Перейти к содержимому

Microsoft Just Gave Your PC an AI Team (We Tested It)

AIofThings

0:00 / 0:00

Microsoft Just Gave Your PC an AI Team (We Tested It)

192 просмотра · 2 недели назад
AIofThings
5 подписчиков
192 просмотра · 2 недели назад
Microsoft just gave your PC an AI team. At its Oct 7 event, Microsoft rebuilt its Windows pitch around "hybrid intelligence": AI agents that run on your own PC when they can, use the cloud when they must, and get their own locked-down account. Then we ran our own AI agent team, locally, for $0. The PC is about to stop being a machine you do everything on yourself — and start being a team that works for you. QUICK FACTS (checked Oct 12, 2026) Microsoft Execution Containers (MXC): generally available on Windows 11; per-agent containment enforced at runtime, plus separate agent identities. Copilot on Copilot+ PCs gets local context, local actions and local models — expected to roll out "in the coming months." Surface Laptop Ultra (Nvidia RTX Spark): from $2,599, Oct 16; the 128 GB version that runs 120B+ models is about $5,899. MAI Code 1.1 Flash runs locally at ~3.3 bits: 70.8% vs 72.6% on SWE-bench Verified (Microsoft's numbers). Microsoft-Decision-1: a 9B decision model, $0.042 per million input tokens, output free; 35x faster than GPT-6 Sol in Microsoft's benchmark. OUR TEST 24 made-up tax files sorted by a local Qwen3-4B model (llama.cpp, 8 CPU cores, no GPU, $0). One agent: 19.6 s, 18/24 Four parallel subagents: 18.7 s, 18/24 Four subagents + three sentences of shared context: 24.1 s, 23/24 One shot each, no edits. The lesson: a team only beats one agent when every helper sees the same picture. CHAPTERS 0:00 Your AI agent gets its own account 0:47 Microsoft's Oct 7 event 1:49 AI has a bill problem 3:06 Copilot gets hands (the Autopilot tax demo) 4:42 MXC: a lock for every agent 6:45 RTX Spark + Surface Laptop Ultra 8:35 MAI Code 1.1 Flash + HydraFusion 9:49 Microsoft Decision-1 11:13 Who pays when your agent messes up? 12:04 Part 2: one AI becomes a team 13:32 Multi-agent AI is already in products 14:42 We ran our own AI team test 16:21 Why it isn't magic yet 17:36 Your PC as a team you manage 18:23 My take 19:12 Your turn Benchmark numbers are the companies' own unless marked independent. Would you give an AI agent its own account on your computer? What would you let it do — and what never? Tell me in the comments. SOURCES Building Windows for hybrid intelligence (Microsoft): https://blogs.windows.com/windowsexpe... Microsoft Execution Containers: https://blogs.windows.com/windowsdeve... Surface Laptop Ultra pre-orders: https://blogs.windows.com/devices/202... Local models + sandboxed tools (MAI Code 1.1 Flash): https://commandline.microsoft.com/loc... Microsoft-Decision-1: https://commandline.microsoft.com/mic... Nvidia newsroom, RTX Spark: https://nvidianews.nvidia.com/news/nv... Nvidia blog, Oct 7 event: https://blogs.nvidia.com/blog/local-a... GitHub HydraFusion: https://github.blog/ai-and-ml/github-... Copilot Home, Code and Autopilot: https://blogs.microsoft.com/blog/2026... Windows Latest on the Copilot+ PC brand: https://windowslatest.com/2026/10/08/... TechRepublic, Surface Laptop Ultra configurations: https://www.techrepublic.com/article/... Nadella + Huang conversation (Windows):    • Satya Nadella and Jensen Huang discuss Hyb...   Full Oct 7 event (Windows):    • Something new is coming from Windows | Oct...   Satya Nadella on X (Oct 7): https://x.com/satyanadella/status/210... Satya Nadella on X (Oct 9, Decision-1): https://x.com/satyanadella/status/210... Microsoft Surface on X: https://x.com/surface/status/21078879... Microsoft AI on X: https://x.com/MicrosoftAI/status/2107... Anthropic, multi-agent research system: https://www.anthropic.com/engineering... Google Cloud, Gemini agent: https://cloud.google.com/blog/product... Microsoft Research, Magentic-One: https://www.microsoft.com/en-us/resea... Why Do Multi-Agent LLM Systems Fail? (MAST): https://arxiv.org/abs/2503.13657 Cognition, Don't Build Multi-Agents: https://cognition.com/blog/dont-build... METR, Time Horizon 1.1: https://metr.org/blog/2026-1-29-time-... OSWorld 2.0: https://osworld-v2.xlang.ai/ #Microsoft #AIagents #SIDecoded