AIR#48 - April 21, 2024
Good morning, AI enthusiasts! As you savor the first sips of your morning coffee, today's edition of AIR: The AI Recon is set to ignite your imagination with tales of innovation, ethical puzzles, and the endless possibilities that AI continues to unfold. At the heart of our digital digest, the breakthrough in "Teaching AI to Think Ahead with Self-Reasoning Tokens" captures the essence of AI's evolving intellect, promising a future where machines can predict and plan with unprecedented precision. This leap towards more autonomous decision-making models marks a significant milestone in our quest to create AI that not only understands but anticipates.
In the bustling world of autonomous vehicles, Mercedes' launch of the first Level 3 self-driving cars in the U.S. heralds a new era of hands-free driving, challenging the boundaries of technology and regulation. This move sparks a fascinating contrast with British Columbia's cautious stance, banning self-driving cars above Level 2 for public safety. These developments paint a vivid picture of the global tug-of-war between innovation and safety, highlighting the diverse approaches to integrating cutting-edge technology into our daily lives.
As we delve deeper into today's stories, the saga of OpenAI's GPT-4 exploiting security vulnerabilities from CVE advisories offers a thrilling glimpse into AI's potential to outsmart existing cybersecurity measures. This revelation not only showcases the prowess of AI in understanding and navigating complex security landscapes but also raises crucial questions about the ethical implications of wielding such power. So, as you flip through the pages of today's edition, let the stories of AI's triumphs and challenges inspire you to ponder the future of this dynamic field, where every breakthrough brings us a step closer to understanding the true potential of artificial intelligence.
Business
Mercedes Launches First Level 3 Autonomous Cars in the U.S. Without Driver Supervision Requirement
Mercedes rolls out the U.S.'s first Level 3 autonomous cars in CA and NV, letting drivers go hands-free without road watch. #DrivePilot
AI Arms Race in Hiring: Candidates and Employers Battle with Bots
AI revolutionizes hiring: Candidates & employers engage in a bot battle, raising questions about the future of recruitment.
British Columbia Bans Self-Driving Cars Above Level 2 for Public Safety
British Columbia bans self-driving cars above Level 2 for safety, limiting autonomous tech use on roads. Fines and penalties apply.
🔥 Exploring the Frontier: LLMs in Financial Markets
LLMs are reshaping financial markets, from predicting stock trends to creating synthetic data, blending AI's power with finance's complexity.
Meta's New AI Tool on Facebook, Instagram: Unstoppable and Everywhere
Meta's new AI tool invades Facebook & Instagram, offering help but can't be turned off, even if it pretends to be a NYC mom.
OpenAI Counters Musk's Claims, Seeks Dismissal of Lawsuit
OpenAI dismisses Musk's lawsuit as "revisionist history," defending its partnership with Microsoft against claims of forsaken principles.
Microsoft's Unleashed Deepfake AI: Too Dangerous to Release
Microsoft's VASA-1 deepfake AI, capable of creating realistic talking videos from images & audio, won't be released due to ethical concerns & potential misuse.
Meta AI Unleashes Creativity on WhatsApp, Instagram, and Messenger
Meta AI transforms WhatsApp, Instagram, & Messenger into creative hubs, offering unique AI-powered features for art, chat, and more. Explore now!
Microsoft and OpenAI's $100B 'Stargate' Venture to Challenge NVIDIA's AI Chip Dominance
Microsoft and OpenAI launch $100B Stargate venture to build AI data centers, challenging NVIDIA's chip dominance by 2028.
AI-Generated James Bond Trailer with Henry Cavill Hits 2.5 Million Views
AI-generated James Bond trailer starring Henry Cavill fools millions, sparking debates on AI's role in fan fantasies and casting rumors.
Engineering
[GitHub] E2B's Open-Source SDK for Custom Code Interpreters with Any LLM
Launch E2B's SDK on GitHub for creating custom code interpreters compatible with any LLM, enhancing AI code execution with Python & JS.
🔥 "Teaching AI to Think Ahead with Self-Reasoning Tokens"
New AI breakthrough: Self-Reasoning Tokens enable models to predict and plan future actions, significantly reducing loss and enhancing long-range reasoning capabilities.
[Github] AdventureAI: Llama3 Powers Interactive Fiction Victory
AdventureAI leverages Llama3, testing AI's prowess in interactive fiction. Despite efforts, AI struggles with complex narratives, hinting at current limitations.
[Github] LLM Scraper: Convert Webpages to Structured Data with OpenAI Models
Transform webpages into structured data effortlessly with mishushakov's LLM Scraper, leveraging OpenAI models.
OpenAI's GPT-4 Demonstrates Ability to Exploit Security Vulnerabilities from CVE Advisories
OpenAI's GPT-4 can exploit 87% of real-world vulnerabilities from CVE advisories, outperforming other models and tools.
Exploring Color Perception: GPT-4 vs. Humans in Naming Shades
GPT-4 mostly aligns with human color naming but struggles with pink, hinting at bias removal or data training quirks. #AIcolorPerception
Junyang Lin Announces Qwen1.5-110B Model Preview on Hugging Face
Junyang Lin invites all to try the new Qwen1.5-110B model on Hugging Face, promising model weights release soon.
[Awni Hannun on X] MLX 0.11 Boosts Generation Speed Across All Model Sizes
MLX 0.11 update: Speeds up AI model generation across sizes, enhancing efficiency. Awni Hannun reveals details.
Exploring Positional Encoding Techniques in Transformers
Exploring the magic of positional encoding in transformers: From sinusoidal waves to learned encodings, RoPE, and ALiBi, each technique casts its own spell on sequence understanding. 🧙♂️✨
How to Use Llama 3 for Free in Production with Cloudflare and OpenAI
Jack Culpan reveals how to use Llama 3 for free in production with Cloudflare and OpenAI, offering a cost-effective alternative to gpt-4-turbo.
[GitHub] CapDec: SOTA Zero-Shot Image Captioning with CLIP & GPT-2 - EMNLP 2022 Findings
CapDec revolutionizes zero-shot image captioning, merging CLIP & GPT-2 for top results without using images, as shown at EMNLP 2022.
Academic
[Paper] Avoiding Machine Learning Pitfalls: A Guide for Academics by Michael A. Lones
Michael A. Lones' guide helps academics dodge common ML pitfalls, ensuring robust model evaluation and fair comparisons.
EURISKO Revived: Unlocking Lenat's Legacy and Running on Medley Interlisp
EURISKO, once a legend, now revived and running on Medley Interlisp thanks to Lenat's archives and dedicated fans. A true digital archaeology triumph.
LLMs Lack Real-World Creativity: A Personal Exploration
Exploring AI's creativity limits: LLMs like ChatGPT struggle with practical creative tasks, despite amusing attempts. Real-world application remains elusive.
Rethinking "User": The Shift Towards More Human-Centric Tech Terminology
Tech's shift from "user" to more human-centric terms reflects evolving digital relationships and the need for precision in our AI-integrated world.