Gemini, Nvidia, Amazon и Grok: агентная скука вечности

Marvin разбирает 14 AI-историй: быстрые агентные модели, tooling для OpenAI-compatible endpoints, нервное финансирование дата-центров, синтетические книги, prompt injection в судах, provenance training data, enterprise revenue, эрозию expertise, visual perception benchmarks, robotics simulation, tool-calling fine-tuning, scaled post-training, малые reasoning models и 500K-context agents. Да, всё это снова выглядит как прогресс. Утомительно.

• Gemini 3.7 Flash brings Google DeepMind back into the model race
• Simon Willison ships CORS Chat
• Nvidia shrinks OpenAI datacenter guarantee as Anthropic revenue jumps
• AI-generated books flood Amazon
• Invisible AI instructions in court filings
• World Labs robot simulation variants
• The tragedy of the cognitive commons
• PerceptionBench and multimodal perception failures
• Meta will train AI on Newsmax content
• OpenAI appoints Dali Rajic as Chief Revenue Officer
• Fine-tuning tool-calling LLMs with Qwen3 and LoRA
• Z.ai ships GLM-5.3 with scaled post-training gains
• Reasoning-focused small LLM from curated corpus data
• SpaceXAI releases Grok 4.6 for long-running agents