Oct 23 AI Daily

xAI launches Grok Bot, AI teammates that sign into your tools and finish the work

xAI launched Grok Bot, AI teammates designed to take on real work: bots can sign in to your tools, use them like you do, take projects from start to end on desktop or iOS, and return for approval when needed. Bots run on encrypted cloud computers with training opt-out, sensitive actions can pass through Auto Review, and enterprise admins can configure DLP, certs, proxies, and network controls. Pricing has Premium, Super ($100/month with 3.5x usage), and Ultra ($200/month with 12x usage) tiers, with a macOS download available.

Also: TechCrunch · AI

Testing 341 European shops: OpenAI's ChatGPT ad pixel fires before consent by default

Testing 341 retail sites (mostly European) embedding OpenAI's ChatGPT ad pixel in a cookie-free headless browser, the author found 224 send a measurement event to OpenAI before any consent, with 95 firing while a cookie banner was still on screen. The cause is the pixel's default: OpenAI's documented install snippet contains only oaiq("init") with no consent call, and a separate oaiq("consent", false) call is needed to gate it; the author verified this control works, so the issue lies in the default rather than the SDK. The piece cites comparable fines such as CNIL's 150-million-euro penalty on shein.com's operator in 2025, and details the methodology, limits, and disclosure process.

Anthropic Donates Agent Skills Standard to the Agentic AI Foundation

Anthropic announced it is donating the Agent Skills open standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which it co-founded with Block and OpenAI and which already houses MCP. Since the standard's publication in December 2025, more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Agent Skills. A skill is a folder with a SKILL.md file plus scripts or reference material that teaches an agent a specific task or domain; the concept emerged from work on Claude Code, and within weeks of its October 2025 launch developers published thousands of skills on GitHub and wrote loaders for Codex, Cursor, and OpenCode.

Also: Simon Willison

Microsoft and Hugging Face release ThinkingBox-Training, turning database grading into RL rewards for agent post-training

Microsoft and Hugging Face released ThinkingBox-Training, an open-source stack that uses executable checks of an agent's final database state as GRPO rewards, applying 50 steps of full-parameter RL to Qwen3.8-27B. On the held-out 507-task ThinkingBox-Bench, pass@1 rose from 51.70% to 60.78% and tasks passing all 20 attempts grew from 38 to 107, with cost per dependable task falling from $24.36 to $8.00; pass@20 dipped slightly from 89.35% to 87.97%. The method's core is offline policy-relative difficulty: probing candidate tasks against the base policy and keeping those in the 0.08–0.90 success bands via a Beta posterior, yielding a 187-task training pool trained with token-mean GRPO on 24 H100s. Data selection, rollout, reward, and Verl integration code are open-sourced at microsoft/thin…

Microsoft Research releases Flint-interactive, a language for creating interactive data visualizations

Microsoft Research released Flint-interactive, a language built on top of Flint for creating interactive data visualizations, enabling agents to produce charts users can explore via clicking, filtering, zooming, highlighting, and cross-chart interactions. The Flint compiler resolves raw triggers such as clicks and drags into semantic events and applies high-level update operators, separating interaction logic from visual logic to cut generation cost and debugging effort. Users can try it in chat via Flint-MCP and MCP Apps, while developers can build dashboards, data stories, and bespoke interactions with Flint Interactive Canvas and the programming API.

AWS benchmarks low-latency speech streaming for AI agents with Kokoro TTS on SageMaker AI

An AWS Machine Learning blog post shows how to host the Kokoro-82M TTS model on a SageMaker AI real-time endpoint and stream sentences to it as an agent generates them, instead of waiting for the full response. In isolated tests this cut time to first audio byte from 329.3 ms to 156.7 ms, and end-to-end bidi mode was 3.5x faster than batch on a deployed AgentCore runtime. The post also covers base64 audio in SSE inflating payloads by about 33 percent plus MP3 frame alignment issues, and offers an alternative that multiplexes binary audio and JSON control frames over the same bidirectional session.