AIR#107 - June 19, 2024

Good morning, AI aficionados! As you sip your morning coffee, get ready to dive into today's edition of AIR: The AI Recon. Leading the headlines is Meta FAIR's latest release of groundbreaking AI research, models, and datasets. This treasure trove is set to foster innovation, creativity, and responsible AI development, making it a must-read for anyone looking to stay ahead in the AI game. And if you're intrigued by the latest in AI performance, don't miss out on Yandex's YaFSDP, which promises to boost LLM pre-training by up to 20%, outperforming existing methods under high memory pressure conditions. It's all available on GitHub for the curious developers out there.

But that's not all! In a fascinating turn of events, China's DeepSeek Coder V2 has outperformed GPT-4 Turbo in coding and math tasks, supporting an impressive 338 programming languages. This open-source marvel is making waves and could be a game-changer for developers seeking robust coding solutions. Meanwhile, the scholarly community is buzzing about a new paper revealing that a single direction in language models can control refusal behavior, exposing vulnerabilities in safety fine-tuning. It's a thought-provoking read that underscores the complexities of AI safety and ethics.

And for those keeping an eye on global tech developments, Ukraine is leveraging AI to tackle a monumental 700-year de-mining challenge, using cutting-edge technology to prioritize landmine removal from Russian invasions. This inspiring story highlights the real-world impact of AI in solving long-standing issues. Whether you're here for the groundbreaking tech updates, the ethical debates, or the 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

Ukraine AI to Clear 700 Years of Russian Landmines
Ukraine uses AI to prioritize landmine removal, tackling a 700-year de-mining challenge from Russian invasions.

China's AI-Powered Sexbots Ready to Hit Shelves
China's AI-powered sexbots, designed for emotional interaction, are set to launch soon, raising ethical and technical challenges.

Bacon Ice Cream and Nugget Overload: McDonald's Pulls AI Drive-Throughs
McDonald's scraps AI drive-throughs after hilarious order fails like bacon ice cream and nugget overload. Tech will be re-evaluated.

Waymo's Robotaxis Are Finally a Real Business for Google
Waymo's robotaxis are now a real business, generating significant revenue in major cities despite ongoing safety concerns.

State of AI Report 2024
State of AI 2024: AI adoption is rising but still has a long way to go. Leadership wants more investment, but real ROI is just beginning.

Engineering

🔥 [GitHub] YaFSDP by Yandex: Faster Sharded Data Parallelism for LLM Pre-Training
Yandex's YaFSDP boosts LLM pre-training by up to 20%, outperforming FSDP in high memory pressure conditions. Check it out on GitHub!

🔥 Every Way to Get Structured Output from LLMs
Discover methods to get structured output from LLMs, from JSON parsing to token constraints. Find the best frameworks for your needs!

China's DeepSeek Coder V2 Outperforms GPT-4 Turbo
China's DeepSeek Coder V2, an open-source AI, outperforms GPT-4 Turbo in coding and math tasks, supporting 338 programming languages.

Intel's OpenVINO 2024.2: Expanded LLM & Quantization Support
Intel's OpenVINO 2024.2 boosts LLMs with expanded support, improved quantization, and new model optimizations for better AI performance.

AI Background Remover in Your Browser | HTML.ZONE
Remove image backgrounds instantly in your browser with HTML.ZONE's AI tool. No downloads needed!

Neural Networks Explained with Minimal Math
Neural networks demystified with simple diagrams: Learn how they map inputs to outputs and the role of hidden layers and activation functions!

Suno Unveils 'Audio Inputs' Feature for Custom Song Creation
Suno's new 'Audio Inputs' lets Pro & Premier users create custom songs from any sound, sparking creativity and community engagement.

[Github] Yandex's YaFSDP: Turbocharges LLM Training, Optimizes GPU Use
Yandex's YaFSDP tool speeds up LLM training and optimizes GPU use, outperforming ZeRO and FSDP methods. Now open-sourced! 🚀

AI-Powered GPT-4 Manga Translator Revolutionizes Reading Experience
AI-powered Manga Translate offers instant, accurate multi-language manga translations, revolutionizing the reading experience for fans worldwide.

[GitHub] BigCodeBench: Advanced Benchmark for Code Generation
GitHub's BigCodeBench offers a new benchmark for evaluating code generation with complex tasks and diverse function calls.

[Paper] Joint Audio and Symbolic Conditioning for Text-to-Music Generation
JASCO: New AI model generates music from text, combining symbolic and audio cues for precise control over musical elements. 🎶

A Tour of Differentiable Rasterization
Explore Differentiable Rasterization: Convert vector graphics to pixel images using JAX, enabling gradient descent for complex image editing!

Academic

🔥 [Paper] Single Direction Mediates Refusal in Language Models
A single direction in language models controls refusal behavior, revealing vulnerabilities in safety fine-tuning.

🔥 Meta FAIR Releases Groundbreaking AI Research, Models, and Datasets
Meta FAIR releases groundbreaking AI research, models, and datasets to foster innovation, creativity, and responsible AI development.

Why Does AI Hallucinate? | MIT Technology Review
AI chatbots hallucinate because they predict words based on patterns, not facts, leading to plausible but incorrect answers.

[Paper] A Robot Walks into a Bar: Evaluating LLMs as Comedy Creativity Tools
LLMs fail as comedy tools, producing bland, biased jokes. Comedians find them restrictive, reinforcing hegemonic views.

[Paper] RLHF Reduces LLM Creativity and Output Variety
RLHF reduces LLM creativity and output variety, limiting their ability to generate diverse, non-toxic content.

[Paper] PlatoNeRF: 3D Reconstruction in Plato's Cave via Single-View Two-Bounce Lidar
PlatoNeRF uses single-view two-bounce lidar to achieve 3D scene reconstruction, overcoming occlusions and ambient light challenges.

UNESCO Report: Generative AI Threatens Holocaust Memory
UNESCO warns generative AI could distort Holocaust history, fueling antisemitism. Urges ethical AI use to preserve historical accuracy.

[Paper] Transcendence: Generative Models Surpass Their Trainers
Generative models now surpass their trainers, achieving superior performance in tasks like chess. Study reveals how AI transcends human expertise.