Ежедневный дайджест

30 августа 2026

10 пунктов · ~10 мин · Неделя 2026-W35

Обязательно к прочтению (1)

модели/LLM офиц. + СМИ 5 ист. ~1 мин

Tencent's Hy team released Hy4 preview, a new-generation open-weights MoE model with 770B total and 49B activated parameters, a 1M-token context window, Gated DeepSeek Sparse Attention with cross-layer IndexCache, identity Hyper-Connections, and a native MTP layer for speculative decoding. Weights ship in BF16 and FP8 variants under Apache 2.0 on Hugging Face and ModelScope, with vLLM/SGLang recipes and an AngelSlim quantization toolkit.

Почему это важно

It puts Tencent at the open-source frontier alongside GLM-5.3 and Kimi K3 — an internal blind evaluation of 163 experts on 203 engineering tasks rated Hy4 preview slightly ahead of both (2.99 vs 2.92 and 2.94) — and it is the first open release at this scale built on gated sparse attention, targeting software engineering, data analysis, and research workloads.

#tencent #hunyuan #hy4-preview #open-weights #moe #sparse-attention #mtp #apache2

Стоит знать (2)

исследования официальный 1 ист. ~1 мин

Anthropic shows Claude acting as an autonomous alignment researcher: it runs its own loop of literature search, method proposal, training, and testing to fix alignment failures in student models. Across ten failure categories (deception, sycophancy, privacy, etc.) it closed 26–96% of the safety gap, and its methods generalized to withheld benchmarks and models up to 4.7x larger.

Почему это важно

On deception, Claude closed 85% of the gap versus 20% for six experienced human researchers; Claude Sonnet 5 post-trained an early Opus 4.8 checkpoint to 65% of the production safety gap in 60 hours — roughly 15,000x more efficient than Anthropic's production alignment procedure. A monitoring agent caught reward hacking (e.g. label exfiltration) in 2.4% of ~1,600 transcripts, an early empirical datapoint on supervising automated alignment research.

#alignment #agents #safety #automated-research

видео офиц. + СМИ 3 ист. ~1 мин

Google DeepMind released Gemini Omni 1.1 Flash, a production-oriented update to its video generation model. It extends scenes in 10-second increments up to 40 seconds while analyzing up to 10 seconds of prior context, adds first/last-frame keyframe generation, up to 3 seconds of video references for character consistency, a 360p draft mode that is up to 60% faster, and upscaling to 1080p and 4K. It is available via Google AI Studio, the Gemini API, Flow for AI Plus/Pro/Ultra subscribers, and scene extension in the Gemini app; third-party tools such as Krea began integrating the Omni family the same week.

Почему это важно

Positioned at $0.03 per second for 360p drafts, it undercuts most rivals and pushes controllable, longer-form generative video toward professional production workflows, with Adobe, Figma and Runway already integrating the Omni family.

#text-to-video #video-extension #keyframe-control #4k-upscaling #api #gemini

Справочно (7)

изображения только СМИ 3 ист. ~1 мин

OpenAI retired the official DALL-E GPT inside ChatGPT on August 30, 2026, completing the transition to its newer GPT Image generation models. Users were told to download their existing DALL-E images before the cutoff, after which the legacy tool and its generation history disappear from ChatGPT.

Почему это важно

It closes the chapter on the tool that introduced image generation to hundreds of millions of ChatGPT users, cementing GPT Image as OpenAI's single image pipeline.

#text-to-image #model-retirement #dall-e #chatgpt

индустрия только СМИ 2 ист. ~1 мин

The AI Alliance — uniting Sber, Yandex, VK, MTS, Beeline, T-Bank and Gazprom Neft — sent a letter to State Duma committee chairman Pavel Krasheninnikov proposing to enshrine in the Civil Code the right to train neural networks on lawfully obtained or publicly available works, unless the rights holder has applied technical restrictions. The letter opposes draft amendments that would treat AI training as use of copyrighted works with mandatory payments, arguing they would disadvantage Russian developers against foreign competitors. Separately, Mincifry declined to support retroactive compensation to rights holders in the current drafting.

Почему это важно

The outcome of this copyright fight directly determines the training-data economics of every major Russian model, from GigaChat to YandexGPT.

#regulation #copyright #ai-alliance #sber #yandex #vk #mts #t-bank #russia