AIR#114 - June 26, 2024
Good morning, AI aficionados! Grab your coffee and get ready to dive into today's edition of AIR: The AI Recon. Leading the headlines is an exciting breakthrough from researchers who have developed AI language models that operate without matrix multiplication. This innovation not only boosts efficiency but also drastically reduces power consumption, with one model running on just 13 watts—the energy of a single lightbulb! It's a must-read for anyone fascinated by the relentless push for more sustainable and efficient AI solutions.
But that's not all! Mozilla is making waves with its latest AI features in Firefox Nightly, offering a privacy-focused, customizable browsing experience. With tools like ChatGPT and Google Gemini now integrated, Firefox users can enjoy enhanced functionality without compromising their privacy. This development is sure to intrigue those who prioritize both cutting-edge technology and data security. Meanwhile, the AI community is abuzz with a new ranking that reveals many so-called 'open source' AI models aren't as open as they claim. This transparency check is sparking important conversations about the true openness and accessibility of AI technologies.
And for those who love a good business update, Waymo has officially opened its autonomous ride-hailing service to all in San Francisco, marking a significant milestone in the deployment of driverless taxis. This expansion is a testament to the rapid advancements in autonomous vehicle technology and its growing acceptance in urban areas. Whether you're here for groundbreaking tech updates, ethical debates, or the latest industry buzz, today's edition is packed with stories that will both intrigue and challenge you. So, sit back, sip your coffee, and let's delve into the dynamic world of artificial intelligence together!
Business
Waymo's Autonomous Ride-Hailing Service Now Open to All in San Francisco
Waymo's autonomous ride-hailing service is now open to all in San Francisco, expanding its driverless taxi operations.
Anthropic: Collaborate with Claude on Projects
Claude.ai now lets Pro and Team users organize chats into Projects, enhancing workflows with curated knowledge and shared team insights.
AI Discovers Rare-Earth-Free Magnet 200x Faster Than Humans
AI discovers rare-earth-free magnet 200x faster, reducing costs and emissions. Revolutionizing electric mobility and more.
Elon Musk Diverts Thousands of Nvidia AI Chips from Tesla to X and xAI
Elon Musk diverts 12k Nvidia AI chips from Tesla to X and xAI, delaying Tesla's supercomputer setup. Tesla shares dip, Nvidia rises.
Why U.S. Tech Must Build AI Weaponry
U.S. tech must lead in AI weaponry to control its purpose and impact, argue Palantir's Karp and Zamiska in The Washington Post.
AI Is Wreaking Havoc on Global Power Grids
AI's massive energy demands are straining global power grids, causing delays, outages, and price hikes. The future of AI may hinge on energy solutions.
Engineering
🔥 Anthropic's New Claude Projects: How to Create and Manage Them
Create and manage Claude projects easily: Set up, add content, and customize instructions. Star projects for quick access.
Researchers Upend AI with MatMul-Free LLMs
Researchers innovate by creating AI language models without matrix multiplication, boosting efficiency and reducing power consumption.
🔥 Researchers Run High-Performing LLM on Lightbulb Energy
Researchers run a high-performing LLM on just 13 watts, the energy of a lightbulb, by eliminating matrix multiplication and using custom hardware.
Mozilla Launches First AI Features in Firefox Nightly
Mozilla introduces AI features in Firefox Nightly, offering privacy-focused, customizable browsing with tools like ChatGPT and Google Gemini.
Sohu – Fastest AI Chip for Transformer Models
Sohu, the fastest AI chip, processes 500k tokens/sec, replaces 160 H100s, and is optimized for transformer models. Transforming AI efficiency!
🔥 Language Models on the Command Line: Simon Willison's Annotated Talk
Simon Willison's talk on using Large Language Models via command-line tools is now annotated and available on YouTube. Check it out!
Linear Algebra Basics for AI/ML – Part 1
Learn Linear Algebra basics for AI/ML with vectors, matrices, and PyTorch operations in part 1 of this series. Dive in and start coding!
Rabbit R1s Break as Elevenlabs API Key Revoked
Rabbit R1s are down as Elevenlabs API key was revoked due to an update oversight.
Sohu: The World's First Transformer ASIC by Etched
Etched raises $120M for Sohu, the first transformer ASIC, promising AI models 10x faster and cheaper than GPUs, revolutionizing real-time AI.
[GitHub] FiddleCube: Generate Ideal Q&A for Testing LLMs
FiddleCube on GitHub generates ideal Q&A datasets for testing, evaluating, and training LLMs, saving manual effort and ensuring accuracy.
Etched's Bold Gamble: The World's Fastest Transformer Chip
Etched unveils Sohu, the fastest transformer ASIC chip, revolutionizing AI with unprecedented speed and efficiency. 🚀
Hackers 'Jailbreak' AI Models to Expose Flaws Globally
Hackers globally 'jailbreak' AI models to expose vulnerabilities and highlight flaws in powerful AI systems.
OpenAI Launches ChatGPT Desktop App for macOS
ChatGPT now on macOS! Seamlessly chat, write, and create directly from your desktop. Download for M1+ Macs. Windows coming soon!
Training a 70B Model from Scratch: Imbue's Open-Source Tools & Learnings
Imbue's 70B model outperforms GPT-4o using open-source tools, clean datasets, and CARBS optimizer. Tools and learnings now available!
Early LLM Performance with AMD Instinct MI300X on Oracle Cloud
Oracle's OCI tests show AMD Instinct MI300X GPUs excel in LLM tasks, promising high performance for large-scale AI workloads.
Claude 3.5 Sonnet: Creating Games Stuck in the Past?
Claude 3.5 Sonnet shows minimal progress in game development over 6 years, as per Vlad Davidoff's comparison on X.
OpenAI Delays Advanced Voice Mode Release
OpenAI delays Voice Mode release to July for improvements, aiming for a fall rollout to all ChatGPT Plus users.
Academic
🔥 Not All 'Open Source' AI Models Are Truly Open: A Ranking
Many 'open source' AI models aren't truly open. Researchers rank models on transparency, revealing smaller players are more genuinely open.
Simulating 500 Million Years of Evolution with ESM3: A Frontier Language Model for Biology
ESM3 simulates 500M years of evolution, generating new proteins and advancing biology with AI, enabling unprecedented protein design.