AI Weekly Digest #8: DeepFake as a Service?

Welcome to this new edition of our AI newsletter, where we bring you the latest updates on artificial intelligence!

• The Main Headlines: Unveiling DeepFloyd IF's breakthrough, empathetic chatbots replacing physicians, create your digital twin with DeepFake as a service, ChatGPT in robots, and HuggingChat—an open-source alternative!

• Beyond the Hype, the story: Can Open Source Offer a Viable Alternative to OpenAI APIs?

• Bonus: 10min Tutorial to test DeepFloyd IF model!

Beyond the Hype, The Story

The Rise of Open Source Alternatives to OpenAI APIs

Open source alternatives to OpenAI APIs have been gaining momentum, with new, cheap and, open models showing remarkable improvements in various domains e.g., NLP, Image Generation, Video Generation, Coding, Multimodal models, etc.

Although OpenAI APIs are still dominating the market, with no language model capable of replacing ChatGPT APIs or GPT4 API, these advancements showcase the potential of open source solutions to compete with or even surpass proprietary models.

Here’s a quick overview of the month or so!

• Large Language Models: Models like Vicuna and HuggingChat demonstrate the capabilities of smaller models (13B-30B parameters) to compete with giants like GPT-3 (175B parameters).

  These open source models are continuously being enhanced with better datasets and more efficient training methods. (It all started with Alpaca model!)

• Image Generation: DeepFloyd IF represents a new generation of image generation models comparable to OpenAI's Dalle2. Notably, it addresses the challenge of incorporating text synthesis within images, setting a new standard for image generation models.

• Video Generation: Nvidia is emerging as leading the open-source charge in video generation with VideoLDM! Building on top of stable diffusion models and adding a temporal dimension to it. And the results are quite impressive!

• Multimodal Models: The innovative MiniGPT4 combines Vicuna and Blip2 to create a powerful multimodal model reminiscent of GPT-4. It's capable of tasks such as image description, code generation based on drawings, and meme explanations, demonstrating the rapid progress of open source multimodal models. (read more on how MiniGPT4 was created as the result of 4 weeks of incremental innovation!).

  

• Coding Assistant: While OpenAI has dominated the coding assistant market with tools like Codex, GitHub Copilot, and ChatGPT, Replit has recently introduced their own Large Language Model for code: replit-code-v1-3b.

  With 2.7 billion parameters (which is rather small!) and support for 20 languages, it outperforms all open source code models on the HumanEval benchmark and was trained in just 10 days! Once again, showcasing that it is becoming simpler and cheaper to compete with OpenAI’s expensive models!

Sam Altman stated recently that it took more than $100M to build GPT4 ; can cheap Open Source initiative really compete?? One thing is clear, they are trying very hard to, and closing the gap very (very!) quickly!

Stay Tuned!

Main Headlines

• Stability AI releases “DeepFloyd IF” and it cracks one of the remaining image generation challenges: generate an image with text based on a prompt! E.g.,

  Prompt: a photo of a full size old rusty sign that says "Deep Floyd Street", photo realism, bokeh, 50mm cine lens, super sharp focus.

  Result: The generated image does include the text indicated in the prompt!

  

• Patient Diagnostics: Chatbots proved to be better and more emphatic than humans. An image is worth a thousand words, heres’ what you need to know illustrated in the image below!

  A study compared responses to patient questions on a public forum by physicians and an AI chatbot, and found that patients preferred the chatbot's responses 79% of the time due to higher quality and empathy.

• DeepFake as a service: create your digital twin for 149$ and in 24h!

  

• ChatGPT in Robots? Boston Dynamic showcases a new integration of ChatGPT in their robots.

• HuggingChat: open source chatgpt alternative (based on OpenAssistant 30B parameter model) hosted on HuggingFace. You can test it here: https://huggingface.co/chat/

Bonus!

You want to test Stability AI’s new DeepFlod IF model? Here’s a 10min tutorial to get you started!

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!