AI NEWS AND VIEWS

[seanmcilroy29]

AI NEWS AND VIEWS

Google DeepMind and 33 labs made the world's largest AI robot training set, combining data recorded from 22 different robots demonstrating more than 500 skills across 150,000 tasks over 1 million episodes including bi-manual and quadruped form factors. The data was transformed into a consistent format, and it's already shortening the learning curve for new robots. This is where the power of AI becomes far more impressive, bridging into the physical world — Link

Hugging Face CEO predicts the future of AI is mostly small AI — Link

Anthropic published a blog explaining a new way of peering inside the black box of LLMs and understanding why they behave the way they do — Link

AI image recognition systems are beginning to resemble human vision performance, including biases and patterns of recognition, implying that these systems may be more closely mimicking human vision — Link

DeepMind makes an AI to automate prompt engineering. So much for that new career path. The moral of the story? Don’t be your career on a bunch of news headlines — Link

“Dead Grandma’s Locket” Prompt-hack convinces Bing to solve a visual security puzzle, demonstrating yet again that LLMs lack context — Link

Researchers fool AI watermarking systems, underscoring that this is not a simple problem to fix — Link

GPT-4 can accurately predict the risk of cardiovascular disease based on common risk factors — Link

OpenAI is looking into making its own chips, No surprise there. Every company that builds its own data centers will likely start looking at buying its own AI chips — Link

Nvidia blog on how AI is helping fight California wildfires. We’ve mentioned this specific use case before, but it's pretty amazing — Link