AI News: Impressive New Model From Unexpected Company

This video surveys a rapid wave of AI advancements, headlined by Thinking Machines Labs’ new model from Mira Murati’s team, which is framed as a genuine leap beyond typical LLM updates due to real-time interruption-aware translation, posture detection that can proactively prevent risky behavior, elapsed-time tracking, and the ability to run tools like web searches while still speaking. It also spotlights OpenAI’s Codex now working remotely from a phone, plus Claude Code’s cleaner multi-agent interface and Crea 2’s Midjourney-like image controls. The second half focuses on Google’s Android and Chromebook push, where Gemini gains page-aware browsing, one-tap form filling, and voice cleanup, while the new “Googlebook” bakes AI into Chrome OS with a context-aware AI pointer that can execute spoken commands, move objects, and even use eye-tracking. The overarching message is that AI is shifting from isolated chatbots to deeply integrated, proactive agents embedded across devices, browsers, and operating systems.

• ▶ 0:12 The new model from Thinking Machines Labs, founded by former OpenAI CTO Mira Murati, is described as a more meaningful leap forward than typical incremental LLM improvements.
• ▶ 1:19 The demos show real-time translation that speaks over the user, plus pause-aware listening that lets the speaker finish naturally and can count specific words on request.
• ▶ 2:56 The model also detects physical posture, proactively interrupts to prevent risky behavior (like taking 80-year-olds mountain biking), and reframes simultaneous speech in real time.
• ▶ 4:45 The model is uniquely time-aware: it can track real elapsed time during a conversation, such as reminding you when 4.5 minutes are up—something ChatGPT and Claude cannot do.
• ▶ 5:00 It supports simultaneous tool calls, letting it search, browse the web, or generate UI while still speaking and listening, then weave those results back into the conversation.
• ▶ 6:21 Sponsor Crusoe’s Memory Alloy retains and reuses context across requests, keeping inference fast even with long prompts, AI agents, or RAG systems.
• ▶ 7:06 OpenAI's Codex now works from a phone, letting users remotely access and query files on their home computer, which Matt calls a "game changer."
• ▶ 9:02 OpenAI's "Daybreak" takes a different security approach than Anthropic's "Mythos": OpenAI runs the scans for you instead of handing over the capability.
• ▶ 9:54 Claude Code's new Agent View consolidates multiple agents and terminal windows into a single, clearer screen for heavy multi-agent workflows.
• ▶ 10:15 Claude Code now has a cleaner layout, which helps when managing multiple agents at once by making it easier to see what's working and what's done.
• ▶ 10:23 Crea AI introduced a new image model, "Crea 2," which offers functionality and controllability similar to Midjourney.
• ▶ 10:31 Key styling features include image-based style prompting, a stylization slider to control output strength, and independent weighting controls when using multiple reference images.
• ▶ 12:56 The section shifts to Google's announcements from that week's Android event, covering new Android updates and Google Books.
• ▶ 13:09 In a demo, Gemini acts on a photo of an event flyer, automatically preparing a tour in the Expedia app and bringing the user to the final booking step.
• ▶ 13:34 In another demo, Gemini attaches the current web page to a prompt, then navigates to Spot Hero, enters an address, and surfaces parking options, taking the user to the final reservation page.
