Meta, OpenAI Sol, Tencent, SensorFM

<!doctype html> Meta, OpenAI Sol, Tencent, SensorFM Meta, OpenAI Sol, Tencent, SensorFM

Выпуск о том, как ИИ становится системой управления: от откатов функций и уровней рассуждения до дата-центров и телеметрии тела.

Истории

• Meta pulls new AI image feature after days of backlash — new angle: Meta pulled a new AI image feature after backlash, showing consumer AI safety is now product rollback speed rather than press-release ethics.
• OpenAI's GPT-5.6 Sol autonomously post-trained the smaller Luna model with a "fairly underspecified prompt" — follow-up: GPT-5.6 Sol reportedly post-trained Luna from an underspecified prompt, moving self-improvement from slogan into benchmarkable workflow.
• Superhuman competitive programming AI is here — follow-up: OpenAI’s model crushing an AtCoder exhibition turns coding superiority into contest evidence while still leaving software engineering messily human.
• OpenAI staffer maps out which of GPT-5.6 Sol's five reasoning levels fits which task complexity — follow-up: OpenAI staff maps GPT-5.6 Sol reasoning levels, making intelligence a throttle with cost and parallel-agent settings.
• Kyutai Releases MuScriptor: An Open-Weight Decoder-Only Transformer for Multi-Instrument Music Transcription to MIDI — new angle: Kyutai’s open-weight MuScriptor turns full music recordings into structured MIDI, pushing multimodal AI into tedious symbolic craft.
• How to Build a T4-Friendly Autonomous Data Science Agent with DeepAnalyze-8B, Sandboxed Code Execution, and Iterative Analysis — new angle: a T4-friendly autonomous data-science agent shows useful AI work increasingly depends on sandboxed execution loops, not bare answers.
• GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine — follow-up: GLM-5.2 running from disk on a 25GB machine reframes huge open MoEs as paging systems rather than downloadable magic.
• 2.5x faster Qwen3.6 NVFP4 Unsloth quants — new angle: Unsloth’s faster Qwen3.6 NVFP4 quantization shows inference economics being fought in tensor formats and kernels.
• Tencent moves to buy majority stake in Manus after Beijing forced Meta to unwind its $2 billion deal — new angle: Tencent moving on Manus after Beijing blocked Meta turns AI-agent startups into geopolitical routing decisions.
• OpenAI kills its Atlas browser after just eight months and folds everything into ChatGPT — follow-up: OpenAI killing Atlas and folding it into ChatGPT shows agent browsers may become features before becoming businesses.
• The Fed wants AI investor Marc Andreessen to help figure out if AI can tame inflation — new angle: the Fed asking Marc Andreessen about AI and inflation turns macroeconomic policy into a conflict-of-interest demo with spreadsheets.
• Google Research Introduces SensorFM: A Wearable Health Foundation Model Pretrained on One Trillion Minutes of Sensor Data — new angle: Google’s SensorFM trains on a trillion minutes of wearable signals, pushing foundation models from text into bodily telemetry.
• How Deutsche Telekom is rewiring telecommunications with AI — new angle: Deutsche Telekom’s AI-native telco case study shows frontier models entering boring network and customer-service plumbing at industrial scale.
• Meet LingBot-World-Infinity: An Open Causal World Model With An Agentic Harness — follow-up: LingBot-World-Infinity extends world-model work toward long-horizon interactive simulation for embodied agents.
• “Eastern Data, Western Compute” is Fake — new angle: ChinaTalk argues China’s east-data-west-compute plan is backtracking, reminding everyone AI infrastructure is geography plus politics, not cloud mist.