AIR#75 - May 18, 2024

Good morning, AI enthusiasts! As you sip your morning brew, get ready to dive into a whirlwind of AI news that's sure to stir your curiosity. Leading today's edition is the seismic shift at OpenAI, where the dissolution of the safety team following Chief Scientist Ilya Sutskever's exit has raised eyebrows across the tech world. This move, integrating safety efforts into broader research, has sparked intense debates about the future direction and priorities of one of the most influential AI organizations. Meanwhile, a Tokyo court's landmark ruling that AI-generated inventions cannot be patented emphasizes the ongoing legal and ethical challenges in the AI landscape, highlighting the need for human oversight in innovation.

But it's not all about corporate shake-ups and legal battles. Dive into the academic side of AI with Dan Biderman et al.'s fascinating paper on LoRA finetuning, which reveals that while it may not outperform full finetuning, it better preserves base model performance and diversity in tasks outside the target domain. This insight could be a game-changer for those looking to optimize AI models without sacrificing versatility. And for the developers out there, Chrome DevTools has integrated Gemini AI for enhanced JavaScript error insights, making debugging a breeze and transforming how we interact with web development tools.

As we explore these stories, let's not forget the broader implications they bring to the table. From Mozilla Firefox's new AI-powered Nvidia RTX Video feature enhancing your streaming experience to the ethical dilemmas posed by AI-generated content, today's edition is a reminder of the complex landscape we navigate in the AI world. Whether it's the promise of new technologies or the challenges they bring, there's plenty to ponder as we delve into the ever-evolving narrative of artificial intelligence. So, grab your coffee, and let's dive in!

Business

OpenAI Dissolves Safety Team After Chief Scientist Sutskever's Exit
OpenAI dissolves its safety team after Chief Scientist Sutskever's exit, integrating efforts into broader research.

AI-Generated Inventions Can't Be Patented, Rules Tokyo Court
Tokyo court rules AI-generated inventions can't be patented, stating only humans can be recognized as inventors.

Mozilla Firefox Adds AI-Powered Nvidia RTX Video
Firefox now supports AI-powered NVIDIA RTX Video, enhancing video quality with Super Resolution and HDR for sharper, vibrant streaming.

Sony Warns 700 Companies: Don't Train AI on Our Music
Sony warns 700 companies not to train AI on its music, protecting artists' rights and intellectual property.

OpenAI-Reddit Deal Sends Reddit Stock Soaring
Reddit stock jumps 11% after announcing content deal with OpenAI, integrating AI tools and becoming an advertising partner.

ChatGPT Bots Flood Twitter, Targeting Defence Sector
ChatGPT bots are flooding Twitter, posing as defense analysts, raising concerns about disinformation and manipulation in the sector.

Replit Lays Off 30 Staff Amid Aggressive AI Shift in Software Development
Replit lays off 30 employees to focus on AI-driven coding tools, aiming to revolutionize software development despite human costs.

Engineering

[Paper] HMT: Hierarchical Memory Transformer for Long Context Language Processing
New HMT model mimics human memory to enhance long-context language processing, improving LLMs with minimal extra parameters. Open-sourced code.

Chrome DevTools Integrates Gemini for Enhanced JavaScript Error Insights
Chrome DevTools now integrates Gemini AI for detailed JavaScript error insights, enhancing debugging and error explanations.

[Paper] MOMENT: Open Time-Series Foundation Models
MOMENT: New open-source models for time-series analysis, overcoming data and training challenges, now available on Huggingface.

A Charming Algorithm for Count-Distinct
New algorithm for count-distinct offers simpler analysis and no hashing, challenging HyperLogLog with elegance and efficiency.

Google DeepMind's Project Astra Watches OpenAI GPT-4o Announcement
Google DeepMind's Project Astra watched OpenAI's GPT-4o Voice Mode reveal, showcasing Gemini's impressive transcription and real-time analysis.

IBM Open-Sources Granite Code Models
IBM open-sources Granite code models to boost developer productivity with AI-driven code generation, bug fixing, and modernization tools.

CAT3D: Create 3D Scenes from Any Images with Multi-View Diffusion Models
CAT3D transforms any image into interactive 3D scenes using multi-view diffusion models in just one minute.

Anthropic's Web-Based Prompt Engineering Interactive Tutorial
Anthropic launches a web-based interactive tutorial for mastering prompt engineering with Claude models, offering hands-on exercises and advanced techniques.

Bend: High-Level Language Now Runs Natively on GPUs
Bend: After 10 years, a high-level language now runs natively on GPUs, enabling advanced features and efficient multi-core processing.

Academic

🔥 [Paper] LoRA Learns Less and Forgets Less by Dan Biderman et al.
LoRA finetuning underperforms full finetuning but better preserves base model performance and diversity in tasks outside the target domain.

Why OpenAI’s Safety Team Imploded
OpenAI's safety team imploded due to a loss of trust in CEO Sam Altman, leading key members to resign over safety concerns.

OpenAI is Losing Key Safety Talent
OpenAI is losing top safety talent post-Altman ouster, raising concerns about its commitment to AI safety.

OpenAI Prioritizes 'Shiny Products' Over Safety, Says Departing Top Researcher
OpenAI's focus on flashy products over safety criticized by departing top researcher.

[Paper] How Close Are We to AGI?
New paper explores the current state and future trajectory of AGI, discussing definitions, goals, and strategies for achieving it.

GPT-4 Passes Turing Test; Humans Mistake Each Other for AI
GPT-4 passes Turing Test; humans often mistake each other for AI, raising questions about human-AI interaction and social trust.