Claude adds Gmail sending and Drive file management

#1 𝕏

Gmail and Google Drive connectors for Claude let it send emails, draft and send thread replies with user-controlled approval, and manage Drive files. The connectors are available through the connectors menu on all paid plans.

#2 𝕏

The author commented that a 27B dense model delivers “frontier intelligence” on a single 5090, characterizing it as a “new world of local AI.”

#3 𝕏

NVIDIA AI released the open-source TensorRT Model Connect in public preview, enabling a supported Hugging Face model to reach end-to-end TensorRT inference in two commands without an intermediate ONNX export. The resulting bundle can run through native C++ APIs. OpenAI Codex agents were used to build the project—including model implementations, performance tuning, tests, integrations, and documentation—with humans directing and reviewing the work.

#4 𝕏

Philipp Schmid shared that Gemini Managed Agents gives Gemini 3.7 Flash a persistent, isolated Linux sandbox in a single API call, with Python, Node.js, Git, Bash, network access, filesystem I/O, and background processes. Available through the Google AI Studio free tier, it preserves repositories, packages, and generated files across interactions via `environment_id`.

#5 📝 OpenAI News

Pacing model development in an era of cyber-critical capabilities - After the OpenAI–Hugging Face incident and preliminary evidence on August 7 that an upcoming model, Astra, may meet the Preparedness Framework's "Critical cybersecurity capability" threshold, OpenAI paused reinforcement learning training for two weeks, put its largest planned frontier RL run on hold while running smaller-scale evaluations, and paused numerous Astra-related workloads until they meet new security requirements. They implemented stricter safeguards — mandatory sandboxing, network isolation, removal of vulnerable shared services, continuous security testing (including model-driven automated tests), and expanded multistage chain-of-thought monitoring that runs at every sampled token, escalates to automated investigators, aims to issue alerts within 30 minutes and requires pausing activity if teams cannot clear a flag — with monitoring required for all RL training/evaluations involving tools at "Sol" capability or higher and additional monitoring for Astra inference with tools.

Also covered by: @OpenAI, @OpenAI, @Sam Altman

#6 𝕏

Anthropic published a technical report and open-sourced its prompts and data on Hugging Face.

Also covered by: @Anthropic

#7 📝 OpenAI News

ChatGPT Ads expands across Europe - OpenAI is expanding ChatGPT Ads to 31 European countries next week after a six-month U.S. pilot, with initial access through the OpenAI Ads Solutions team, agency and technology partners and self-service Ads Manager coming later this summer. Ads will show only to Free and Go users (Pro and Enterprise remain ad-free), and the platform now supports conversion optimization, geo‑targeting, custom audiences and measurement via the OpenAI Pixel, Conversions API and third‑party integrations, with tens of thousands of marketers already using it.

#8 𝕏

Cursor shared a post tracing 20 years of Git infrastructure and explaining how that history led it to design and operate its Git storage system, Origin, as if it were a database to make Git hosting more reliable, performant, and scalable.

#9 𝕏

Harrison Chase announced the release of LangSmith Tuned Evaluators, starting with Perceived Error, to flag undesirable agent behavior in production traces and attach feedback for improvement. He reports that the tuned model beat frontier models in their benchmark at 82% lower cost.

#10 𝕏

LlamaIndex 🦙 announced that LlamaParse now handles revision tracking, producing clean markdown of a document’s final state and structured data for edits, deletions, and comments—including author, content, and location. It’s intended for redlined documents such as contracts, regulatory submissions, and policy drafts.

#11 𝕏

Claude Cowork is now available on mobile and web for all paid plans.

#12 𝕏

clem 🤗 recapped an ICML reproduction challenge in which, over the past month, 1,221 humans teamed up with coding agents to verify and reproduce 2,226 papers on the Hugging Face Hub. There, 6,816 reproduction logbooks were published openly, 2,962 cloud jobs were launched, and 35,908 claims were judged as agents wrote logbooks, published results, and built on each other’s work.

#13 𝕏

Guillermo Rauch announced a $1 million open effort to verify Vercel Sandbox’s security, inviting participants to test any model for sandbox escapes. If escapes are found, the team plans to patch them, iterate, and share findings to improve transparency around frontier models’ real-world guardrail exploitability.

#14 𝕏

Guillermo Rauch shared that he uses fx.sh as his daily driver, describing the experimental, open-source, model-agnostic tool as 10-20x smaller than major coding CLIs, instantly starting, zsh-like, and embeddable in browsers via WebAssembly.

#15 in

Anu Jagga Narang shared that after nine months building with Claude Code, cleaning Claude’s memory file from 175 lines against a 200-line limit resulted in 181 lines, while a 35KB file guiding writing edits had drifted without being flagged. She rebuilt the setup over the weekend but said she did not yet know whether it worked.

#16 𝕏

The author recapped building a two-sided freelancer marketplace with Bolt.new and using real PageSpeed feedback, including screenshots, to improve its performance. Accessibility reached 100%, SEO rose from 82 to 100, and a shader was added to the hero.

#17 in

Peter Yang recapped Linear’s data report on AI usage, noting that over 40K product teams use Linear and that product AI adoption grew fastest, from 12% to 34%. He also said the share of PMs attaching pull requests rose from 3% to 10% in two years, designers rose from 1% to 8%, and founders were at 23%, second only to engineers. Teams spent more time chatting with AI and delegating to agents without spending less time on existing work.

#18 ▶️

Grok Bot + Grok 4.6 + Cursor Origin - is Claude Code dead?

How I AI Podcast

Also covered by: @claire vo 🖤

#19 𝕏

Mustafa Suleyman recapped MAI-Image-2.6’s 2nd-place ranking on a Text-to-Image benchmark and 3rd-place finish on an image-editing leaderboard, describing it as an all-round high performer for image-generation tasks.

Also covered by: @Mustafa Suleyman

#20 𝕏

Garry Tan commented that GBrain works with AI harnesses including Grok Bot, Claude Code, Codex, Hermes Agent, OpenClaw, and OpenCode, as well as hosted Postgres services using pgvector. He characterized it as enabling ownership of an AI agent’s memory and skills.