-latency responses and task-based model selection. This system can manage emails, calendar invites, create presentations, and code, integrating with hundreds of apps like Notion and Discord. It boasts a custom long-term memory system and planned computer vision capabilities for real-time assistance. The developer seeks feedback on its viability and potential collaborators.

OpenAI, Meta, and SpaceXAI are poised to challenge Anthropic by highlighting cost efficiency. This comes as businesses are increasingly scrutinizing their AI expenditures. These three prominent AI developers recently released new, more advanced models, suggesting a competitive shift in the AI market where cost-effectiveness will be a key differentiator for attracting business customers.

Anthropic, an AI startup, has hired Tom Blomfield, a co-founder of Monzo and a prominent figure in UK tech. Blomfield will join Anthropic's compute team, taking a leave of absence from YC. This move highlights the ongoing competition for AI talent and Anthropic's continued recruitment of high-profile individuals.

PrismML, a Khosla-backed startup, announced a breakthrough by shrinking Alibaba's 27-billion-parameter Qwen 3.6 AI model to run on an iPhone 17 Pro. This compressed version, under 4GB, retains full functionality for complex chat, reasoning, and coding, with all parameters active simultaneously. The open-source model will be available next Tuesday, potentially shifting AI processing from the cloud to devices and changing AI economics.

A letter titled "We Must Act Now," signed by nearly 200 economists, including 15 Nobel laureates and Anthropic's Jack Clark, warns of rapid AI-led job displacement. The signatories urge policymakers to better understand and respond to potential disruptions caused by artificial intelligence.

BuzzRadr users are actively discussing GPT-5.6, a new model from OpenAI. The conversation centers around its deployment safety, as detailed in a provided PDF, and its technical specifications, available through the OpenAI API documentation. With 317 points and 196 comments, the community is deeply engaged in analyzing the implications and capabilities of this latest iteration.

Open weight AI models are facing a critical policy test in the US, potentially relegating them to a permanent "second-class citizen" status. Anthropic is spearheading a campaign against Chinese models, citing concerns about distillation. This policy action could significantly impact the future and accessibility of open AI models.

Users of Codex may have noticed the removal of a 5-hour usage limit. This change is not a bug, but an intentional update. The previous restriction on usage time has been lifted, indicating a potential shift in how the platform manages user access and engagement. This development suggests a more open or flexible approach to user interaction with Codex.