🥟 Chao-Down #344 Latest Stable Diffusion image generation model has a hard time generating people, Spotify launches in-house creative agency to test AI-generated ads, Microsoft recalls "Recall"
Plus, a look at the growing popularity of AI tools among students and teachers.
Are AI image generators getting worse? At least for generating images with people, it seems like aggressive data filtering is to blame.
Stability AI’s latest release, Stable Diffusion 3 Medium, has faced ridicule and disappointment by the AI image-synthesis community.
The model, which aims to generate images from text prompts, struggles with rendering human anatomy accurately, often producing bizarre and incorrect visuals, particularly with hands and feet. The regression in quality seems due to the overzealous filtering of adult content from the training data, which inadvertently removed non-offensive human anatomy examples.
As with many things with AI, data is king. And bad data (or insufficient data) will produce bad output
-Alex, your resident Chaos Coordinator.
What happened in AI? 📰
New Stable Diffusion 3 release excels at AI-generated body horror (Ars Technica)
Microsoft to delay release of Recall AI feature on security concerns (MSN)
Spotify announces an in-house creative agency, tests generative AI voiceover ads (TechCrunch)
AI is getting very popular among students and teachers, very quickly (CNBC)
What Gen Z wants from AI policymakers (Semafor)
Giant Chips Give Supercomputers a Run for Their Money (IEEE Spectrum)
Always be Learnin’ 📕 📖
End-to-end LLM Workflows Guide (anyscale.com)
Can LLMs invent better ways to train LLMs? (sakana.ai)
microsoft/generative-ai-for-beginners: 18 Lessons, Get Started Building with Generative AI (Github)
Projects to Keep an Eye On 🛠
openrecall/openrecall: OpenRecall is a fully open-source, privacy-first alternative to proprietary solutions like Microsoft's Windows Recall. (Github)
NVIDIA Releases Open Synthetic Data Generation Pipeline for Training Large Language Models (NVIDIA Blog)
twentyhq/twenty: Building a modern alternative to Salesforce, powered by the community. (Github)
The Latest in AI Research 💡
VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers (arxiv)
Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization (arxiv)
Easy Problems That LLMs Get Wrong (arxiv)
The World Outside of AI 🌎
The Health Benefits of Relaxation (TIME)
Why Do Rich People Love Quiet? (The Atlantic)
Why doesn’t the US have better sunscreen? (Vox)
‘Brainrot’ Is the New Online Affliction (The New York Times)
Solar-Powered Planes Are Ready to Take Off (And Fly for Months at a Time) - WSJ
Why Is Everyone Getting Sick? Behind the Global Rise in RSV, Flu, Measles - (Bloomberg)
One Last Bite 😋
