AIR#109 - June 21, 2024
Good morning, AI enthusiasts! As you sip your morning coffee, prepare to dive into today's edition of AIR: The AI Recon. Leading our headlines is the fascinating shift from LangChain to modular building blocks for AI agent development, as highlighted by Octomind. Their switch underscores the importance of flexibility and productivity in AI projects, making it a must-read for developers looking to streamline their workflows. Meanwhile, Claude 3.5 Sonnet is turning heads in the business world with its blazing speed and enhanced vision capabilities, promising to redefine what we can expect from AI models. This is an exciting development for anyone keeping tabs on the latest advancements in AI performance.
But that's not all! Tim O'Reilly's thought-provoking piece on "AI's Original Sin" delves into the complex intersection of AI and copyright, proposing a balanced approach that could pave the way for more sustainable and transparent AI practices. It's a compelling read for scholars and anyone interested in the ethical dimensions of AI. And speaking of groundbreaking research, Character.AI's innovative inference optimizations have slashed costs by a staggering 33x while handling an impressive 20,000 queries per second, showcasing the relentless push for efficiency in the AI industry.
On a more practical note, IMG.LY's use of ONNX Runtime with WebGPU is revolutionizing browser-based background removal, achieving near real-time performance at 20x faster speeds. This is great news for developers working on web applications that require quick and efficient image processing. Whether you're here for the latest tech innovations, ethical debates, or practical applications, 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
🔥 Claude 3.5 Sonnet: Frontier AI with 2x Speed and Enhanced Vision
Claude 3.5 Sonnet: 2x faster, enhanced vision, and top-tier intelligence. Available now on Claude.ai, iOS, and major cloud platforms.
Apple Is Bringing A.I. To Your Personal Life, Like It or Not
Apple introduces "Apple Intelligence," a cautious AI for iPhones, integrating deeply into personal lives but raising privacy and control concerns.
AI Data Analyst: Chat with Your Data for Instant Insights | Narrative BI
Chat with your data using Narrative BI's AI Data Analyst for instant, actionable insights—no complex queries needed. Try it free for 7 days!
Qualcomm AI/Copilot PCs Fail to Deliver
Qualcomm's AI/Copilot PCs fall short of expectations, with issues in performance, compatibility, and security overshadowing the hype.
The Rise of AI-Powered Killer Robot Drones
Eric Schmidt's startup builds $400 AI drones for autonomous targeting and attacks, sparking global security concerns.
Wired Confirms Perplexity Is Bypassing Efforts by Websites to Block Its Crawler
Wired reveals Perplexity bypasses website blocks, scraping data despite claims of respecting robots.txt. Ethics and transparency questioned.
OpenAI Co-Founder Ilya Sutskever Launches Rival AI Startup Safe Superintelligence Inc
OpenAI co-founder Ilya Sutskever launches Safe Superintelligence Inc, a new AI startup focused solely on building safe superintelligence.
London Premiere of AI-Written Movie Cancelled After Backlash
London cinema cancels AI-written film premiere after backlash over replacing human writers with AI.
Engineering
🔥 Why We Stopped Using LangChain for AI Agents
LangChain's rigid abstractions complicated our AI agent development, so we switched to modular building blocks for simplicity and productivity.
[GitHub] Local Voice Assistant: Ollama + HF Transformers + Coqui TTS
Local voice assistant "June" uses Ollama, Hugging Face Transformers, and Coqui TTS for privacy-focused, offline voice interactions.
Optimizing AI Inference at Character.AI: Pioneering AGI Efficiency
Character.AI boosts AGI efficiency with groundbreaking inference optimizations, cutting costs 33x since 2022 and handling 20,000 queries/sec.
20x Faster Browser Background Removal with ONNX Runtime | IMG.LY
IMG.LY's new ONNX Runtime with WebGPU enables 20x faster browser-based background removal, achieving near real-time performance.
Is LMDeploy the Ultimate Solution? Benchmarking LLM Inference Backends: vLLM, LMDeploy, MLC-LLM, TensorRT-LLM, and TGI
LMDeploy leads in decoding speed and low latency for Llama 3 models, making it ideal for high-load scenarios, while vLLM excels in low TTFT.
OpenPipe MoA: Outperform GPT-4 at 1/25th the Cost
OpenPipe's MoA models outperform GPT-4 at 1/25th the cost, offering significant improvements in synthetic data generation and fine-tuning.
[GitHub] Jan: Offline, Open-Source ChatGPT Alternative for Your PC
Offline, open-source ChatGPT alternative "Jan" runs locally on your PC with multiple engine support. Explore at jan.ai.
[GitHub] Groqnotes: Structured Audio Notes with Groq, Whisper, and Llama3
Generate structured notes from audio with Groqnotes, using Groq, Whisper, and Llama3 for fast, organized transcription. 🎧📖
Garry Tan Calls for "Perplexity Meets Kindle" Product; Nouswise Delivers
Garry Tan's call for a "Perplexity Meets Kindle" product is answered by Nouswise, offering limitless books, notes, and citations. Check it out!
Peneterrer – The ChatGPT for Web App Security Testing
Peneterrer: The AI-driven tool for web app security testing, offering workflow creation and periodic reports. Secure your site effortlessly!
Introducing Claude 3.5 Sonnet—Anthropic's Most Intelligent Model Yet
Meet Claude 3.5 Sonnet: Anthropic's smartest model, twice as fast, one-fifth the cost. Try it for free!
OTel and AI Observability Tool for Startups: Iudex
Iudex offers startups a quick, cost-effective observability platform with easy setup, real-time service health, and natural language search.
Academic
How to Fix “AI’s Original Sin” – Tim O’Reilly
Tim O'Reilly proposes a balanced approach to AI and copyright, suggesting fair use, transparency, and a sustainable business model for all parties.
[Google DeepMind] Generating Audio for Video with V2A Technology
Google DeepMind's V2A tech uses video pixels and text prompts to generate synchronized, realistic soundtracks for silent videos.
[Paper] RAR-B: Reasoning as Retrieval Benchmark by Xiao, Hudson, Al Moubayed
RAR-b benchmark reveals retriever models struggle with reasoning tasks, highlighting a gap in current LLM capabilities. Fine-tuning helps.
Ex-OpenAI Star Sutskever Aims for Superintelligent AI with New Company
Ex-OpenAI scientist Ilya Sutskever launches Safe Superintelligence, Inc. to safely develop superintelligent AI beyond human capabilities.
How to Detect ChatGPT Confabulations
Oxford researchers find a way to detect when ChatGPT is making up info, improving trust in AI-generated responses.