• ▶ 14:04 A Gemini button is coming to the Chrome browser on Android, mirroring the existing desktop feature, and will activate the assistant directly within the browser.
• ▶ 14:14 The key upgrade is that Gemini will have the context of the web page the user is currently viewing, enabling it to understand page content and take action without switching apps.
• ▶ 14:15 Gemini can fill out forms with a single tap by drawing on stored personal information, such as passport and driver's license details.
• ▶ 14:23 Voice input now cleans up spoken text instead of transcribing it verbatim.
• ▶ 14:26 The system automatically removes filler words like "um" and "uh" for smoother, more polished text.
• ▶ 14:36 It intelligently handles self-corrections and false starts, editing out mid-sentence mistakes (similar to Whisper Flow on desktop).
• ▶ 14:47 The presenter shifts focus "straight into Android phones," transitioning to that topic.
• ▶ 14:49 Google announced "a handful of other features," but these are not detailed in the video.
• ▶ 14:50 Viewers are directed to the news post linked in the description for the full list of additional announcements.
• ▶ 14:57 Google introduced the "Google book," positioning it as the next evolution of the Chromebook.
• ▶ 15:10 The Google book comes with a new operating system designed specifically for AI, though it still runs on Chrome OS.
• ▶ 15:18 The device is essentially a Chromebook with AI features baked directly into the system, similar to AI features on Android phones.
• ▶ 15:26 AI is being "baked into everything" across Android and Google's broader ecosystem.
• ▶ 15:29 Google IO is happening next week, and the narrator will be at the event.
• ▶ 15:32 He expects hands-on time with the announced tools and features, and will report back afterward.
• ▶ 15:32 Google is reimagining the mouse pointer as an AI-driven tool, combining mouse movements, highlighting, and spoken instructions instead of typing commands.
• ▶ 16:25 Users can perform real actions—moving objects, editing text, and reorganizing content—without ever touching the keyboard.
• ▶ 16:35 The system blends mouse movement, highlighting, and voice commands into a single fluid interaction model, turning the pointer into a contextual AI agent.
• [16:43–16:48] The AI pointer can execute natural language commands, like fetching recipe ingredients and adding them to a shopping list.
• [16:49–16:59] The pointer is context-aware: hovering over a note lets the AI interpret ambiguous words like "this" as the actual text note, enabling actions like "Make this orange."
• [17:03–17:23] It supports hands-free head tracking and multimodal generation, e.g., creating an image that combines a menu's content with a bird's artistic style.
• ▶ 17:25 The demo is starting to feel like "the Jarvis we've been promised from Iron Man," though mid-air finger gestures aren't there yet.
• ▶ 17:34 Eye-tracking plus a verbal command ("Move this thing to that over there") lets the AI perform the action without a mouse.
• ▶ 17:43 The host wonders how far we are from dragging fingers around to move objects with commands like "move that to there."
• ▶ 17:47 The system will know exactly what you're pointing at, enabling highly accurate, context-aware pointer tracking.
• ▶ 17:50 This functionality is built into the new Chromebook laptop experience.
• ▶ 17:53 It's likely only a matter of time before the feature expands to a wider range of devices beyond Chromebooks.

