A week in Generative AI: Llama 3, Atlas & Dove
A week in Generative AI: Llama 3, Atlas & Dove
News for the week ending 21st April 2024
Big news from Meta this week with the announcement of Llama 3 which will ship in Meta AI across all their products in the coming months. Microsoft announced a new model called VASA-1 that could enable photorealistic digital AI assistants from any photo. Boston Dynamics announced their latest, all electric, Atlas robot. You know I’m excited for robotics this year 🤖! Dove also released an amazing short film about beauty bias in generative AI. Check it out.
Meta introduces Llama 3: The most capable openly available LLM to date
This week Meta introduced Llama 3, which was broadly hinted at during Meta’s recent AI day in London. This announcement is big news and really significant for two reasons:
Firstly, the open source community now has access to a GPT-4 class model. This is going enable much more capable research and start-up product development. Jim Fan, of Nvidia fame, has a good comparison of Llama-3-400B’s capabilities that really brings this to life.
Secondly, Meta are embedding Llama 3 everywhere. As The Verge reports, Meta AI (powered by Llama 3) will be put everywhere across Instagram, WhatsApp, and Facebook. Almost overnight this will put Llama 3 in the hands and pockets of billions of users.
Mark Zuckerberg wants Meta AI to be “the most intelligent AI assistant that people can freely use across the world” and with Llama 3 Meta feels like they are now at that point.
Microsoft announces VASA-1 which generates talking faces from images
This is pretty impressive and pretty scary at the same time. Give the model a still image and an audio clip and it will create a video of the person in the image saying everything in the audio clip 🤯. The quality of these videos is so realistic that unless you’re looking for it you wouldn’t notice that it’s artificially generated.
VASA-1 can create the video in close to ‘real-time’ with a latency of only 170ms. This means that this approach could be used to turn a large language model like ChatGPT into a photorealistic digital avatar. So, later this year when you get your first digital AI assistant you can give it a face and a voice. Anyone’s face and anyone’s voice as long as you have a photo of them and a 15 second sample of their voice.
Things are going to get weird!
Boston Dynamic’s Atlas humanoid robot goes electric
First we got a video saying farewell to Atlas, and then the next day we got a glimpse of the new Atlas, which has removed all the hydraulics and is now built around fully electric motors.
Techcrunch has some great coverage of this news, including an interview with Boston Dynamic’s CEO. A core part of the philosophy behind the new Atlas model is not to replicate human motion, but to improve upon it, which is exactly what you see in the video.
Can’t wait to see what new Atlas has in store for us as Boston Dynamics shows us more later this year!
Dove releases short film challenging bias in generative AI
“One of the biggest threats to the representation of real beauty is Artificial Intelligence. With 90% of the content online expected to be AI-generated by 2025, the rise of AI is a threat to women's wellbeing, with 1 in 3 women feel pressure to alter their appearance because of what they see online, even when they know it's fake or AI generated.“
Enough said.
AI Ethics News
How cheap, outsourced labour in Africa is shaping AI English
India's election overshadowed by the rise of online misinformation
DARPA Tests AI Dogfighting (But Won't Say If the Human Pilot or Computer Won)
Adobe’s ‘Ethical’ Firefly AI Was Trained on Midjourney Images
Hugging Face releases a benchmark for testing generative AI on health tasks
Long Reads
Stanford University - AI Index Report 2024
“The future is already here, it’s just not evenly distributed.“
William Gibson