MPT-7B, a generative LLM made the sensation of the week on YouTube:
https://www.youtube.com/watch?v=O9Y_ZdsuKWQ
https://www.youtube.com/watch?v=8jIWOczSDAM
https://www.youtube.com/watch?v=EvM0A6d_KSA
Click on the videos to see their tests of this model.
Or test it yourself...
This model created some hype because of its large number of context tokens (65k-token, twice the amount of ChatGPT) and because of its licence (commercial use allowed).
If you want to test these models yourself, you can go to these pages:
https://huggingface.co/spaces/mosaicml/mpt-7b-chat (best)
https://huggingface.co/spaces/mosaicml/mpt-7b-instruct (usually faster, but the results of the other are better imo)
Both models were uncensored when i tested them. If you want to know more about this model, read this webpage of MosaicLM (profile on huggingface link):
https://www.mosaicml.com/blog/mpt-7b (link about MPT-7b)
My opinion is that these models don't beat chatgpt but that their results are better than the other do-it-at-home models.
More models with even better performances are coming:
This tweet claims that a concept that allows up to 2 million of tokens exist:
https://twitter.com/_akhaliq/status/1650308865555148800
With an important information: "This augmentation maintains the base model’s memory size at 3.6 GB in our experiments"
With even a linked github: https://github.com/booydar/t5-experiments
i didn't test this other model (yet?)
---
Stability AI released a model of generative AI that is able to handle text
You have probably seen these pictures with some gibberish text that spoils the fact that a picture was generated by an AI.
Well we are may be already after that...
Introduced in this blog post, Stability AI has just released DeepFloyd, a new text-to-image model that can now integrate text better in pictures.
You can try this model here:
https://creator.nightcafe.studio/deepfloyd-image-generator (you will need a login)
or here:
https://huggingface.co/spaces/DeepFloyd/IF
(here you get some example of test prompts)
The model is free but is not allowed for commercial use.
I have tested it, and it works so-so, with some mistakes and a small resolution (initially), but it is very promising...
I can guess that now the world is going to evolve at a very fast pace...
This project has a Github page
---
Next step: text-to-video
News website Axios reports that Runway AI decided to release a part of their toolkit to the public so "anyone can turn images, text or video clips into 15-second reels."
Click here to go to the website of Runwayml
Previously this year "Runway AI" had already organized a "AI film festival".
If you are interested to know who are the winners, click on this link below:
https://aiff.runwayml.com/
Runway also has a YouTube channel:
https://www.youtube.com/@RunwayML/videos
Test of the tool of Runwayml: https://www.youtube.com/watch?v=Wq8wU2iOJoo