- Modern Chaos
- Posts
- MC.12: Kyutai's AI Vision, From Screenshot to Syntax, and DeepMind's Music AI Experiments
MC.12: Kyutai's AI Vision, From Screenshot to Syntax, and DeepMind's Music AI Experiments
This week, discover the new research lab Kyutai backed by €300 million, grasp the magic of turning screenshots into code, and experience DeepMind's innovative fusion of AI with the art of music composition.
Samuel ROY
November 23, 2023
modern chaos issue 12
Modern chaos is a newsletter exploring tech and AI through the journey of a dev agency shifting from services to product design. We share our notes, analysis and experiments.
Hello everyone!
Like many, I've been caught in the recent drama at OpenAI. It's sparked much speculation. Is it ego, an AGI breakthrough, or something else? The truth is unclear. This reminds us of Cloud services' wisdom: don't rely on one provider. This is true for applications and Language Learning Models (LLMs) alike. Diversifying sources and providers is crucial.
Cheers,
In this issue:
📺 YouTube: Our Approach to Responsible AI Innovation
🌟 Awesome-GPTs: Curated List of GPTs Built on OpenAI
🖼️ Screenshot-to-Code: Convert a Screenshot to HTML/Tailwind/JS Code
📊 Hallucination Leaderboard: LLM Performance in Summarizing
🎶 Google DeepMind: Transforming the Future of Music Creation
🤖 Kyutai: A New AI Research Lab with a €330 Million Budget
🧠 Orca 2: Teaching Small Language Models How to Reason
💻 DeepSeek Coder: A New Open-Source Code LLM in English & Chinese
Tech updates & tools
Youtube: Our Approach to Responsible AI Innovation
Google will introduce new content labels to inform viewers when the content they’re viewing is synthetic. Creators will be required to disclose any use of AI tools, even if it justs partially alteration. The labels will be prominently displayed on sensitive contents, such as elections. In my opinion, this is a step in the right direction. MORE
Awesome-GPTs: Curated List of GPTs Built on OpenAI
It’s great to see the work of other AI enthusiasts and have the ability to try it out immediately. Plus, if you like one of them, it’s easy to retrieve the prompt and create your own personal version. MORE
Screenshot-to-Code: Convert a Screenshot to HTML/Tailwind/JS Code
A simple app using GPT-4 Vision to generate code and images from a given screenshot. I’m curious to see how paid services like v0.dev will survive when so many open-source alternatives are emerging. VIDEO | MORE
Hallucination Learboard: LLM Performance in Summarizing
This is a public LLM leaderboard computed using Vectara's Hallucination Evaluation Model. It evaluates how often an LLM introduces hallucinations when summarizing a document. GPT-4 is at 3%, and Mistral 7B at 9.4%. MORE
Google Deepmind: Transforming the Future of Music Creation
Have a sneal peek at the AI-related music experiments taking place in Youtube. They showcase Dream Track for Shorts, generative music mimicking various artists, and Music AI tools where you just have to sing to turn your idea into music. Impressive. MORE
Kyutai: A New Ai Research Lab with a €330 Million Budget
I talked last week about how France is betting on AI. This news is another strong signal. This lab is aiming to democratize AI and will open-source and disclose everything, including the data used. MORE
Orca 2: Teaching Small Language Models How to Reason
Orca 2 significantly surpasses models of similar size and attains performance levels better than models 5-10 times larger, as assessed on complex tasks that test advanced reasoning abilities in zero-shot settings. Model 7B & 13B available. MORE | HUGGINGFACE
DeepSeek Coder: A New Open Source Code LLM in English & Chinese
Another state-of-the-art code model to try. It’s available on Ollama and features various model size from 1.3B to 33B. MORE
Reply