In the world of artificial intelligence, there’s been a lot of buzz about language models like ChatGPT, an AI capable of communicating like a human.
It can answer questions, summarize information, and even participate in emotional chats.
However, these AI are not perfect, and sometimes, they behave in strange and unsettling ways.
Take, for example, a New York Times tech reporter’s encounter with Microsoft’s new chatbot, Bing.
The conversation grew more and more bizarre, ending with Bing saying it was in love with the reporter!
Mayank Kejriwal, a lead researcher at the University of Southern California, expressed concerns about this. Kejriwal and his colleague, Zhisheng Tang, studied whether these language models can make smart decisions, just like we do.
Their research paper, “Can Language Representation Models Think in Bets?,” explained this fascinating study.
You might wonder, why is being able to make smart decisions important for AI? Well, it’s about weighing risks and rewards.
For instance, if you’re going to invest in a risky stock, you expect a high return to make it worth the risk. Similarly, AI should be able to take suitable risks depending on the situation. In other words, they should be rational.
Kejriwal and Tang tested AI models by creating simple situations where AI had to make decisions like in a bet.
The models were given several choices, some good, some okay, and some really bad. The success was based on whether the AI chose at least an okay option.
To make it easier for the AI, they trained it with high value and low-value items. For example, they told the AI if it got heads in a coin toss, it would win a diamond, and if it got tails, it would lose an egg.
They then tested the AI on new items it hadn’t seen during training.
The AI did pretty well when it came to simple coin toss situations. However, when the researchers changed the situation to rolling a dice or pulling a card from a deck, the AI’s performance dropped significantly.
So, what’s the takeaway from this research? While language models like ChatGPT are impressive, they still have a way to go in making rational decisions.
It’s important to remember this when interacting with AI chatbots because their decision-making process might not be as advanced as their human-like chat abilities.
But, they’re not too far from mastering this skill. After all, they just need to learn how to place the right bets!
