AI Blog How many announcements can you fit into a week?

How many announcements can you fit into a week?

AI Blog

Issue #32

08/11/2023

Our own announcements!

Yesterday our friendly SleedAI bot received some very important updates. It now has GPT-4-Turbo with 128K context window (!!!) AND Dall-E 3! 🎉

We are excited and thrilled! We’ve been waiting for this moment for quite some time and we want to thank OpenAI for deciding to make announcements and finally make their new APIs available.

What does this mean in practice?

Much better results in terms of reasoning, creativity and use of Greek, which he can correct, just ask him to do proof reading.
The images produced by the Dall-E 3 have significantly better quality. For the sake of argument:

OpenAI dev-day announcements!

On 6.11.2023 the OpenAI Developer Day took place. You can and we recommend you to watch the keynote:

We’ve been told a lot of things, so let’s go through them one by one.

First, there was a launch of the GPT-4-Turbo as a preview. This model is the new & cheaper generation of the GPT-4 and has the world’s knowledge until April 2023.

Also, its context window is at 128K tokens (about 300 pages)! Practically this means that it “remembers” for much longer while you chat with it and doesn’t start saying invalid things after a few messages.

This is a really impressive number, it even surpasses Claude, which until recently had the largest context window at 100K.

GPT-4-Turbo also accepts images as input to its API. Therefore, it can be leveraged to create applications that do generate captions or moderate content.

As we said above, you can find it in the SleedAI bot, but it’s also available in the playground, which has changed its UI a bit as well.

On the top left, next to Playground, we see a toggle that says Chat. If we click on it it will display 4 options.

Complete and Edit are legacy, meaning they will soon be no longer available, so you can ignore them. We are interested in the Assistants option. This is a new feature offered by OpenAI. Basically they are bots (agents, the official name) that you can integrate into your own apps. Inside the playground, you are given the option to build and test them. We played a bit with them yesterday and made an HR AI Assistant that has knowledge about our internal processes. It needs a bit of coaxing on prompts to respond, since we didn’t do the necessary data formatting on the file we uploaded, but it gives very satisfying results.

You can also create your own Assistant by selecting the corresponding tab from the side menu.

Click +Create will display the corresponding side menu.

Fill in:

Name
Instructions, i.e. the system prompt to follow each time
Model
Extra tools, e.g. files from which you want it to derive its answers or functions or Code Interpreter
- Code Interpreter: writes and executes Python, can create graphs and charts, and manipulate files with a variety of data and formatting.

Important reminder

OpenAI does not utilize API and Playground requests in the training of its models.

Assistants are a HUGE breakthrough. They are a no-code option for making AI agents and are directly related to the next announcement, custom GPTs. You are given the option to make a custom ChatGPT, which will combine your own instructions, extra knowledge from files you gave it, and whatever skill the ChatGPT has even now.

Their creation is also no-code and you can share them with other users. In fact, later this month, OpenAI will open the GPT Store, where verified builders will be able to make their GPTs available.

And we’re not done with the announcements yet.

Dall-E 3 became available in API and we obviously ran to add it to the SleedAI bot. The difference with Dall-E 2 is huge.

Prompt used: white fur monster in a kitchen, night time, digital art

We have new text-to-speech APIs. We couldn’t find out if Greek is supported as a language, but our guess is probably not, as it sounded a bit strange in our test.

OpenAI also joins the list of companies offering Copyright Shield to their customers. That is, it agrees to pay the costs incurred if someone faces legal claims about copyright infringement and if that someone has ChatGPT Enterprise or uses its developer platform.

In conclusion, what we have to say is that OpenAI is trying to be an AI provider for all kinds of services – and for now it is succeeding. In less than 1 year, it has 100 million weekly users and 2 million developers leveraging its APIs. “Build for the future” is something we say/hear often in the AI world. And after the day before yesterday this “building” becomes extremely easier and more accessible.

PS. There was such a fuss that yesterday OpenAI had an outage on all their products. Probably because of the massive requests that were made.

Resources

AI Insider 📰

New AI campaign from Socialab for Kotsovolos. It consists of 5 static posts with cute monsters depicting different categories of devices.
We have an important announcement from Midjourney, who released a style tuner feature. It allows you to personalize the images you produce. In your prompt you add the parameter –style + the code of your own style.
Meta does not allow the use of its AI tools in Ads Manager on ads that are related to politics, elections, social issues, employment, health, financial services or pharmaceuticals.
Elon Musk announced Grok, the AI chatbot that he told us would “understand the universe”. If this objective reminds you of anything, it’s the same as most AI in sci-fi books/movies/series. Grok will be available to premium X subscribers and has real-time access to content published on X.
Runway released a new upgrade to Gen-2 (their AI video maker) significantly upgrading their outputs. A simple and quick test that we ran:

Also, Runway seems to have built its own AI video editing machine.
Performance Max gains AI capabilities for creating creative assets. It can suggest and produce headlines, descriptions, and images, as well as edit existing images.

Learning Bytes 🧐

Elon Musk made some bold statements to the UK Prime Minister during the AI summit, such as that in the future because of AI no one will have to work.
The key takeaways of the aforementioned summit:
- New AI models must be tested before they are released.
- The Bletchley Park conference was internationally attended and an international declaration on the risks of AI was signed.
- There was clear US influence.
- Elon Musk’s presence at the conference brought additional publicity and highlighted the need to address the threats of AI.
- There were different views on the risks of AI, but all agreed that addressing immediate threats, such as disinformation, is a priority.
- Countries are moving at different paces in regulating AI, with some advocating the need for immediate action and others for a more gradual approach.
Handy guide on how to leverage AI in Google Sheets.