Sora arrives ✨🎥

So much happening in GenAI this week, I thought I'd do a quick recap.

Memory for ChatGPT

OpenAI let slip that they are testing the release of 'memory' for ChatGPT. This feature will allow the AI to pick up details about you from different contexts you provide it, things like your living arrangements, if you have children, what your likes and dislikes are.

This sounds great in theory and might lead to much better responses for laypeople. However, I have some reservations about this. In much the same way that custom instructions can sometimes pollute your chat with unnecessary context, there is a danger that inappropriate context might be added to the chat without your direction. While this might make it easier for those with less knowledge to get better results, it might make it harder for those of us (AI Accelerator graduates!) to keep our context clean and achieve the better results we are accustomed to.

I'm going to be testing this in detail upon release and will let you know what I find.

Gemini 1.5 Incoming

So, Google decided this week would be a good time to tell everyone about the upcoming release of the new Gemini 1.5 model. We've been testing Gemini Ultra (Pro) for the last week, and the results are good; some of the outputs are more useful than ChatGPT+, however, it still struggles with some reasoning tasks and is unreliable for some basic reasoning tasks. Hopefully, this release will help to address these limitations and give GPT-4 some serious competition.

In this technical demo (although take it with a pinch of salt after the blue duck incident), Google showcased the use of automated transcription and video understanding using their claimed new 1 million token context window. The results are impressive, and hopefully, this will make its way to the Gemini Pro account soon so we can test it out.

You can watch the demo here.

Sora

Last, but by no means least, OpenAI unleashed Sora on the world. This text-to-video model has blown minds around the world with its insanely good video quality. However, it isn't just the videos that have captured the attention of the AI community. In a research post, OpenAI explains how they have found emergent abilities of world-building simulations, which, if true, might well be a pivotal point in the evolution of AI.

In this video, I explain why this is such a big deal.