The latest AI news - too many to list

It’s been a while.

I’ve been reflecting on where I fit in this AI realm and how I can best contribute in a meaningful way as we keep moving forward at lightning speed. One of the ways is to reboot this newsletter with a slight course correction, including a name change to Unreal AI.

There is so much happening in AI every single day. It can be super overwhelming and hard to keep up - it has been for me too. I need clarity and simplicity - maybe you do too.

I will attempt to bring just that with this newsletter so it’s easier to keep up and get what you want and need to know about all the crazy, nonstop AI action. Quick to read, simply explained, easy to understand.

Thank you for hanging around during this quiet spell. If you feel this direction is not what you signed up for, I understand. Thanks for being part of this journey.

Having said all that, let’s get to it.

There is A LOT to cover.

The Latest

Midjourney (quietly) released a new Alpha website yesterday, that is, with image generation.

But. The alpha site is in its early testing phase (it can look very different when it’s shipped to all), and currently, it is only available for Midjourney users who have generated 10000 or more images.

P.S. The website image generation is NOT v6. Midjourney is really hoping to release v6 before Christmas! Keep an eye out for the community rating party (where we get to rate image pairs), which is the next step before the release of v6. Here’s today’s office hours recap.

OpenAI announced a partnership with Axel Springer, whose media brands include publications of the likes of Politico, Business Insider, BILD, and Welt. ChatGPT Plus users will soon have real-time information from the publications, and OpenAI will be able to use Axel Springer’s content in training its LLMs.

Initial reactions from commenters are mixed.

OpenAI has managed to find more GPUs, which means you can now again subscribe to ChatGPT Plus to access GPT-4, which, in my opinion, is still the best out there, even with issues. (at the moment, anyway - competition is intensifying)

Speaking of language models:

Microsoft announced Phi-2 - a small language model with common sense, language understanding, and logical reasoning. MS says it outperforms models up to 25x larger. It is currently only available in the Azure AI Studio model catalog for research purposes. Phi-2 is said to have been primarily trained on synthetic data (artificially created data).

Mistral, a French-built open-source AI model, Mixtral 8x7B, a high-quality sparse mixture of experts model (SMoE), is making waves.

It’s said to match or even surpass GPT-3.5 and Llama 2 with the ability to run it locally on devices without dedicated GPUs - which is the big deal IMO. Also notable is that it comes without guardrails for safety. Curious? You can try Mistral via Together.ai.

And here’s a Mistral demo on the phone by Nick Dobos.

One more: How to install Mistral/ChatGPT-type LLM locally on your Apple device.

Krea AI is now in open beta, available for all, and it’s free for everyone.

Kaiber’s video2video Transform 2.0 is now available for all.

Stability AI introduces Stable Zero123 Translation: a 3D object generation from single images. Available on Hugging Face for researchers and non-commercial users.

Google DeepMind announced Imagen 2 - their most advanced text-to-image generator. High quality, photorealistic outputs, etc. But it seems to be another meh announcement - since it’s only available to a very limited number of people in their VertexAI Google Cloud platform.

Meet Pi. Inflection AI released Pi, your personal AI. Similar to ChatGPT Voice, you can talk with it on your devices. Available for the public. I gave it a whirl yesterday, and although similar to GPT Voice but funner, at this point, I (subjective, I know) prefer GPT Voice (even with the plethora of issues as of late). You can choose one of six voices to chat with you. Maybe I’m too old haha (Gen X), as I found it to be more geared to the younger crowd.

You can now use Claude with Google Sheets - a seamless integration between the two. You will need Claude's AI key to install Claude for Sheets. If interested, click on the link for an easy integration tutorial.

Meanwhile…

Where did Optimus get its dance moves? From its father, of course.
Tesla Optimus & Elon dance moves from DogeDesigner.

Raise your hand if you have used the word ‘hallucinate’ this year more than ever. Dictonary.com named it the word of 2023. Let’s celebrate hallucinate.

Sports Illustrated fires CEO after controversy of AI-generated content, including AI-generated bios and authors.

Meanwhile, the New York Times hires an editorial director for AI to lead AI implementation and establish gen. AI guidelines, “how we do and do not use generative AI”.

Quick Explainer

What is RAG?

RAG, or Retrieval Augmented Generation, is an extra step between your input and the LLM model. In this step, RAG pulls external data in for context rather than relying on data that is in the LLM.

Why does it matter? RAG can improve the quality of generative AI responses. Think, injection of real-time information, leading to more accurate responses and fewer hallucinations in your queries, e.g. when using an AI chatbot.

Of course, it isn’t as simple as that. Here are two sources you can check out for more information about RAG. Snorkel's AI article and Oracle's article

Ok, I’m coming up for air.

You may see some adjustments to the format in the next couple of weeks, and I’m not sure yet whether this newsletter will be a daily thing or a dropping a few times a week.

One thing is for sure - we are not going to run out of AI news and action any time soon!!

brb,

Alie