• ▶ 17:56 The speaker notes that a previously discussed feature will also be available "in other devices as well," extending the announcement beyond the immediately demonstrated platform.

• ▶ 17:59 He shifts the video's pace, explaining he has "a handful more things" to show, characterizing them as "smaller updates" that he will run through "really quickly in a rapid" — setting up a fast-paced summary of the remaining items.

• ▶ 18:26 Anthropic changes Claude subscriptions for third-party tool use to monthly credits that bill at API rates once exhausted, which users call a "massive nerf" that burns out in hours.
• ▶ 19:48 Anthropic passes OpenAI in business adoption for the first time, rising to 34.4% of businesses while OpenAI fell to 32.3%.
• ▶ 20:12 Anthropic expands into verticals with Claude for legal (MCP connectors/plugins) and Claude for small business, adding pre-built agents across finance, sales, marketing, HR, and more.
• ▶ 21:19 Anthropic’s Claude helped a user recover a 12-year-old Bitcoin wallet by finding an old wallet file on their dumped college computer, which decrypted the mnemonic.
• ▶ 22:02 Meta added incognito chat inside WhatsApp and began rolling out Muse Spark, its most powerful LLM, across Meta AI apps with faster voice and smarter glasses.
• ▶ 23:32 Digg relaunched as an AI-powered news aggregator by Kevin Rose and Alexis Ohanian, surfacing top stories and rising trends from leading AI voices on X.
• ▶ 26:42 Host discovered an AI workflow using Claude Code that can generate a full multimedia scene from a single terminal prompt.
• ▶ 26:42 The automated process creates 3D models, Gaussian splats, and ambient looping sounds from one GitHub repo after cloning and running Claude Code.
• ▶ 27:08 Host calls it “the type of stuff” he loves AI for and plans to test it himself, possibly making a separate video about it.
• ▶ 27:36 Rivian's new AI assistant uses "Unified Intelligence" to embed AI into the vehicle with full awareness of its status and diagnostics.
• ▶ 27:43 The assistant enables natural-language voice controls, like adjusting heated seats, reading/dictating messages hands-free, and acting as an interactive vehicle manual (e.g., step-by-step tire changes).
• ▶ 28:02 Matt sees this as a preview for the entire auto industry, where future vehicles will let drivers describe problems and the AI will use car-specific diagnostics to guide troubleshooting.
• ▶ 28:30 Figure Robotics is running a 24/7 livestream of its robot working continuously; originally planned as an 8-hour event, it was extended indefinitely after hitting the 8-hour mark.
• ▶ 28:47 The stream has been live for over 34 hours, with the robot sorting 43,000 packages during that time.
• ▶ 28:55 The robot’s task is simple—flipping package labels face-down onto a conveyor belt—but the host notes the remarkable part is its endurance: performing the repetitive action almost non-stop for nearly 35 hours.
• ▶ 29:17 Matt plans to attend Google I/O in person for hands-on demos and to meet AI folks, but stresses all announcements are unconfirmed rumors.
• ▶ 29:46 A rumored Gemini 3.2 Flash model could arrive, offering about 92% of GPT-5.5's coding/reasoning ability while being 15–20x cheaper and faster.
• ▶ 30:08 A new "Gemini Spark" agent is rumored—a 24/7 assistant that learns from user behavior and works with connected apps and skills.
• ▶ 30:29 Expect news on Google AR glasses at I/O, but the host has "no clue" if they'll actually appear.
• ▶ 30:49 The glasses were demoed live on stage at last year's I/O, yet still haven't been released to the public.
• ▶ 31:00 Don't expect an official glasses launch this year—likely just a progress update, plus more Android and "Google book" news.
• ▶ 31:16 The week featured a bunch of smaller updates rather than one dominant story.
• ▶ 31:23 The host was really impressed by Thinking Machines’ release and wants to get hands-on with it.
• ▶ 31:27 He is super excited about the ChatGPT app coding feature, which lets him steer coding work from his phone instead of sitting at his computer.
• ▶ 31:14 Bigger coverage is coming next week, pointing to the upcoming Google I/O event.
• ▶ 31:34 The speaker is excited about AI-based coding, noting they are no longer coding in the traditional sense.
• ▶ 31:37 Due to attending Google I/O next week, they plan to publish only one video—the end-of-week news roundup.
• ▶ 31:48 They confirm the end-of-week recap will cover everything learned at Google I/O plus other major AI news from that week.
• ▶ 31:53 The host wraps up by reminding viewers the roundup covers the week's AI news outside of Google announcements, and encourages likes and subscriptions to stay informed.
• ▶ 32:04 The channel's mission is to "drink from the fire hose" of daily AI news and break it down to what most people need to know, filtering out noise, hype, and junk.
• ▶ 32:23 Instead of daily coverage, the channel offers one weekly roundup with everything viewers need to know, so they don't have to track AI news all week long.

• ▶ 0:00 Intro and Live AI Model Demos (0:00 - 4:31) - Intro plus live demos of translation, pause-aware interruption, slouch detection, and language reframing.
• ▶ 4:31 Time-Aware Demo and Sponsor (4:31 - 7:04) - Time-aware model capabilities and tool-use demo, followed by the Crusoe managed inference sponsor.
• ▶ 7:04 OpenAI and Anthropic Tool Updates (7:04 - 10:21) - Codex mobile and remote wiki access, OpenAI Daybreak, and Claude Code Agent View.
• ▶ 10:21 Image Tools and Google Announcements (10:21 - 18:04) - Crea/Krea image tools, Google Android updates, Google Book, and AI-powered Chromebook news.
• ▶ 18:04 Anthropic Plan and Business News (18:04 - 21:16) - Claude Code limit/subscription changes, credit change reactions, business adoption milestone, and legal/small-business launch.
• ▶ 21:14 More AI News and Creative Stories (21:14 - 26:46) - Claude Bitcoin wallet recovery, Meta/Notion/Dig updates, Made with AI prank, and World Labs 3D tool.
• ▶ 26:46 Creative AI, Robotics, and Google I/O Preview (26:46 - 32:49) - Claude Code 3D/sound generation, Rivian AI assistant, Figure robot livestream, and Google I/O expectations.