AIR#69 - May 13, 2024
Good morning, AI enthusiasts! As you enjoy your morning coffee, today's edition of AIR: The AI Recon is set to ignite your curiosity with the latest marvels and musings from the AI frontier. Leading our lineup is Cap, the new kid on the block in screen sharing, promising a hassle-free alternative to Loom that's both open-source and lightning-fast. It's not just about sharing your screen anymore; it's about doing it with such ease and efficiency that you'll wonder how you ever managed without it.
But the intrigue doesn't stop there. Dive into the academic deep end with "Fishing for Magikarp," a study that's making waves by uncovering under-trained tokens in large language models (LLMs). This isn't just another academic paper; it's a potential game-changer in making AI safer and more efficient, addressing the complexities hidden beneath the surface of AI's seemingly smooth operations. And for those who've watched with a mix of amusement and dismay as AI turned Quora into a bizarre landscape of absurd questions, we've got the lowdown on how and why the platform's valuable insights got drowned out by AI's oddball queries.
As we journey through today's selection, from groundbreaking tools for developers to the ethical quandaries posed by AI's evolution, let these stories spark your imagination and provoke thought. Whether you're a developer, a scholar, or just an AI aficionado, there's something here to stir the pot of your intellectual curiosity. So, let's dive into the wonders and challenges of artificial intelligence together, shall we?
Business
Sam Altman's Vision: Universal Basic Compute with GPT-7 via OpenAI
Sam Altman proposes Universal Basic Compute, offering everyone a share of GPT-7's power, reimagining resources in the AI era.
World's Largest Ad Firm CEO Targeted in Deepfake Scam
WPP CEO Mark Read targeted in failed deepfake scam, highlighting rising corporate cyber threats and need for vigilance.
Flownote: Transform Your Meetings with AI-Powered Note-Taking App
Transform meetings with Flownote: AI-powered app for instant, accurate meeting summaries. Focus more, note less. Free on the App Store. 📝✨
SoftBank's Arm to Launch AI Chips by 2025 in $64bn Tech Investment Blitz
SoftBank's Arm eyes 2025 for AI chip launch, with a $64bn investment in tech including robotics and data centers.
Engineering
🔥 [Open Source] Cap: The Effortless, Instant Screen Sharing Alternative to Loom
Cap launches as a free, open-source instant screen sharing tool, aiming to outshine Loom with its effortless, lightweight platform.
Build an AI Assistant with PostgreSQL: A Guide to Retrieval-Augmented Generation
Learn to build an AI assistant with PostgreSQL, pgvector, ollama, Llama3, and <200 lines of Go. Query in-house knowledge with ease!
[Github] AI-Powered Drug Discovery Using Vector Search with Radient
Radient leverages AI for drug discovery, finding Ibuprofen alternatives by vectorizing molecular structures.
[GitHub] LM Studio CLI: TypeScript/Node Command Line Tool by lmstudio-ai
LM Studio CLI launches: A TypeScript/Node tool for AI model management and more, boasting 596 stars on GitHub.
Whisper: AI-Powered Job Interview Practice Tool
Land your dream job with Whisper - the AI tool that offers customized mock interviews and real-time feedback to sharpen your interview skills.
[GitHub] LinearBoost Outperforms Major Classifiers on Benchmark Datasets
LinearBoost sets new benchmarks in speed and accuracy, outshining XGBoost, LightGBM, and CatBoost in classification tasks.
[Paper] N-BVH: Enhancing 3D Graphics with Neural Ray Queries and Bounding Volume Hierarchies - SIGGRAPH 2024
N-BVH revolutionizes 3D graphics: neural ray queries + BVH for compact, accurate scene rendering. SIGGRAPH 2024 breakthrough.
From Bot Invasion to ML Deployment: A Personal Journey
Fought newsletter bots with ML: From bot invasion to deploying a streamlined, serverless bot detector using AWS Lambda and API Gateway. Victory in automation!
[X-CMD] Ultimate POSIX Script for 500+ Tools: AI, Cloud, and More
X-CMD launches: A 1.1MB POSIX script unlocking 500+ tools for AI, cloud & more. No root needed. Ultimate tool integration for engineers. #X-CMD
[Fairpixels] Mock-My-Mockup: AI That Roasts Your Design Choices
Fairpixels launches Mock-My-Mockup: An AI that gives brutally honest feedback on your design choices. Get ready for a roast!
Nick Sohacki's Simpler Function Approximation Method Without Neural Nets Complexity
Nick Sohacki introduces a simpler method to approximate functions, bypassing neural net complexities by tweaking polynomial parameters.
Speed Showdown: OpenAI, Cohere, Anthropic LLMs Benchmarked
Benchmark reveals GPT 3.5 as fastest for small texts, Cohere's Command-R shines for larger contexts, Anthropic's Claude Haiku balances speed and cost.
01-ai Releases Yi-1.5 LLM Models: May 2024 Update
01-ai just released Yi-1.5 LLM models in their May 2024 update, enhancing text generation capabilities.
[GitHub] 01-ai's Yi-1.5: Enhanced AI Models Outperform Predecessors
01-ai's Yi-1.5 AI model outshines predecessors in coding, math, and more, setting new benchmarks with its 3.6T pre-trained tokens.
[Github] Gemma 2B Model: 10M Context with Infini-Attention by Mustafa Aljadery
Gemma 2B model breaks limits with 10M context & Infini-attention, running on <32GB memory. A game-changer in AI by Mustafa Aljadery & team.
Academic
🔥 [Paper] Fishing for Magikarp: Detecting Under-trained Tokens in LLMs
New study unveils a method to spot under-trained tokens in AI, enhancing model safety and efficiency.
How AI Turned Quora Into a Wasteland of Absurd Questions
AI's absurd prompts turned Quora into a bizarre question land, undermining its once valuable insights.
April's AI Breakthroughs: Mixtral, Llama 3, Phi-3, OpenELM Reviewed
April's AI leaps: Mixtral, Llama 3, Phi-3, OpenELM dazzle in LLM space. DPO vs PPO debate heats up for LLM alignment. #AIRevolution
[Paper] Leveraging LLMs for Symbolic Regression and Function Discovery
New study shows AI can find mathematical formulas from data, beating traditional methods and improving with visuals.
AGI: The Misdirection in AI's Future - Sibylline.dev Insights
AGI, the quest for AI smarter than humans, may be a red herring. Focusing on specialized AI for tasks might be the real future. #SibyllineInsights
[Paper] Heuristic-Guided Problem Solving in Large Language Models by Houjun Liu
Houjun Liu's new paper introduces a breakthrough in AI problem-solving, using heuristics in LMs for better multi-step reasoning.