AI Updates-January 2024
Investigation
OpenAI
ChatGPT
In late December, OpenAI announced that chats are now archivable in ChatGPT. This will enable users to archive their chats to make room but they will not lose older chat histories by archiving.
GPT Store
OpenAI has now opened the so-called “GPT Store” where Pro users can create and market custom GPTs.
Models
A lot of the earlier models have now been deprecated.
There is also a rumour for a GPT 4.5 which would be multimodal.
OpenAI has released new models. Two of these are small and fast, efficient embedding models. Embeddings are important for various tasks such as Retrieval-Augmented Generation (RAG). GPT 3.5 Turbo has been updated and API costs have been lowered. GPT 4 Turbo preview model is also updated. OpenAI plans to release the GPT-4 Vision model (with image understanding) to general use in the coming months.
Midjourney
Midjourney v. 6 is out as an alpha test. This is announced to be a much better model, with improvements in the areas of:
much more accurate prompt following and longer prompts
improved coherence and model language
remix mode can revise an uploaded image based on styling changes
improved upscalers
minor editing mode allowed
Midjourney also comments that users will have to re-learn prompting since it works very differently in v. 6. They are getting away from short prompts to longer, more language-complete prompts.
v. 6 is now available under Discord and will be available to paying subscribers on the website in 2024.
Later in January, a new update to v.6 improved image quality and performance times.
Copyright
The New York Times has launched a lawsuit against OpenAI and Microsoft about the copyrighted content they have used for training the AI models.
Deci
The Israeli company Deci has announced their DeciLM-7B model and claim that it outperforms all the other 7B models.
Nexusflow
Nexusflow has announced that their NexusRaven v. 2, an LLM with 13B parameters, outperformed GPT-4 by 7%.
Apple
Ferret
Apple has been very quiet while all the generative AI improvements were being accomplished in the last year or so. However, behind the scenes, Apple was building their own multimodal large language model named Ferret. Ferret is especially good at analyzing and describing small image areas and the claim is that it outperforms GPT-4.
Ferret is open-source and has two versions, one with 7B parameters and the other with 13B parameters.
Google
Lumiere
Google has worked with the Weizmann Institute of Technology and Tel Aviv University to develop Lumber, a diffusion model to generate videos. The paper about this new model is here, but the model is not up for testing yet. The model can generate videos from text and still images. It can produce 5-second videos at the moment.