Gemini recap: sub-agents; Claude Dashboards, Motion are in beta

#1 𝕏

Sundar Pichai recapped the introduction of a new Gemini agent at a Gemini at Work gathering with 700+ Google Cloud customers at NASA Hangar One. He described it as a universal work agent that connects to enterprise context, handles knowledge and media tasks, writes and runs code, creates sub-agents, and orchestrates across multiple models.

#2 𝕏

Claude Dashboards and Claude Motion are in beta, allowing users to ask Claude to turn data into live dashboards and ideas into animated explainers.

#3 𝕏

Anthropic announced OSS Scanner, a free service using its frontier models to periodically scan opted-in open-source projects for vulnerabilities. Reports will include a proof-of-concept, explanation, and suggested fix.

#4 𝕏

Google Research shared results in The Lancet with BIDMC Medicine from evaluating its research medical system, AMIE, in a real clinic. Across 100 patient interactions, AMIE had zero safety stops and matched doctor diagnoses in 90% of cases.

Also covered by: @Sundar Pichai

#5 𝕏

Anthropic says an unnamed astrophysicist worked with Claude Science to create what it describes as the first complete ultraviolet map of the sky.

#6 📝 Anthropic News

Building on our commitment to American scientific discovery - Anthropic announced on Oct. 8, 2026, that it committed $150 million over three years to the Genesis Mission, a federal initiative, to deepen its support for American scientific research.

Also covered by: @Demis Hassabis

#7 𝕏

NVIDIA shared how H-Company uses NVIDIA Dynamo to optimize VLM serving for computer-use agents.

#8 𝕏

Aravind Srinivas shared that Perplexity Decider ranks first on DecisionBench, with quoted benchmark results reporting 93.9% accuracy and the lowest cost.

Also covered by: @Aravind Srinivas

#9 𝕏

Claude Docs, Slides, and Design are out of beta and available on every plan, including Free. Teams and Claude can edit the same document, deck, or design together.

#10 𝕏

Thariq shared ai-newtab, built on Claude Managed Agents to generate code in a cloud sandbox using Sonnet 5.5. Each run costs approximately $0.75, and the project is available on GitHub.

Also covered by: @Thariq

#11 𝕏

Santiago demonstrated Typesense, an open-source search engine, searching 32 million records in 30 milliseconds and linked to a GitHub project owned by the typesense account.

#12 ▶️

What Changes When AI Goes to Production | Katja Sirazitdinova, NVIDIA

SyntaxGTM

#13 𝕏

DeepLearning.AI shared its free course “Building AI Assistants with On-Device Memory,” developed in partnership with Qdrant and taught by Dylan Couzon. It covers building assistants with on-device memory that requires no network connection or remote server.

#14 𝕏

Santiago shared an unnamed creative-work harness that connects AI models to Blender, DaVinci Resolve, After Effects, Ableton, and Unity to build videos, graphics, games, and other projects, operating similarly to Codex or Claude Code.

#15 𝕏

Rowan Cheung shared four copy-and-paste prompts for GrokBot routines, saying a recent update added native X monitoring. The routines cover AI workflow discovery every weekday at 9:15am in his time zone with at most three results sent to Slack; X brand-mention digests every weekday at 9:00am covering the last 24 hours; category monitoring every four hours while skipping posts older than 24 hours; and follow-up tasks for contacts added in the last 14 days, checked every weekday at 8:30am in his time zone without texting, emailing, or calling anyone.

#16 in

Anu Jagga Narang created a /save skill that asks Claude to store key conversation decisions, but found recall inconsistent and stale. Citing Zhou et al.’s TEPA paper—where append-only memory succeeded 21% of the time after answer reversals versus 31% with no memory—she now asks Claude to validate saved information, sometimes multiple times for important work.

#17 𝕏

LlamaIndex 🦙 shared a breakdown of why Markdown is its default parsing output: it preserves headings, lists, and tables for easier debugging, while HTML is used for tables with merged headers. A parser can capture every word yet still lose which column a number belongs to, forcing the model to guess.

#18 𝕏

Cursor demonstrated that users can ask a follow-up question in the same chat, and /visualize will answer with a new chart.

#19 𝕏

Philipp Schmid says SynthID Detector is now publicly available at synthid.com, supporting Nano Banana 2.1, OpenAI, NVIDIA, and Kakao, with Apple support coming soon. Users can upload or paste image, video, or audio files to scan for watermarks from Google or its partners, with files deleted immediately afterward.

#20 𝕏

Hugging Face encouraged readers to support open science by trying Paper Pages, sharing context about new Paper Reproductions involving an agent-based ML Intern tool.