|
$ ai-news --date 2026-10-11 | mail
AI News — Sunday, October 11, 2026
// the most viral AI topics of the last 24 hours, curated for your stack
$ cat top-story.md
Open WebUI 0.12.0 adds realtime voice calls with animated 3D avatars, authenticator-app two-factor authentication, and shared chat replies so several users can collaborate in one conversation. Skills now support multiple files with version history and ZIP/JSON import-export, and the model editor and knowledge-base file browser were redesigned.
Why it matters: This is the frontend of your RAG stack: 2FA and per-model reasoning controls matter for multi-user deployments, and the knowledge base as a file browser makes managing RAG documents easier. Skim the release notes before upgrading.
// source: Open WebUI GitHub releases
## 🛠️ Tools & Repos
Open-source (MIT) single-user assistant that deploys to your own Cloudflare account with npx create-talorys@latest, with chat, memory, tasks, notes and scheduled reminders. It reached 266 points on Hacker News. Why it matters: A low-effort template for a private agent with no extra servers or telemetry. // GitHub / Hacker News
A policy-driven execution layer that contains agent workloads at process, session, WSL or MicroVM level, with file-system and network policies and Enforcement, Learning and Permissive modes. Why it matters: Shows where agent sandboxing is heading; Learning mode is a handy way to work out what an agent really needs. // Windows Developer Blog
## 🤖 Agents & MCP
Hooks get onFailure: "block", and prompt/agent hooks now block what they should. Subagents gain autoCompactWindow, the Read tool gets allow_large, and remote MCP connections no longer reconnect in tight loops after outages. Sonnet 5.5 cache-read pricing dropped to $0.10 per million tokens. Why it matters: The hook fixes close real policy gaps, and cheaper cache reads lower the cost of long agent sessions. // Claude Code GitHub releases
Over three months, agents decompiled a game to C++ and reached 99% function presence and 83% byte-exact matches, after the team added byte-matching verification as an objective pass/fail signal. Why it matters: The lesson carries over to any agent workflow: give agents an objective verification harness. // momo5502.com / Hacker News
## 📰 Worth Reading
Uses constrained decoding to make a small LLM (Qwen3-1.7B) pick one of a fixed set of options in a single pass. Accuracy was about 59% on CommonsenseQA, rising to 62% after fine-tuning, and temperature scaling corrects the model's overconfidence. Why it matters: A cheap pattern for routing or classification steps in a RAG pipeline with a local model. // nishtahir.com / Hacker News
Simon Willison built a newsletter index page for his blog almost entirely by voice while cooking dinner, working with an AI assistant. Why it matters: A practical example of hands-free agentic coding. // simonwillison.net
$ echo "automatically researched on 2026-10-11"
> ai-news · news.reichenberg.ruhr
[ unsubscribe ]
[ view in browser ]
|