- Chaos Theory
- Posts
- 🥟 Chao-Down #15: Diffusion models are found to generate copies of their training data
🥟 Chao-Down #15: Diffusion models are found to generate copies of their training data
Plus, Google and Meta double down on Generative AI following earnings.
While today’s headlines are all about the arms race brewing among the tech giants around generative AI, today we feature an interest result from research.
In this paper, researchers found that Stable Diffusion and Google’s Imagen models could generate nearly identical copies of data that they were trained on, including images that are copyrighted and photos based on real people.
The fact that these models can replicate and memorize the original datasets pose real concerns on legitimate use especially if they were built without the subject’s knowledge or permission. To be fair, getting these results are extremely rare (authors claim to see this happen only 0.03% of the time), but given the wide distribution of these image-gen models, such outcomes are bound to happen.
-Alex, your resident Chaos Coordinator.
What happened in AI? 📰
Google is holding an event about search and AI on February 8th (The Verge)
Meta will make generative AI a priority this year following latest earnings call (Axios)
The race of the AI labs heats up (The Economist)
Inside ChatGPT’s Breakout Moment and the Race for the Future of AI (Forbes)
Always be Learnin’ 📕 📖
OpenBioLink/ThoughtSource: A central, open resource for data and tools related to chain-of-thought reasoning in large language models. (Github)
Building a GPT-3 app with Next.js and Vercel Edge Functions (Vercel)
Projects to Keep an Eye On 🛠
salesforce/logai: LogAI - An open-source library for log analytics and intelligence (Github)
Maroofy - Search for any song and it'll use the song's audio to find similar-sounding music. Powered by an AI model trained on 120M+ songs (Twitter)
The Latest in AI Research 💡
Large Language Models Can Be Easily Distracted by Irrelevant Context (arxiv)
How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech (arxiv)
Progressive Prompts: Continual Learning for Language Models (arxiv)
The World Outside of AI 🌎
Eye drops recalled after US drug-resistant bacteria outbreak (ABC News)
J.P. Morgan Asset Management Bought a Forest for $500 million (WSJ)
One Last Bite 😋
One consequence of ChatGPT, the influx of hustle influencers claiming people can get rich by using the tech to automatically generate affiliate link-ridden, click-baity content. Is this the future we want? 🧵
— Alex Chao (@alexchaomander)
3:54 PM • Feb 3, 2023