🤖 Clone Robotics Built a Muscle-Flexing Android (Literally)

/

Clone Robotics introduced Protoclone, a full-body humanoid built with synthetic muscles and over 500 sensors. It’s got a skeleton that mimics all 206 human bones (more or less), runs on pneumatic muscles, and even sweats with a water-cooling system. Yeah, sweating robots are a thing now.

It’s not walking yet—it still needs support—but it's a glimpse into a future where robots might cook your dinner or vacuum your floor. People online are torn: creepy? Impressive? Both?

And if that wasn’t enough…

🥊 Robot Boxing Is Happening

Unitree Robotics is prepping for a full-on robot boxing match called “Iron Fist King: Awakening.” Their G1 bots will throw (slow) punches at each other in a livestreamed event. Is it goofy? Yep. Will it get views? Also yep.

🏃‍♂️ Robots Running Marathons (For Real)

Beijing is hosting a half-marathon where robots run with humans. That’s over 13.1 miles (or 21.1 kilometers) of bipedal robot madness. It's not about speed yet, but endurance. Companies are using this as a stress test for joints, motors, and real-world AI control. Some bots can hit speeds up to 7.5 mph (or 12 km/h)—wild.

🧠 DeepSeek Just Leveled Up AI Reasoning

A new model from DeepSeek, GRM (Generative Reward Model), now lets AI judge and critique itself before answering. Using something called SPCT (Self-Principled Critique Tuning), it basically learns its own values and uses them to score its own responses.

With repeated sampling and a “meta reward filter,” it outperforms GPT-4o on certain tasks—all with a smaller 27B model. Open-source whispers are swirling, and DeepSeek might be prepping to release its chatbot “R2” built on this tech.

🧠 NVIDIA Drops the Efficient Powerhouse: Nemotron Ultra

NVIDIA’s new Nemotron Ultra is a beast—253B parameters—but optimized like crazy. It runs on just one node with 8x H100s, thanks to architecture tricks like skipping attention and compressing layers.

Coolest part? It has a “reasoning on/off” switch. Flip it for deep logic, math, and code. Turn it off for quick replies. It outperforms bigger models like DeepSeek R1 while being open source and cheaper to run. Developers, this one’s for you.

🦙 Meta Quietly Unleashed Llama 4—and It's a MONSTER

Meta released Llama 4 Scout and Maverick—open-weight models with massive token windows (up to 10 million!), native multimodal support, and performance rivaling GPT-4... using half the compute.

Scout runs on a single H100. Maverick uses 128 experts in a Mixture-of-Experts setup. And the real game-changer, Llama 4 Behemoth (2 trillion parameters), is still training. Rumors say it could crush GPT-4.5 and Claude 3.7 when it lands.

📸 Google’s Gemini Live Gets Real-Time Vision with Project Astra

Google and Samsung teamed up to roll out Gemini Live’s real-time visual AI. Point your phone’s camera at anything, and the AI tells you what’s going on—analyzing fingers, landmarks, or entire websites via screen share.

It’s part of Project Astra, and it's going live now on Pixel 9 and Galaxy S25 devices. Gemini 2.5 Pro is also rolling out with more API access and cheaper token pricing. Google’s clearly going for that AI-assistant throne.

🧠 ChatGPT’s Memory Just Got an Upgrade

OpenAI made a big move too—ChatGPT now remembers everything you’ve ever told it. That means your conversations can feel more personal and connected over time. It’s rolling out to Pro and Plus users (unless you're in the EEA/UK), and you can opt out if that freaks you out.

🧵 TL;DR – This Week's Biggest Moves

• Clone Protoclone – Human-like android with synthetic muscles and 500+ sensors. It sweats, flexes, and aims to do your chores one day.

• Unitree Boxing Match – Humanoid robots are set to fight in a livestreamed event. Slow punches, big entertainment.

• Robot Half-Marathon – Beijing will host robots running 21km alongside humans. Endurance test for real-world navigation.

• DeepSeek GRM – New AI model critiques and scores its own answers. Outperforms GPT-4o in some benchmarks using less compute.

• Nemotron Ultra – Nvidia’s efficient 253B model with a “reasoning on/off” switch. Beats bigger models on a single 8xH100 setup.

• Llama 4 – Meta’s Scout and Maverick offer massive 10M-token windows and open weights. Behemoth (2T params) is still training.

• Google Project Astra – Real-time visual AI via Gemini Live on Pixel 9 and Galaxy S25. Point your camera, get instant feedback.

• ChatGPT Memory – ChatGPT now remembers past conversations for smarter, more personal interactions.

That’s a wrap for this week! If you’re loving this format, reply with “🔥” or forward it to a friend who wants to stay ahead of the AI curve without watching full videos.

Catch you next week,
– Azem @ AI Revolution

P.S. Our free AI Avatar course is still live — https://www.skool.com/ai-content-accelerator
Join the AI Content Accelerator and start building with these tools today.



HOW DID YOU SEE TODAY NEWSLETTER

🌟🌟🌟🌟🌟 Excellent

🌟🌟🌟 Average

🌟 Good