AIR#98 - June 10, 2024
Good morning, AI aficionados! As you sip your morning coffee, prepare to dive into today's edition of AIR: The AI Recon, where the world of artificial intelligence is buzzing with stories that are both groundbreaking and thought-provoking. Leading the headlines is a fascinating new research paper that eliminates the need for matrix multiplication in large language models, reducing memory usage by up to 61% while maintaining top performance. This innovation could revolutionize the efficiency of AI models, making it a must-read for anyone excited about cutting-edge AI advancements.
In other thrilling news, teams of LLM agents have now been shown to exploit zero-day vulnerabilities, significantly improving detection and planning capabilities. This development underscores the dual-edged nature of AI, offering both incredible potential and new challenges in cybersecurity. Meanwhile, the open-sourcing of Labmlai's "Inspectus" library for visualizing LLM attention is set to enhance our understanding and insights into how these models operate, making it an essential tool for developers and researchers alike.
But it's not all about technical marvels and cybersecurity feats. Over in the business world, San Francisco's AI-fueled comeback is attracting billionaire investors, betting big on the city's future despite market challenges. This revival highlights the transformative power of AI in shaping economic landscapes and driving innovation. Whether you're here for the groundbreaking research, the security updates, or the economic insights, 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
First NHS AI-Run Physiotherapy Clinic Launches This Year
NHS launches first AI-run physiotherapy clinic to reduce wait times with same-day digital appointments, improving patient access and care.
California and States Threaten AI Revolution with Overregulation
California and other states' overregulation could stifle AI innovation and competitiveness in the U.S., risking global leadership.
San Francisco's AI-Fueled Comeback Attracts Billionaire Investors
San Francisco's AI-driven revival attracts billionaires betting big on the city's future, despite market challenges.
Apple's AI Push: Home Robots, AR Glasses, and Smart AirPods Unveiled
Apple's AI leap: Unveils home robots, AR glasses, and smart AirPods at WWDC, pushing boundaries of tech innovation.
I made a GPT-4 assistant to master Shopify stores
Created a GPT-4 assistant for Shopify: resolves 70% of support queries, syncs real-time data, easy setup, and customizable. Boost your store now!
Apple WWDC 2024: AI, iOS 18, macOS 15, Siri Updates & More
Apple WWDC 2024: AI, iOS 18, macOS 15, and major Siri updates. Introducing 'Apple Intelligence' for iPhone, iPad, and Mac.
High-Flyer Capital Management: China’s Quant Fund-Turned-AI Pioneer
High-Flyer Capital Management: From quant fund to AI pioneer, leading China's tech innovation.
The Quest for Human-Like Robots
Humanoid robots like Agility's Digit and Boston Dynamics' Atlas are now aiding in real-world tasks, enhancing versatility and efficiency.
Free AI Tool Translates PDFs into 100+ Languages
Translate PDFs into 100+ languages for free with PDFTranslate.ai, preserving layout and images. Trusted by 600M+ users worldwide.
Apple Intelligence: How iPhone's AI Will Work
Apple's WWDC 2024 will unveil "Apple Intelligence," integrating on-device and cloud AI to enhance daily tasks on iOS, iPadOS, macOS, and WatchOS.
Engineering
[Paper] Teams of LLM Agents Exploit Zero-Day Vulnerabilities
Teams of LLM agents can now exploit zero-day vulnerabilities, improving detection and planning by up to 4.5x over previous methods.
🔥 [GitHub] labmlai's LLM Attention Visualization Library "Inspectus" Now Open-Sourced
Labmlai's "Inspectus," a library for visualizing LLM attention, is now open-sourced on GitHub. Enhance your AI insights today!
[GitHub] Ultravox: 200ms Voice LLM by Fixie AI
Fixie AI's Ultravox: a multimodal LLM with 200ms response time, merging audio and text without ASR, built on Meta's Llama 3.
[GitHub] nanoGPT Pipeline in a Spreadsheet by dabochen
nanoGPT in a spreadsheet? Yes! dabochen's project visualizes GPT's inner workings using simple matrices and calculations. Check it out!
[GitHub] Llamanet: Zero-Setup, Zero-Dependency OpenAI Replacement Powered by Llama.cpp
Llamanet: Instantly switch OpenAI apps to Llama.cpp with zero setup or dependencies. One line of code, no 3rd party apps needed!
[GitHub] PaddleOCR: Multilingual OCR Toolkits by PaddlePaddle
PaddleOCR offers lightweight, multilingual OCR toolkits supporting 80+ languages, deployable on various devices.
[Video] Andrej Karpathy: Reproducing GPT-2 (124M) from Scratch
Andrej Karpathy releases a 4-hour YouTube tutorial on reproducing GPT-2 (124M) from scratch, covering everything from network build to optimization.
Text ChatGPT in the UK: Next.js Subscription Starter for SMS and WhatsApp
Text GPT-4 via SMS/WhatsApp in the UK with flexible pricing. Free trial available. Email and Telegram support coming soon!
Build RAG Apps with Jina Embeddings v2 on Amazon SageMaker JumpStart
Build RAG apps with Jina Embeddings v2 on Amazon SageMaker JumpStart for advanced, cost-effective AI solutions. Deploy models with ease!
Academic
🔥 [Paper] Scalable MatMul-Free Language Modeling
New research eliminates MatMul in large language models, reducing memory usage by up to 61% while maintaining top performance at billion-parameter scales.
[Paper] ChatGPT is Bullshit
ChatGPT's inaccuracies aren't hallucinations but bullshit: indifferent to truth, designed to seem human-like, not to convey accurate info.
Propagandists Exploit AI: OpenAI and Meta Expose Global Influence Operations
OpenAI and Meta reveal propagandists from Russia, China, Iran, and Israel using AI for influence campaigns, urging transparency and vigilance.
AI Will Become Mathematicians' 'Co-Pilot'
AI to become mathematicians' "co-pilot," aiding in proof verification and collaboration, revolutionizing traditional solitary math work.
[Paper] Machine Learning Without a Processor: Emergent Learning in Nonlinear Electronic Metamaterials
Analog metamaterials achieve machine learning without processors, enabling fast, low-power tasks like XOR and nonlinear regression.
Microsoft's Bing Chat AI Spreads Election Misinformation
Microsoft’s Bing Chat AI spreads election misinformation, making up false scandals and polling numbers. Microsoft struggles to fix it.