- The Rundown AI
- Posts
- Nvidia's reveals an open AI model
Nvidia's reveals an open AI model
PLUS: Ultra-realistic humanoid robots are coming
Rowan Cheung
June 17, 2024
Sign Up | Advertise | Tools | AI University
Welcome, AI enthusiasts.
Nvidia already has the AI chip industry in the palm of its hand — and now it’s pushing the open-source LLM sector forward as well.
The chipmaker’s latest ‘Nemotron’ model family unlocks a new world of training data abilities, paving the way for even more powerful AI systems. Let’s dive deeper…
In today’s AI rundown:
Nvidia’s new open-source AI models
Chinese startup creates realistic robots
Transform presentations into blog posts
Tiny LLM matches GPT-4 on math
6 new AI tools & 4 new AI jobs
More AI & tech news
Read time: 4 minutes
LATEST DEVELOPMENTS
NVIDIA
Image source: Nvidia
The Rundown: Nvidia just introduced Nemotron-4 340B, a family of open-source language models designed to generate high-quality synthetic training data and build powerful AI applications across industries.
The details:
The three models (Base, Instruct, Reward) form a ‘pipeline’ for creating synthetic data to train new, powerful LLMs.
Instruct creates high-quality synthetic training data (and was trained on 98% synthetic data), while Reward filters the data for the highest-quality examples.
The Nemotron-4 models match or exceed open-source competitors like Llama-3, Mixtral, and Qwen-2 across a variety of benchmarks.
NVIDIA also released Mamba-2 Hybrid, a selective state-space model (SSM) that surpassed similar transformer-based LLMs in accuracy.
Why it matters: Nvidia just provided a free, open-source model family that not only matches the capabilities of some of the top rivals in the space, but also excels at crafting synthetic data needed to continue leveling up new LLMs. The chipmaking giant is an AI powerhouse of many talents.
TOGETHER WITH HUBSPOT
The Rundown: Unlock the future of marketing with HubSpot’s AI Prompt Library — a free ebook filled with powerful curated prompts to elevate your content and drive higher conversions.
With HubSpot’s AI Prompt Library, you can:
Access prompts crafted by HubSpot's top marketing minds
Create engaging content that resonates with your target audience
Unlock creative and unconventional marketing strategies to set you apart
Download the free ebook now and start transforming your marketing efforts today.
EX-ROBOTS
Image source: Yahoo
The Rundown: A Chinese startup called Ex-Robots is developing ultra-realistic humanoid robots that can closely mimic human facial expressions and emotions.
The details:
Ex-Robots’ lifelike humanoid robots can smile, frown, and replicate subtle facial movements using small motors embedded in the silicone faces.
The startup uses multimodal AI and algorithms to enable the robots to recognize and express emotions based on perceiving their environment.
Creating one humanoid robot currently takes 2-4 weeks, with costs as high as $275k per unit.
The company said near-term use cases include museum exhibits, with eventual roles in healthcare, therapy, and child services.
Why it matters: The advancement of ultra-realistic robots combined with the power of AI is about to create some bizarre new frontiers. Will humans be creeped out by an actual human-looking humanoid over mechanical options like Tesla’s Optimus? It’s unclear if the world is ready… but we’ll find out soon.
AI TRAINING
The Rundown: Canva recently introduced an AI feature called ‘Magic Switch’ that automatically converts your presentations into compelling blog posts.
Step-by-step:
Sign up for Canva Pro and access the Magic Studio here (free trial available for one month).
Open your presentation in Canva.
Click on “Resize & Magic Switch” in the top bar.
Select “Transform into Doc” and select “Blog Post”.
In a few seconds, it’ll generate an entire blog post draft based on the presentation 🎉
PRESENTED BY WEBAI
The Rundown: webAI’s platform, Navigator, empowers users to build, develop, and deploy AI-powered solutions quickly and efficiently — regardless of ML resources or expertise level.
With Navigator, you can:
Train, build, and deploy AI solutions in one environment
Easily connect to data sources and deploy to any device on your network
Maintain privacy using models trained on your own machine
Build using the tools right for you with our no-code and full-code functionality
Get early access to Navigator and start building and harnessing the power of AI today.
AI RESEARCH
Image source: Midjourney
The Rundown: Shanghai AI Laboratory researchers just developed a new algorithm called MCT Self-Refine, which allowed a relatively small 8B-parameter model to achieve GPT-4 level performance on complex math.
The details:
MCT Self-Refine combines an 8B LLaMa model with Monte Carlo Tree Search, an AI technique famously used by Google DeepMind to master the game of Go.
The algorithm constructs a search tree, refining answers through a series of processes and self-evaluation.
On the GSM-Hard math word problem benchmark, infusing the Monte Carlo Tree Search took the 8B model from 25.5% accuracy to 45.5%.
The model also achieved 96.7% on the GSM8K benchmark, outperforming significantly larger models like GPT-4, Claude, and Gemini.
Why it matters: The research shows it’s possible to boost math capabilities without massive scale — and GPT-4 level performance with a model trained on 200x less parameters is an impressive feat. If the approach proves to be a more efficient path to advanced reasoning, we could be on the cusp of a new wave of model acceleration.
NEW TOOLS & JOBS
🤖 Brave Leo AI - An AI assistant in Brave browser to stay up-to-date with search results
💼 Linkedin AI Job Coach - Find opportunities and generate applications
🔎 Leap AI Lead Research - Enter an email to find out everything about that person
🚀 StoryBlock - A CMS to get value from content faster with AI
🧠 Huly - Open-source replacement for Linear/Jira, Slack, Notion, and more
🎨 Abridge - Senior Product Designer
💻 Palantir Technologies - Software Engineer, New Grad
🧪 Fiddler AI - Staff AI Scientist
🤖 Meta - Software Engineer - ML Systems
QUICK HITS
Microsoft is officially delaying the launch of its new Copilot+ Recall feature, now set to come as a preview feature available initially to testers in the Windows Insider Program.
Sam Altman has privately revealed that OpenAI is considering transitioning its nonprofit governance structure to a for-profit benefit corporation.
McDonald’s is removing its AI-powered drive-thru ordering system tested in over 100 restaurants following viral videos showing customer frustrations and technical issues.
Edward Snowden responded to OpenAI’s addition of former NSA director Paul Nakasone to its board, saying the move is a 'willful, calculated betrayal of the rights of every person on earth.’
MIT unveiled a new study showcasing a 3D technology platform capable of mapping the human brain.
Tesla shareholders are suing Elon Musk for diverting talent and resources to xAI, citing a breach of fiduciary duties and unjust enrichment.
THAT’S A WRAP
SPONSOR US
Get your product in front of over 600k+ AI enthusiasts
Our newsletter is read by thousands of tech professionals, investors, engineers, managers, and business owners around the world. Get in touch today.
FEEDBACK
How would you rate today's newsletter?Vote below to help us improve the newsletter for you. |
If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.
Reply