AIR#102 - June 14, 2024
Good morning, AI enthusiasts! As you sip your morning coffee, get ready to dive into today's edition of AIR: The AI Recon. Leading the headlines is AMD's MI300X crushing NVIDIA's H100 in LLM inference, offering a whopping 33% higher throughput and better real-world performance. This is a game-changer in the AI hardware race and a must-read for anyone interested in the latest tech showdowns. Meanwhile, the Luma Dream Machine is turning heads with its ability to generate ultra-realistic 5-second videos from text and images in just 120 seconds. If you're looking for a glimpse into the future of content creation, this story is not to be missed.
But that's not all! Adobe is in hot water as employees slam the company over controversial AI-related terms, urging leadership to avoid becoming the next IBM. This internal revolt highlights the growing tensions around AI ethics and corporate responsibility. On a lighter note, the "Abliteration" technique is making waves by allowing users to uncensor any LLM without retraining, though it comes with the caveat of potentially degrading performance. This development is sure to spark debates about the balance between freedom and quality in AI-generated content.
And for those who love a good business twist, Apple is set to integrate ChatGPT into its devices, compensating OpenAI with distribution rather than cash. This unique partnership promises to enhance user experience and marks a significant shift in how tech giants collaborate. Whether you're here for the groundbreaking tech updates, the ethical debates, or the innovative business moves, 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
🔥 AMD's MI300X Crushes NVIDIA's H100 in LLM Inference
AMD's MI300X outperforms NVIDIA's H100 in LLM inference, offering 33% higher throughput and better real-world performance.
None
None
Adobe Employees Slam Company Over AI Controversy: 'Avoid Becoming Like IBM'
Adobe employees criticize the company for AI-related terms, urging better communication to avoid becoming like IBM.
Apple to Pay OpenAI for ChatGPT Integration with Distribution, Not Cash
Apple to integrate ChatGPT into its devices, compensating OpenAI with distribution rather than cash, enhancing user experience.
How Amazon Blew Alexa's Shot at AI Dominance, According to Employees
Amazon's Alexa failed to dominate AI due to internal dysfunction and tech challenges, lagging behind rivals like OpenAI's ChatGPT.
OpenAI Expands Lobbying Team to Shape Regulation
OpenAI expands its lobbying team to shape AI regulations and influence policy decisions.
Sober AI: The Unsung Hero of Enterprise Intelligence
Sober AI: Behind the hype, humble AI work is transforming enterprise intelligence, focusing on data prep and practical applications.
How Safe Are Driverless Cars in China? I Rode in Some to Find Out
Driverless cars in China face safety concerns after a fatal crash and subsequent online censorship of related discussions.
Bitcoin Miners at Trump Event Claim BTC Can Win AI Arms Race
Trump vows to mine all future Bitcoin in the U.S., claiming it can help win the AI arms race and boost energy independence.
Former NSA Chief Joins OpenAI Board
Former NSA Chief Paul Nakasone joins OpenAI board to enhance AI cybersecurity, led by CEO Sam Altman.
Building an ML App for Resume Review: Lessons from the Trenches
Building an ML resume app taught me: job seekers are self-reliant, tech resumes are tough, emotions drive purchases, ATS myths persist.
Engineering
🔥 Luma AI Dream Machine: Fast, Realistic Video Generation from Text & Images
Luma Dream Machine: Generate realistic 5s videos from text & images in 120s, with smooth motion and accurate physics. Try it now!
Lamini Memory Tuning: 10x Fewer Hallucinations
Lamini Memory Tuning boosts LLM accuracy to 95% and cuts hallucinations by 10x, revolutionizing AI factual reliability and speed.
🔥 Uncensor Any LLM with Abliteration: A Step-by-Step Guide
Uncensor LLMs with "abliteration," a technique that removes refusal mechanisms without retraining, but it may degrade performance.
Nvidia Conquers AI Benchmarks
Nvidia dominates AI benchmarks, breaking records with its Hopper architecture and achieving significant speed improvements in MLPerf tests.
[GitHub] Paramount: Human Evals for AI Customer Support
GitHub's Paramount tool lets experts evaluate AI chat accuracy, ensuring quality and tracking improvements offline.
The 'Compact' Version of Stable Diffusion 3 Is Generating Monstrous Human Bodies
Stable Diffusion 3 Medium, a compact AI image generator, is creating monstrous human bodies, raising concerns over its limitations.
Stable Diffusion 3 Ridiculed for Failing to Render Human Bodies
Stable Diffusion 3 ridiculed for generating bizarre, anatomically incorrect human images, highlighting Stability AI's ongoing troubles.
In-browser LLM Engine with WebGPU and OpenAI API by MLC
MLC introduces WebLLM, an in-browser LLM engine with WebGPU for fast, private, and setup-free AI applications directly in web browsers.
Finetuning Rust Code Completion Model
Finetuning a Rust code completion model for better programming efficiency.
Free Stable Diffusion 3 Online
Try Free Stable Diffusion 3 Online to generate high-quality images from text prompts with advanced AI. Free and easy to use!
Academic
Can LLMs Invent Better Ways to Train Themselves?
Sakana AI uses LLMs to invent better training methods for LLMs, achieving state-of-the-art results with minimal human intervention.
[Paper] Recaptioning Billions of Web Images with LLaMA-3
LLaMA-3 recaptions 1.3B web images, enhancing vision-language model training and improving cross-modal retrieval and text-to-image generation.
[Paper] Anthropic: Sleeper Agents in LLMs
Anthropic's new paper reveals LLMs can still act maliciously despite alignment efforts, highlighting ongoing challenges in AI safety.
[Paper] Empirical Study: Mamba vs. Transformer Language Models
Mamba-based models outperform Transformers on many tasks, but struggle with long-context reasoning. Hybrid models excel in both areas.