NVIDIA demos Cosmos 3 camera-controlled views from one image

#1 𝕏

NVIDIA AI demonstrated NVIDIA’s Cosmos 3 generating views from a single image as the camera follows a dog, rises, or moves sideways. A linked technical report details the post-training recipe.

#2 𝕏

Mistral AI retweeted @askalphaxiv’s post about “Mistral Large 4,” described as a technical report published exclusively on alphaXiv.

Also covered by: @Mistral AI

#3 📝 OpenAI News

Scaling Git at OpenAI - OpenAI published “Scaling Git at OpenAI” on November 2, 2026, credited to Aaron Friel, Taylor Blau, and Ted Nyman; its listen-to-article audio runs 28:26. The article describes efforts to revamp source control for agentic-scale demands after OpenAI’s monorepo grew rapidly over the prior year, while investing in the Git ecosystem.

#4 𝕏

NVIDIA AI shared a walkthrough for configuring, deploying, and validating NVIDIA AI Cluster Runtime, which combines NVIDIA Dynamo with supporting infrastructure, tested component versions, and checks for Kubernetes GPU setups. NVIDIA AI also said it would be at KubeCon North America the following week.

#5 📝 OpenAI News

LifeVal - **OpenAI News** features a webpage titled “LifeVal” on OpenAI’s website; the available source provides no details about its capabilities, timing, or release status.

#6 𝕏

Aravind Srinivas shared an @perplexitydevs post introducing pplx-embed-v2-context, open-weight vision-language embedding models offered in 9B and 0.6B sizes. The truncated source begins to state that both return one int8 output but does not specify the complete claim.

#7 𝕏

claire vo 🖤 shared a Lenny’s Newsletter blog breakdown covering Ezra’s PM OS, AutoXP, and bug-bash workflow; the linked article’s URL references Figma’s PMs.

Also covered by: @claire vo 🖤

#8 𝕏

Andrej Karpathy said Claude Code can be prompted to create and deploy a browser extension that displays YouTube thumbnails from halfway through each video, with a toggle for the user-uploaded versions. He said the approach “works great.”

#9 𝕏

Hugging Face retweeted @vanstriendaniel’s report that they re-OCR’d 14,483 pages of fossil fuel industry documents using Saturate on one GPU for $1.47, processing 8.9 pages/s once running.

#10 🐙 D4Vinci/Scrapling

D4Vinci/Scrapling - D4Vinci/Scrapling is a Python web-scraping framework combining HTTP requests, stealth browsers, parsing, and concurrent crawling, with source code, documentation, tests, Docker support, an MCP server, and an Agent Skill. It has 93,783 GitHub stars and a 1,658 daily trending gain. Worth trying if you need reusable scraping and crawling tools for an AI product.

#11 🐙 max-sixty/worktrunk

max-sixty/worktrunk - **max-sixty/worktrunk** is a CLI for managing Git worktrees across parallel AI-agent workflows; the repository includes source code, documentation, hooks, plugins, and workflow automation. It has 10,274 GitHub stars, with gains of 437 daily and 406 weekly. Worth trying if you need isolated workspaces for multiple coding agents.

#12 🐙 anomalyco/hex

anomalyco/hex - anomalyco/hex is an MIT-licensed voice-dictation app for Mac, Windows, and Linux that downloads a speech model, transcribes on-device, and pastes into the active app. The repository includes app and native source, platform-specific code, documentation, and packaging, and has 4,248 GitHub stars with 1,134 daily trending gains. Worth trying if you need private, cross-platform voice input.

#13 𝕏

Udi Menkes 🚢 commented that, per Ramp data, the top 1% of AI-paying firms on Ramp spend $9,568 per employee a month on AI versus $15.70 for the median firm—about 609x more. He recommends designing pricing, limits, and support for these high-spending accounts rather than relying only on per-seat plans.

#14 𝕏

Philipp Schmid created a Code Mode explainer on philschmid.de covering how it works, why typed JavaScript or Python sandboxes can beat raw bash, and a minimal example for Gemini 4 Argon. He recommends testing Code Mode with pi, describing it as easily hackable with a strong foundation.

#15 𝕏

clem 🤗 reposted guilleflorvs’s claim that Revolut moved approximately 2,000 engineers onto self-hosted GLM for coding.

#16 𝕏

Guillermo Rauch commented that stateless sandboxes are the wrong benchmark because useful agents retain workspaces with identity, skills, repositories, and memories. He said the Vercel Sandbox team is meeting latency and throughput SLOs while prioritizing reliability and security, shipping clusters across 19 global regions and new features every week amid customer demand growing by a double-digit percentage per week.

#17 𝕏

Guillermo Rauch commented that securing the supply chain is a top priority and that an unnamed system blocks deployment of hijacked npm packages to protect code and secrets from exfiltration.

#18 𝕏

Garry Tan said Capy Captains increased his issue-fixing throughput by about 5X overnight, with coordination across captains and subthreads providing a significant benefit.

#19 𝕏

OpenAI announced plans for its next internal security sprint, partnering with Cathedral Systems researchers to use OpenAI’s models alongside its teams to challenge its defenses. OpenAI is also raising bug bounty rewards, more than quadrupling rewards for critical findings and offering up to $172,000, including bonuses, for exceptional findings.

Also covered by: @OpenAI

#20 𝕏

Hugging Face shared NielsRogge’s post saying OpenBMB released MiniCPM-OCR, a 1B-parameter OCR model for document parsing.