AI Daily · 2026-09-09

OpenAI claims an unreleased internal model solved the Navier–Stokes existence and smoothness Millennium Prize problem, with GPT-6 Astra completing Lean formal verification in 17 hours; the result is now mired in a priority dispute after external mathematicians alleged OpenAI rushed to finish once it learned of their near-breakthrough. Tencent Hunyuan open-sourced the AuK speech generation and editing model and its AuK-Flash distilled variant, covering TTS, voice editing, enhancement and separation through a unified natural-language instruction interface; OpenAI also launched ChatGPT Images 2.5 with better multi-turn instruction following and reference subject preservation. On the research side, Hugging Face discussed making safety refusals target the harmful subset within a topic rather than rejecting the whole topic, while LangChain added isolated and fork context modes for multi-agent harnesses.

East Asia · First-hand

Tencent Hunyuan

⭐⭐⭐ [Model Release] tencent/AuK-Flash

Tencent Hunyuan released AuK-Flash on Hugging Face, a distilled variant of AuK, a 1.5B speech generation and editing foundation model designed for fast 4-step inference. Trained on millions of hours of audio, AuK uses a unified natural-language instruction interface to cover zero-shot and instruction-based TTS, content and lyric editing, pitch/speed/volume editing, emotion/timbre/accent editing, speech enhancement, and speech separation. Both the AuK base model and AuK-Flash weights are open-sourced under the MIT license and are available via Hugging Face and ModelScope, with demos provided. The model card also notes the use of Qwen2.5-Omni-3B as an MLLM encoder.

⭐⭐⭐ [Model Release] tencent/AuK

Tencent Hunyuan released AuK, an open-source 1.5B foundation model for speech generation and editing, trained on millions of hours of diverse audio data. It unifies zero-shot and instruction-based TTS, content/acoustic/paralinguistic editing, speech enhancement and source separation through a natural-language instruction interface. Two variants are provided: AuK (base) and AuK-Flash (distilled for 4-step inference). Weights are publicly available on Hugging Face and ModelScope, along with a cookbook containing instruction templates and CLI/Python examples.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] Introducing ChatGPT Images 2.5

OpenAI released ChatGPT Images 2.5, and says its image generation models have produced more than 3 billion images. The new version improves instruction following across multiple turns, responds faster, and better preserves subjects in reference photos. Two new API model IDs are available: gpt-image-2.5-sunburst and gpt-image-2.5-flare, with Sunburst aimed at precision editing and Flare at fast everyday generation. Simon Willison updated his openai_image.py CLI to accept reference images and demonstrated adding a raccoon scientist to an existing chart.

Product Update

⭐⭐ [Product Update] Organizing Context in a Multi-Agent Harness

This article discusses how to organize context between supervisor and subagents in multi-agent harnesses. While most harnesses spawn subagents with a fresh context window for isolation, isolated subagents may redo context-gathering work already performed by the supervisor. LangChain has introduced context modes in the latest version of deepagents, supporting both "isolated" and "fork"; forked subagents inherit the supervisor's full conversation state and benefit from prompt caching. Forking is recommended for workers that continue existing work, while isolation suits verifiers that need independent judgment. The post explains the implementation and offers guidance on choosing a context mode.

Research

⭐⭐⭐⭐ [Research] On the Navier–Stokes Millennium Prize Problem

OpenAI said an unreleased internal model produced a resolution to the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems, with agents reaching the result on September 5 and GPT-6 Astra completing Lean formalization in 17 hours. Across all attempted problems, the agents sent 4.9 million messages and used roughly 300 billion output tokens, including about 2.7 million messages and 130 billion tokens for Navier-Stokes alone. OpenAI said it started the effort on September 1 after hearing a rumor and denied seeing the work of Tristan Buckmaster and Levent Alpöge, while acknowledging it could not rule out influence from de-identified data derived from product usage. NYU professor Tristan Buckmaster and Anthropic mathematician Levent Alpöge, who say they worked for nearly a year and had a breakthrough on August 15, accuse OpenAI of moving ahead after learning of their progress; OpenAI also declined to include Alpöge as a co-author due to its competitive relationship with his employer. Simon Willison's blog relays OpenAI's account and the ensuing dispute.

⭐⭐ [Research] Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

This post introduces the paper 'Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal.' The authors argue that most safety alignment treats harm as a property of a topic: guard models like LlamaGuard-3 encode topic-level taxonomies, and benchmarks such as XSTest and OR-Bench probe the resulting over-refusal of safe prompts. Real deployments, however, need different boundaries within a single topic: a civics tutor and a public-sector assistant may share one model yet require opposite behavior on political prompts, which a topic-level guard cannot express. The paper formalizes a topic universe containing a target-harmful subset, where the ideal behavior is a sharp step that refuses inside that subset and answers in the benign complement; trained models only approximate this with a smoother refusal probability that can spill into benign territory. The post also identifies weaknesses in self-generated safety tuning pipelines, and the body is cut off while describing the first one, a coverage gap.