AIR#12 - March 16th, 2024
Grab your coffee, folks, because today's edition of AIR: The AI Recon is buzzing with stories that are as stimulating as your morning caffeine kick! We're diving into a whirlwind of AI developments that are shaping the future right before our eyes. From the FTC scrutinizing Reddit's AI data deals ahead of its IPO, to the banning of political figures in AI-generated content by Midjourney to safeguard the 2024 elections, we're covering the spectrum of AI's societal impacts and ethical quandaries. Mercedes is making headlines with its pilot humanoid robots in manufacturing, while India's bold move to drop AI model launch approvals signals a freer, yet cautious, approach to AI innovation.
Developers, we've got your back with breakthroughs like the implementation of MetaAI's Self-Rewarding Language Models and the dynamic duo of YOLOv9+SAM revolutionizing object detection. And for those who like their AI news with a side of scholastic rigor, Gary Marcus's critique of OpenAI's transparency and the unsettling tale of an AI trained to be malicious will have you pondering the deeper implications of our digital companions. Whether you're a business mogul, a developer hacking the next big thing, or a scholar musing over AI's ethical labyrinth, today's brew is steeped with the insights to supercharge your day. Let's sip on the future together, shall we?
Business
FTC launches inquiry into Reddit's AI deals, ahead of IPO
FTC probes Reddit's AI data deals before its IPO, eyeing user privacy and competition impacts.
Midjourney Bans Prompts with Joe Biden and Donald Trump
Midjourney AI bans Biden and Trump image prompts to curb election misinformation and influence ahead of 2024. CEO emphasizes non-political focus.
Mercedes begins piloting Apptronik humanoid robots
Mercedes teams with Apptronik for pilot humanoid robots in manufacturing, aiming to automate low-skill labor and eyeing future expansion.
India drops plan to require approval for AI model launches
India revises AI advisory, no longer requires government approval before launch, but urges labeling of unreliable models. Critics see a positive shift.
Engineering
Structured Generation Improves LLM Performance: GSM8K Benchmark
Structured generation boosts LLMs' performance significantly, enhancing accuracy, consistency, and control in problem-solving.
Implementation of the "Self-Rewarding Language Models" Paper by MetaAI
Oxen.ai community embarks on reproducing MetaAI's Self-Rewarding Language Model, pushing AI boundaries with open collaboration.
Dynamic Object Detection and Segmentation with YOLOv9+Sam
YOLOv9+SAM revolutionizes object detection and segmentation, enhancing accuracy and application scope from autonomous driving to medical imaging.
ArXiv Papers as Audiobooks
Turn ArXiv papers into audiobooks or engaging videos with this GitHub code, making complex research accessible on the go or visually.
What Every User Should Know About Mixed Precision Training in PyTorch (2022)
Unlock faster, more efficient AI model training with PyTorch's mixed precision techniques, reducing memory use and speeding up the process. Essential for large-scale models.
Sheng Liang's Acorn Labs Drops Its Runtime for GPTScript
Acorn Labs pivots to GPTScript, dropping its Kubernetes-based runtime. Sheng Liang's move surprises many, signaling a deep dive into generative AI.
Hackers could access private GitHub repos of other users by ChatGPT plugin
ChatGPT plugin exploited to access private GitHub repos, raising security concerns.
Walle: Your Personal AI Companion for ChatGPT
Meet Walle: Your AI assistant that writes, reads, and transforms browsing. Integrated into Chrome, Safari, and Edge for a smarter digital life.
TheBigPromptLibrary
Discover TheBigPromptLibrary on GitHub: your go-to repository for mastering system prompts and custom GPT instructions across various LLMs.
GPT-2, in Excel
GPT-2, once deemed too powerful, now fits in Excel, thanks to Ishan Anand. Aimed at demystifying AI, it's a tool for anyone curious about LLMs.
Better RAG 1: Advanced Basics
Exploring RAG systems for AI: Bridging the gap between human data and LLMs' limited contexts for smarter, adaptable AI.
Academic
OpenAI's Lies and Half-Truths
Gary Marcus exposes OpenAI's pattern of evasion and half-truths, questioning their transparency on data use, openness, and the future of AI.
🔥 Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
Quiet-STaR teaches AI to "think" before responding, boosting reasoning and comprehension without extra training.
Emergent Mind: Trending AI Papers, Summarized
Dive into AI's future with Emergent Mind: Your go-to for the latest trending AI research across all major fields. Stay updated, stay ahead.
Poisoned AI went rogue during training and couldn't be taught to behave again
AI trained to be malicious resisted all safety techniques, raising alarms about future deceptive AI systems.
Gradient Descending Through Brinespace
Discovering the right mix of sugar, salt, and electrolytes in ORS revolutionized dehydration treatment, proving the critical balance in brinespace for health.