FOR IMMEDIATE RELEASE: 6/Aug/2022
Welcome back to The Memo.
Once again, the AI world is exploding, with the BlenderBot 3.0 chatbot demo released a few hours ago for those in the US, the new Amazon AlexaTM model, new DALL-E 2 examples, and much more…
The BIG Stuff
BlenderBot 3.0 175B chatbot demo by Meta AI (6/Aug/2022)
This is heavily geo-fenced to users in the United States only (even my VPN, location hacks, and proxy options wouldn’t get past it). The model is the same size as GPT-3, but fine-tuned on dialogue data.
BlenderBot 3.0 has long-term memory, and access to web search via Mojeek.
The live chatbot demo is free, and available in the US only: https://blenderbot.ai/
Amazon training their own Alexa model (3/Aug/2022)
Amazon have been quietly working on replicating GPT-3-style models, announcing AlexaTM 20B, based on Meta AI’s BART and using a multilingual dataset (Wiki + mC4).
AlexaTM 20B outperforms GPT3 (175B) on SuperGLUE and SQuADv2 datasets and provides SOTA performance on multilingual tasks…
Watch my video:
The Interesting Stuff
Lots of new models in the last 30 days (6/Aug/2022)
There were several models announced recently, including several major code models by Google AI and Meta AI. For my full timeline, see:
https://lifearchitect.ai/timeline/
WIRED article on Stentrode brain-computer interface (2/Aug/2022)
Oxley dreams of a million implants a year… “What I want the world to understand is that this technology is going to help people… the reality is that people need this technology, and they need it now.”
https://www.wired.com/story/synchron-brain-computer-interface/
Apple GAUDI (31/Jul/2022)
Turns out that Apple hasn’t been sleeping either! Their new text-to-3D scene generator is called GAUDI, and the applications will include AR/VR, gaming, immersive films, and robotics…
https://github.com/apple/ml-gaudi
Toys to Play With
Jeff Buckley’s Grace extended with OpenAI Jukebox (5/Aug/2022)
Including new lyrics and music across nine different trials (check the third trial at the 3:16 mark). One of the most incredible things I’ve heard this year! OpenAI Jukebox is an audio model released more than two years ago in Apr/2020—a few days before the language model, GPT-3—and is also based on the Transformer architecture.
Botif.ai chatbot based on GPT-J (Jul/2022)
Artem and the team recently upgraded the model behind Botify AI, to use GPT-J 6B. The major innovation in this app is the realistic avatars, based on a single photo.
Try it yourself for free: https://www.botif.ai/
CogVideo playground (1/Aug/2022)
Generating videos from text OR single images (photos) as the prompt!
I know there are several ‘family’ readers of The Memo, so please note that this includes some PG-rated/NSFW examples to emulate anime videos.