Howdy! It’s Barsee again.

Happy Thursday, AI family, and welcome back to AI Valley.

In today’s edition:

• 🤖 Boston Dynamics and Toyota’s Humanoid Robots

• 🧠 Nvidia casually dropped an Open-Source 70B model

• 📱 Real-time Motion Capture with a smartphone

• 🤖 Plus trending AI tools, posts, and resources.

Ready, set, go…

TOGETHER WITH LAMBDA

🦾 New NVIDIA H100 SXM instances deliver up to 51% higher performance compared to PCIe versions

Lambda deployed 1x, 2x, and 4x NVIDIA H100 SXM GPU instances in its public cloud, offering new options for AI Developers looking to access high-end GPUs without committing to full-node 8x configurations.

Benefit from:

• Up to 51% higher performance compared to PCIe versions

• More instance types to right-size your compute capacity

• Affordable, on-demand access to powerful NVIDIA H100 SXM GPUs

ROBOTICS

🤖 Boston Dynamics and Toyota’s Humanoid Robots

On Wednesday, Boston Dynamics and Toyota announced a partnership to accelerate the development of general-purpose humanoid robots that can more efficiently tackle real-world challenges.

New advancements in robot learning:

• Toyota researchers have created "Large Behavior Models" (LBMs) that operate similarly to AI models like ChatGPT, enabling robots to learn physical tasks by simply observing human behavior.

• Using this method, Toyota engineers have successfully taught robots tasks like pouring liquids and using kitchen tools.

Why does this partnership matter?

• Toyota will test its advanced AI models on Boston Dynamics’ Atlas robot, leveraging the robot's impressive physical capabilities to gather critical data that will help refine the AI models even further.

• In exchange, Boston Dynamics will gain from Toyota's AI expertise, incorporating Large Behavioral Models (LBMs) to improve Atlas's ability to manage complex tasks and make smarter decisions.

NVIDIA

🧠 Nvidia casually dropped an Open-Source 70B model

Nvidia has unveiled a new fine-tuned, open-source large language model (LLM) called Llama-3.1-Nemotron-70B-Instruct. This model is now outperforming industry giants like GPT-4o and Claude 3.5 Sonnet on key benchmarks.

Here’s a breakdown:

• Built on Meta’s Llama 3.1 70B model, Nemotron has been enhanced by Nvidia using advanced techniques such as Reinforcement Learning from Human Feedback (RLHF).

• It scores highly on alignment benchmarks, achieving 85.0 on Arena Hard, 57.6 on AlpacaEval 2 LC, and 8.98 on GPT-4-Turbo MT-Bench.

• Despite its smaller size of 70 billion parameters, Nemotron is surpassing competitors like GPT-4o and Claude 3.5 Sonnet across various metrics.

How to get it?

Nvidia has open-sourced the model along with the reward model and training dataset on Hugging Face, and a preview is available on their website.

MOTION CAPTURE

📱 Real-time Motion Capture with a smartphone

Engineers at Northwestern University have recently introduced MobilePoser, a system that enables full-body motion capture using just a mobile device, eliminating the need for specialized rooms or expensive equipment.

What makes it unique?

• MobilePoser utilizes three sensors—accelerometers, gyroscopes, and magnetometers—that are already embedded in portable devices like smartphones, smartwatches, and wireless earbuds.

• To improve the accuracy of these sensors, the team developed a custom AI algorithm, trained on a large dataset of simulated sensor readings based on high-quality motion capture data.

• By combining sensor data, machine learning, and physics, MobilePoser accurately tracks a person's full-body pose and translates that into their 3D movements in real-time.

Why does it matter?

• MobilePoser is significant because it paves the way for exciting immersive gaming experiences and health applications.

• Plus, this technology could improve indoor navigation, addressing a major limitation of GPS, which only works outdoors.

QUICK HITS

• Mistral releases Ministral Edge LLMs optimized for laptops and phones. It generally outperforms Google’s Gemma 2 2B and 9B and Llama 3.2 3B and 3.1 8B on benchmarks. (link)

• AMD has released a new AI chip, MI325X, which will ship in Q4. It is rumored to beat NVIDIA’s H200. (link)

• OpenAI released a study to evaluate fairness in ChatGPT—it analyzed how GPT responds to users based on their names and found there were occasional differences in responses. (link)

• AI-powered social media app promises to ‘shape reality’. (link)

• Reddit has launched AI keyword-targeting with dynamic audience expansion, multi-placement optimization, AI keyword suggestions, and unified targeting flow. (link)

• X (formerly Twitter)’s ToS Update Allows AI Training on Your Data. A royalty-free permission to use, reproduce, modify, and the right to use your content for AI training for Grok. (link)

• Meta introduces TPO, a new method that helps AI models "think" before responding, similar to OpenAI's o1 approach. (link)

TRENDING TOOLS

• Glato AI > Create video ads with expressive AI creators from a product URL. (link)

• Napkin > It turns your text into visuals so sharing your ideas is quick and effective. (link)

• AI Haggler > Web app that checks hotel availability and negotiates the best prices for you. (link)

• SheetMagic > AI + web scraping in Google Sheets. (link)

• Strella > It conducts customer interviews so you can gather their insights. (link)

• Reworkd > AI agents will understand web pages and automatically generate code to extract the exact data. (link)

COOL FINDINGS / RESOURCES

• 100+ incredible ChatGPT statistics & facts in 2024. (link)

• 7 useful tips on building with Cursor. (link)

• A series of video tutorials by Meta AI on how to build with Llama models across Linux, Windows, Mac, and more. (link)

• Apple study exposes deep cracks in LLMs’ “reasoning” capabilities. (link)

• How AI is removing legal obstacles that slow down business. (link)

DAILY DOSE OF CONTENTS

1/ How lightspeed invests in AI and beyond.

2/ Cops are using high-tech grapplers to stop high-speed chase.

THAT’S ALL FOR TODAY

That’s all for today’s issue, folks.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ Like what you see? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.