📰 News
-
Qwen-Image-2.1 says 7B beats the closed labs — Alibaba. Generate and edit. Native RGBA, up to ten reference photos. Runs on a 3090. Their own bench, independent numbers still pending. Research license; commercial use needs a separate ask. → The Decoder
-
Tencent’s Gander keeps talking while it works — A “cerebellum” handles the conversation second by second. A swappable “brain” can be Codex or Claude Code. Tests used a GPT-5.6. Interrupts users 8% of the time vs GPT-Realtime at 13.5%. Task accuracy lags. Weights coming. → The Decoder
-
Runway wants video as a live stream — Don’t prompt and wait. Talk, and the first frame is already moving. GWM-1. Small visual errors compound frame to frame. With Nvidia, first frame under 100ms. No ship date. → The Decoder
-
Daily AI use in the U.S. more than doubled in six months — Epoch AI × Ipsos. Almost every day: 8% → 19%. Once a week: 17% → 10%. August asked per product, so the real number may be higher. → The Decoder
-
Fake students have to make real mistakes — Microsoft and Illinois, StudentSim. Qwen3-4B. Beats a prompted GPT-5.4 at chess, English, and math. A chess tutor trained against it scored best with human masters. → The Decoder
-
Trump wants an AI Force and an AI czar — Modeled on Space Force. High IQ only. Claims AI could be 25% of GDP. No new rules, just existing criminal and civil law. Data-center pushback framed as a left-wing attack. → The Decoder
-
Your agent aced it. Will it do it again? — IBM. GPT-4.1 ReAct on AppWorld: Mean@5 77.4%, Pass^5 only 53%. A 24.4-point consistency gap. Flat decision points flip. Guidelines cut the gap in half. → Hugging Face
🔥 GitHub Trending
-
trycua/cua — Desktops for computer-use agents. Drivers, cloud desktops, benches. Code, APIs, and a GUI in the same task.
-
coder/coder — Self-hosted cloud dev environments plus agents. Terraform workspaces. No API keys inside the workspace.
-
higgsfield-ai/higgsfield — Multi-node GPU orchestration. ZeRO-3 and FSDP. Train models with billions to trillions of parameters.
-
docling-project/docling — PDFs, Office, email — ready for gen AI. MCP server. Runs locally.
💬 Community
-
Quoting voxium — Big company. Specs, code, tickets: all Claude Code. 12–13 hours of hitting Enter. L1 through L7. Shipping isn’t the bottleneck, so why are we slow?
-
openarm — Open-source 7-DoF humanoid arm for contact-rich work. Full bimanual system $6,500. CAD through Isaac Lab.
-
A Letter from a Machine Learning Engineer — Anonymous mail from someone who says they train frontier models. No moat. Open weights at Fable 5.1 in a year. The recipient can’t verify it either.