What’s happening with AI? Researcher explains why you can look forward to more creative NPCs, competition for ChatGPT and hot dog tomatoes

After the tiny Christmas break in December, January brings even more new things, especially around language models. (Source: stock.adobe.com – Narumol)

Progress is exploding in the world of artificial intelligence. I can hardly keep up with reading publications; there are so many curious, practical and fascinating things.

That’s why I’m offering you a brief overview again and trying to show a mix of the most important, bizarre and interesting innovations in the field of AI.

Language models as NPCs

What happened: Last April we reported on a virtual village that was simulated using a language model. Similarly, there is now an approach that brings language models into open-world games.

A ChatGPT-like language model makes the decisions of an NPC. A range of environmental information is included, such as where the NPC is and whether he is brave or fearful. The neural network in the background makes a decision based on this, which the language model then describes.

Overview of the components of the AI ​​NPC. (Source: Ming Yan et al. in “LARP: Language-Agent Role Play for Open-World Games”)

Why is that important? Have you ever been told by the fifth guard in a row in Skyrim that an arrow in his knee ended his adventuring career? This is exactly what can be avoided in the future, making the game worlds appear more lively.

It also relieves game designers of their work, as they don’t have to plan hundreds of possible interactions that the player may never see in the end. Instead, a language model simply takes over the interactions. Developers then just have to be careful that the language models don’t get into any mischief.

Ultimately, this post shows again that language models exhibit very general capabilities that can be used for everything from translations and Christmas cards to better NPCs in video games. As long as the developers are careful that the language models don’t get up to all sorts of mischief.

Samsung presents new smartphones – and relies entirely on AI

What happened: Samsung not only presented the new S24 Ultra in its live presentation, but also relies entirely on AI. Among other things, Google is on board as a partner. Users can now, among other things, simply mark things on their screen to let Google’s AI tell them more about it.

There is also a slow-motion mode for videos, which automatically upscales recorded videos to more frames per second so that you get smooth slow-motion recordings. In addition, Samsung also shows real-time translation when making calls.

Why is that important? It’s hard to miss how manufacturers are placing increasing focus on AI features and evolving them from gimmicks to practical tools. Things that were purely research projects a few years ago are finding their way into finished products.

With the translator shown by Samsung, your first science fiction dreams can also be fulfilled. By combining it with existing voice imitation technology, a real-time translator could be built that comes very close to the language barrier-breaking Babel fish from “The Hitchhiker’s Guide to the Galaxy.”

This is how you try it: The new AI features are initially reserved for users of a few selected Android phones, but I would be surprised if Google and Apple didn’t add corresponding features to other phones very soon.

Talking over each other makes language models smarter

What happened: Someone asks a question, and several people in the room take turns answering it – each with one word. The answer is better than any one person could provide. While this sounds strenuous and unlikely for humans, it works extremely well for language models like ChatGPT.

With alternatingly generated snippets of words, researchers from England were even able to beat ChatGPT with three smaller language models. By the way, small here means around 14 billion parameters, downright tiny compared to ChatGPT’s 175 billion parameters or Google’s 1.6 trillion-parameter Gemini.

A better result was achieved with less than a tenth of the arithmetic operations.

Why is that important? Not only are the combined models massively smaller than the current top dogs, they are also significantly faster, because only one of the models has to be used per scrap of words. Instead of 175 billion arithmetic operations for the next scrap of words as with ChatGPT, you don’t even need “only” 14 billion, but only around 5 billion parameters. At the latest when you want AI like Google’s Gemini to run on your cell phone, your battery will thank you for it.

This is how you try it: You need a group of people, ask them a question, and then have them take turns… All joking aside, to try something like this you’ll need a bit more practical knowledge and you’d have to have several such language models lying around yourself. But I am sure that such technology will soon find its way into one of the larger models.

52:00

The game graphics of the future will be breathtaking – thanks to AI

Did you forget what you discussed with your boss right after the Zoom meeting? Is your mechanical keyboard driving gaming partners crazy in voice chat? Would you like automatic subtitles for the video of your presentation?

Krisp could be of help to exactly these problems. The practical tool simply closes between your input and output and therefore works for basically any software that runs over audio channels, such as Zoom and Discord.

If desired, it can be used to record and summarize conversations, suppress noise both in your background and from your conversation partner, and much more.

And the best thing about the whole thing: You can use it quite extensively in the free version, even if additional functions such as combining more than two meetings per day are only possible for paying users.

Kurznews

Not a genetically modified chicken or a hot dog-tomato hybrid, just a mistake in the AI ​​model: Generated recipe images on Instacart. (Source: Instacart/Business Insider)

Aside from these larger topics, there are numerous smaller news items:

• The food delivery service Instacart recently started suggesting AI-generated recipes, including images, to its users. However, there are all sorts of misfires, such as chickens with two bodies, as Business Insider reports. From the painful experience of my research, I can confirm that recipes and pictures for cocktails cannot be presented without errors.
• After users complained that various GPT variants such as ChatGPT or GPT-4 had become increasingly lazier when creating lines of code and simply would not complete a task, OpenAI now wants to tackle the problem.
• Editing images should become even better in the future: Current technology first turns a part of the image into a 3D object in order to then be able to control it. This allows you to move a person’s arm in a photo, for example, without first deleting the arm and then regenerating it, as before, and losing details in the process.
• Google introduces its new video model that generates video from text descriptions. Not only does this model deliver sharper videos, it can also animate given images, such as paintings or comics. However, it is not freely available.

Did you miss our review for last month? Then you should catch up on this:

Our AI review for December: Faster images, nicer editing for everyone

What AI news did you see in January? Will Samsung’s announced AI features influence your choice of your next smartphone, or do you perhaps even find something like this rather scary? Have you ever tried an AI-generated recipe? Tell us your opinions and experiences gladly in the comments!