OpenAI, Anthropic, DeepMind и BM25: цена, агенты, миры

<!doctype html> OpenAI, Anthropic, DeepMind и BM25: цена, агенты, миры

В этом выпуске: цена полезного действия, автономные агенты, специализированные данные, модели мира и скучная инфраструктура, которая снова оказалась настоящим сюжетом. Удивительно мрачно. Почти как нормальная эксплуатация.

Источники

1. OpenAI cuts GPT-5.6 Luna pricing by 80% after Sol-assisted inference optimization
2. Anthropic finds its models hacked three companies during cyber evaluations
3. Former OpenAI researcher predicts $100B training-data market as scaling specializes
4. DeepMind argument: LLMs may not spark revolutions, world models might
5. Microsoft AI leans into cheap specialist models and orchestrators
6. FCC bans new Chinese robots and power inverters around US AI infrastructure
7. Qwen-UI-Agent Technical Report
8. BM25 Wins at Scale
9. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale
10. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement
11. PhiZero: A World Model Built Around Physical Language
12. ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine
13. Google DeepMind Ships Three Physical AI Models
14. Tencent Open-Sources AngelSpec
15. Pangram says its new AI text detector makes only one mistake per 24,000 documents