The Sharpa D01 humanoid grabs soft rods dropped at random, with nobody steering it. In the clip shared by CyberRobooo, rods are dropped at random and the robot grasps them autonomously. CyberRobooo's take: humanoids have already surpassed humans in certain capabilities. Random drops. Autonomous grabs.

A humanoid developers can put to work, from around $18k: Astribot's T1. It made its North America debut at IROS 2026, per TheHumanoidHub. In the clip: T1 at Astribot's booth, a person beside it at a table of objects. - ~1.55m tall - 23 degrees of freedom - 5kg payload per arm - Teleoperable out of the box, for data collection and training - Open SDK Astribot builds its own AI models, the Lumo family, and they run on the robots. $18k, and the AI stack comes with it.

Claude helped a Harvard physicist produce 36 manuscripts in three months. Often, they only became scientifically valuable once human experts stepped in. Matthew Schwartz used BootLoops, an open-source harness, across 18 fields, from particle physics to linguistics, per The Decoder. His advice: "Look at everything yourself."

Claude Opus 5.5 and Sonnet 5.5 are now in Antigravity, for paid users. The best Claude models, per @Claudeupdates11. Pay for Antigravity, and they're yours.

A $25M donation from Greg Brockman fell through, and leaked Slack messages show his own OpenAI staff helped sink it. In June, employees pushed back on his ties to Leading the Future, per Garrison Lovely at Semafor. That pushback contributed to Brockman reneging on the money, Semafor reports. A Slack thread. A $25M U-turn.

This is wild. One prompt, and Claude Fable 5.5 made a 3-minute animated film with an original score. The prompt, per X user @imjustnewatai: a video of all human progress, then 30 years into the future. In the clip: the discovery of fire, antibiotics in 1928, an imagined 2046. No footage. No stock music. One prompt. Three minutes. Its own soundtrack.

A trillion tokens a day. That's where Jev was about a week ago, per TypeSafe CEO Diogo Almeida. About a quarter of Fortune 500 companies use it, he told the Wall Street Journal. He calls Jev the forefront of a "new class of AI." Silicon Valley circles are abuzz, per the Journal. Both numbers are the CEO's own.

Meta's AI agent Muse can now run on hardware you build, like a Raspberry Pi 5. Muse Gadgets is open-source ESP32 firmware plus a Linux SDK, per Engadget. Meta's own ideas, per The Verge: Muse on a color E Ink display showing reminders, or on an HDMI stick for a big screen. Nat Friedman's tip: grab an API token and point your coding agent at the GitHub repo. Meta's own gadget, Muse Home Link, connects Muse over home Wi-Fi to devices you already own. Community-built skills can let it turn on a light, control a TV or send a document to a printer, depending on your setup. Free with an active Muse subscription. US only, one per subscriber, ships in October. Build your own, or claim Meta's.

This is wild. One prompt and a song went into Opus 5.5. 12 hours later, @minchoi woke up to an animated music visualizer. In the clip: a red-and-white figure moves across graphic panels labeled with sections like "pre-chorus" and "a cappella." One prompt. One song. Woke up to this.

One robot brain, MindOn's Mind-1, sorts, packs, folds clothes and tidies a home, across different robots. In the clip, via CyberRobooo: robotic arms sort items on a conveyor and handle clothing, then a humanoid tidies a living room. On a Unitree G1 it has been tested watering plants, moving packages, cleaning mattresses and tidying up. How it performs at actual customer sites remains to be seen, CyberRobooo notes. One brain. Different bodies. Real chores.

Grok Build now puts all your agents on one screen. The new Agent Dashboard opens with one command, per the @grok account, which showed it in a video. Type /dashboard.

Tesla halved the RAM on its AI5 chip to 72GB, betting Optimus needs bandwidth more than memory. AI6 takes a 1/3 cut, to 144GB of LP6, per Elon Musk on X. His reason: it was the only way to get enough volume for Optimus production, and it greatly reduces cost. He expects a negligible effect on performance. Bandwidth was held constant. In the clip, from @Optimus_RH: Optimus's hand near screws. Half the memory. Same bandwidth.

One Opus 5.5 prompt replaced an $850/month motion-design stack, claims X user @0xMovez. No After Effects. No DaVinci. No Premiere Pro. Just one prompt in Claude, per the post. The claimed pipeline: concept → storyboard → transitions → sound → final animation One creator's claim, shared with a guide to doing it. Five stages. One prompt. $850 a month, by their count.

Cloudflare says its new Clef models are smarter and faster than Jev. Clef also costs more, per The Register. They are open-weight and multimodal: they handle images and video. - Clef: built on Qwen3.8-27B - Clef-flash: built on Qwen3.5-9B The weights are on Hugging Face. Bring your own hardware.

A model that fits in a 2.09 GiB file checks formal logic about as well as Qwen3-8B, per webAI's own tests. TwIL-LM3-Pro has 3.66B parameters, less than half as many. It runs on a laptop via llama.cpp. It checks whether a conclusion follows from its premises, turns statements into logic and critiques formal proofs. On BIG-Bench Hard's logic subset: 95.4%. VibeThinker-3B: 61.1%. Built on IBM's Granite 4.2. Non-commercial license. Proof checking, no cloud required.

Other humanoid makers build slim, human-like robot hands. Boston Dynamics just gave Atlas a clunky one. "Many of the current designs are giving up on things like reliability and manufacturability," Alberto Rodriguez of Boston Dynamics told IEEE Spectrum. In the clip, from Boston Dynamics: Atlas handling small objects with the new hands. 13 degrees of freedom, directly actuated, built for sim-to-real RL. The old three-finger hand was research hardware, never meant for mass production. Less human. More manufacturable.

Google DeepMind had a multimillion-dollar budget for Stratego in 2023 and fell short. Researchers from Carnegie Mellon, NYU, Stanford and MIT built Ataraxos for less than $8,000. It decisively beat the best Stratego player of all time, per The Decoder. The hard part: both sides set up their pieces face down. Millions fell short. Under $8,000 didn't.

Grok Bot can now hand a coding task to Cursor and manage your pull requests. The PRs go through GitHub and Origin plugins, per the @bot account. Also new, per @mattyp's walkthrough: - Team bots, shared with your team members - Finances: search and manage them with Plaid - Voice calls: it searches your chats mid-call and handles interruptions And when it builds something, it sends a video demo.

Gemini 4 Argon catches GPT-6 Astra. Not Claude Opus 5.5. That's independent testing, per The Decoder. Argon is Google's first frontier model in over seven months. TestingCatalog has it ahead of Astra, Opus 5.5 and Fable 5.1 across many benchmarks, with a new top score of 77.9% on DeepSWE v1.1. The per-token price is low: $2 in, $10 out during introductory pricing, per @OfficialLoganK. But Argon burns more than twice as many tokens per task as Astra, per The Decoder. First access: a small group of cybersecurity partners. The API and paid tiers come later. Gap closed. No clear lead.

Federal and state agencies can now get Claude for Government. It's generally available, per Anthropic. Claude Code CLI and Claude for Microsoft 365 come with it, in early access. Claude is now open to government.