Google AI announces Gemini 4 Argon's 1M-token output limit

#1 𝕏

Google AI announced Gemini 4 Argon, a frontier model built for deep reasoning across complex, long-horizon workflows in software engineering, legal and finance work, and cybersecurity defense, with an expanded 1M-token output limit. Argon is rolling out to trusted cyber defenders in the Fairwind Program, with broader availability planned.

Also covered by: @Philipp Schmid, @Google DeepMind, @Google DeepMind, @Sundar Pichai, @Logan Kilpatrick, @Sundar Pichai, @Demis Hassabis, @Logan Kilpatrick

#2 𝕏

Cognition announced that it is the first customer running NVIDIA Vera Rubin, powered by CoreWeave. On SWE-2 inference, Cognition says the new chips deliver about 4.8x more token throughput than GB200 at the same decode speed.

#3 📝 OpenAI News

Disrupting a coordinated model-distillation campaign - OpenAI News recapped how OpenAI identified and disrupted a coordinated campaign designed to extract protected reasoning from its models, characterizing the activity as consistent with adversarial distillation. The earliest observed activity occurred in the first week of July.

#4 𝕏

Aravind Srinivas says contextual embedding models are being open sourced, describing them as state-of-the-art and best-performing in turbopuffer’s context-bench.

#5 𝕏

Santiago recapped using Dots to check in family flights, transfer his available PayPal balance, automate analytics screenshot email replies, sort a spreadsheet, and summarize task-related emails—all within a few hours. He added that he wants Dots to work on his Linux laptop.

Also covered by: @Peter Yang, @How I AI Podcast, @Peter Yang, @Greg Isenberg

#6 𝕏

Aravind Srinivas announced open access to an email-based task delegation agent, with no Perplexity account required. For a limited time, users can forward or CC computer@perplexity.com to delegate tasks for free; the agent works in the background while retaining the email context.

#7 𝕏

NVIDIA AI shared a tutorial showing how one prompt built and deployed a manufacturing-line visual AI agent with alerts, video search, and incident reports in under 30 minutes using the new Build Vision AI skill in NVIDIA VSS Blueprint 3.3.

#8 𝕏

Cognition shared that @crosbylegal, with 12 engineers supporting 50+ attorneys, put Devin on first response to investigate 20–40 bugs and alerts a day. Devin cut Sentry noise by 50% and kept on-call to one engineer as the legal team grew 5x.

#9 𝕏

NVIDIA AI shared a hands-on walkthrough of NVIDIA NeMo Relay, developed with Nous Research and @Teknium, that collects Hermes Agent traces across two example scenarios and displays calls, retries, and full traces in Arize Phoenix. The accompanying blog examines how Nous used traces and task results to evaluate fixes across repeated runs.

#10 𝕏

Demis Hassabis announced the release of open-source SynthID Bio tools, which enable watermarking of AI-generated proteins so researchers can build on the work. The research was published in Nature.

#11 in

Claire Vo recaps a Jev episode in which John Lindquist demonstrated about a dozen things builders can make with the TypeSafe AI decision model, including multi-step routing, messy-data deduplication, model-versus-model blitz chess, and a real-time presentation coach. Vanta sponsored the episode, which was available live on YouTube.

Also covered by: @Dharmesh Shah, @How I AI Podcast, @Sebastian Raschka, @Sebastian Raschka

#12 ▶️

Did a 50 year old military secret just solve agent prompt injection?

Fireship

#13 𝕏

Santiago shared TwIL-LM3-Pro, an open-source 3.66B-parameter model that scores 95.4% on BIG-Bench Hard’s logic subset and is small enough to run locally on a laptop. He highlighted the potential of specialized post-training for task-specific small models.

#14 𝕏

Qwen3.8-27B is now accessible via @nebiustf. The 27B dense model is described as ready for multi-step workflows, including building agents and deep research.

#15 𝕏

Guillermo Rauch commented that service operators can add their services to Connect to reach 20M+ developers and the billions of agents they’ll ship. He said Connect helps agents and apps avoid static credentials, making integrations easier and more secure.

#16 in

Greg Isenberg shared his view that building a ChatGPT plugin is a strong risk/reward bet, claiming 1.2 billion people use ChatGPT weekly and that, as of the day before his post, it recommended plugins during conversations. He advised turning recurring workflows into plugins, matching descriptions to user phrasing, shipping five, and doubling down on the one gaining the most traction. He predicted that recommendation slots would become crowded within 12 months.

#17 ▶️

Can AI Predict The Future? - Sikt Intelligence (my new startup)

All About AI

#18 in

Guillermo Rauch announced that Vercel is joining the SAP Commerce Cloud ecosystem and working with SAP to improve how coding agents and developers interact with SAP APIs, MCPs, and CLIs to support commerce experiences worldwide.

#19 𝕏

Google Research recapped the CDC’s announcement that Google’s science AI model ranked highest among 39 eligible models for forecasting flu-related hospital admissions during the 2025-26 flu season. Google developed the forecasts using Empirical Research Assistance, an AI tool that generates computational solutions across scientific fields.

#20 𝕏

A post shared a prompt for building a checkout page in bolt.new where clicking “Complete order” fires a confetti burst from the button.