AI Weekly Digest #16: API based open source LLM alternative to OpenAI is out!

Hello, tech enthusiasts! This is Wassim Jouini and Welcome to my AI newsletter, where I bring you the latest advancements in Artificial Intelligence without the unnecessary hype. Let's dive into this week's news and explore the practical applications of AI across various sectors.

• The Main Headlines:

  • Midjourney 5.2 releases new groundbreaking features,

  • Stable diffusion announced an update to its AI image generator SDXL, the largest open source Image generator,

  • Automatic Image Segmentation can now be done in real time thanks to Fast Segment Anything

  • Grammys won’t accept AI generated content in the competition?

• Beyond the Hype: MPT30B released for commercial use offering the first API based open source LLM alternative to OpenAI!

• Bonus: Access SDXL online, and test uncrop, an open source equivalent to photoshop’s generative fill!

Beyond The Hype

MPT-30B is released for commercial use! How does it compare to other models and is it cheaper than OpenAI?

Blog Post Annoucing the release: https://www.mosaicml.com/blog/mpt-30b

Why is it a big deal?

• Proprietary vs. Open Source: OpenAI is dominating the market with very few competitors, such as Google’s PaLM, which is in public preview, and Bard, which is still in beta.

• Data & Model control: The release of MPT30 as both an infrastructure and an API is enticing for businesses seeking full control over their data and models; it offers a streamlined process for training and utilizing their models online.

• Model Settings: With 30 billion parameters, this model is capable of running on a single A100 GPU. It offers an 8k-token context window, which is the largest compared to open-source models, and even surpasses the 4k-token limit of GPT-3.5-turbo-0613. For comparison, Falcon40B is currently limited to a 2k-token context window.

• Performance: the blog post release states that it is comparable to GPT-3 davinci-003 (which is very decent and can be used for various tasks such as categorization, data extraction, sentiment analysis and so on). It also states that it is competitive with other open-source models such as LLaMa-30B and Falcon-40B.
  However: it doesn’t say much on how well it performs on different languages. My assumption is that its performance would drop very quickly on languages other than English.

What does it mean for the market?

• Pricing: it’s hard to have a fair price comparaison since Mosaic Inference packages are based on a cost per “GPU hour” and OpenAI is based on the number of tokens used. Yet the blog post states that a same operation (e.g., serving a fined-tuned model via API) would be x15 cheaper through their infra compared to OpenAI’s GPT-3 APIs.

• Worth it? if you are relying on a davinci fine-tuned models, obviously this would offer some significant savings! Nevertheless, GPT3.5 turbo is 20 times cheaper than davinci models and performs much better without fine-tuning ; it can be worth comparing both approaches first. Note that GPT3.5 turbo has a 16k version as well, which can further extend its capabilities.

Conclusion: This is a significant development primarily because the open source community is rapidly gaining momentum in competing head-to-head with OpenAI. MPT30B is merely the initial step. We can expect to see more advanced, efficient, and cost-effective models emerging soon, positioning them as viable alternatives to OpenAI models.

Main Headlines

#1 Midjourney 5.2 is out bringing new features!

Mid-Journey announced version 5.2 with new features including aesthetics, more variation in generations, and a new shortened command to identify non-impactful words in your prompt. The most interesting feature is the zoom out option, which allows you to zoom out on any image

• Zoom Out: an image is worth a thousand words! This features allows you to zoom out of an image → the AI regenerates the image while extrapolating the rest of the scene! Check these results.

• Shorten : the /shorten command (same as /imagine to generate an image) takes a prompt as an input and return a shorter version. It helps you get insights on what is actually used o generate your images!

#2 Stable Diffusion announced the release of SDXL 0.9

SDXL 0.9 is an upgrade to the Open Source Stable Diffusion text-to-image model, and it outshines its predecessor with significantly improved image and composition detail! (you can test it, see link in the bonus below).

#3 Automatic Image Segmentation can now be done in real time thanks to Fast Segment Anything

Meta released “Segment Anything Model” (SAM) in April 2023, a model capable of segmenting any image automatically. It offered the ability to segment an image based on a prompt as well (e.g., find dog in the image would return the position and mask for the available dog in the image!).

The challenge? making it blazing fast to allow for real time applications.

This is what FastSAM is for with a x50 speed improvement!

#4 Grammys won’t accept AI generated content in the competition?

The Recording Academy, the body responsible for the prestigious Grammy Awards, has recently updated its rules concerning eligibility for nominations. According to the new regulations, only human creators are eligible for the music awards.

In a nutshell, Grammys won't nominate AI generated content unless there is "proof that a real person meaningfully contributed to the song too."

Should we add an AI category? what do you think?

Bonus!

Stable Diffusion SDXL features can be tested for free online.

The model will soon be available as open source and via APIs, meanwhile, you can test it here! Have Fun!

That’s it for today! If you made it this far, I’d appreciate a quick feedback 😋! I know there’s room for improvement! So don’t hesitate to share with me the things you liked and those that you didn’t.

Have a great Sunday and may AI always be on your side!