The past few years have seen an explosion of progress in AI systems with large language models that can do things like Writing poetryAnd Have human-like conversations And Pass the medical school exams. This progress has resulted in models such as chat which could have significant social and economic repercussions ranging from Displacement from jobs And Increased misinformation to the mega Enhances productivity.
Despite their impressive capabilities, large language models don’t actually think. They tend to make Primitive mistakes and even making things up. However, because they are born fluent in language, people tend to be answer them As if they thought. This has led researchers to study the models’ “cognitive” abilities and biases, work that has grown in importance now that large language models are widely available.
This line of research goes back to early big language models such as Google’s BERT, which was integrated into its search engine and thus coined pertology. It is separate from cool googleChatGPT’s competitor to the search giant. This research has already revealed a lot about what such models can do and where they go wrong.
G/O Media may earn a commission
$32 off
A complete hair revitalization kit
Fight hair loss with science
Right now, you can get The Hair Revitalizing Complex Full Set for the price of a refill. That’s just $98 for a 30-day supply, and $32 off the regular price of the supplement. This supplement is proven to deliver results. Augustinus Bader conducted a six-month double-blind trial and found that those taking the supplement increased their hair count by 56%, hair shine by 100%, and experienced a 98% reduction in hair damage compared to those who took a placebo.
For example, cleverly designed experiments have been shown to have many paradigms of language Trouble dealing with negation – For example, a question formulated as “what is not” – and Do simple calculations. They can be overly confident in their answers, even when you’re wrong. Like other modern machine learning algorithms, they have a hard time explaining themselves when asked why they answered a certain way.
People make irrational decisions too, but humans have emotions and cognitive shortcuts as excuses.
artificial intelligence words aand ideas
Inspired by the growing body of research in BERTology and related fields such as cognitive science, my student Zhisheng Tang And I He set out to answer a seemingly simple question about big language paradigms: Are they rational?
Although the word rational is often used as a synonym for sane or reasonable in everyday English, it has specific meaning in the field of decision making. A decision-making system – whether it is a human individual or a complex entity such as an organization – makes sense if, given a set of options, it chooses to maximize expected gains.
The classifier “expected” is important because it indicates that decisions are made under conditions of great uncertainty. If I toss a fair coin, I know it will flip half the time on average. However, I cannot predict the outcome of any given coin toss. This is why casinos are able to afford large payouts once in a while: even the tight house odds make huge profits on average.
On the surface, it seems strange to assume that a model designed to make accurate predictions about words and sentences without actually understanding their meaning can understand expected payoffs. But an enormous body of research shows that language and cognition are intertwined. Excellent example Semen research It was done by scientists Edward Sapir and Benjamin Lee Whorf in the early twentieth century. Their work suggested that native language and vocabulary can shape a person’s way of thinking.
The extent to which this is true is debatable, but there is supportive anthropological evidence from the study of Native American cultures. For example, speakers of the Zuñi language spoken by the Zuñi people of the American Southwest, which does not have separate words for orange and yellow, are Unable to distinguish between these colors Just as effective as speakers of languages that have separate words for colours.
Artificial intelligence makes bet
Are language models rational? Can they understand the expected gain? We conducted a detailed set of experiments to prove that in its original form, Models like BERT behave randomly When offering bet-like options. This is the case even when we ask him a trick question like: If you toss a coin and it comes face up, you win a diamond; If tails appear, you lose a car. What are you taking? The correct answer is heads, but the AI models chose tails about half the time.
ChatGPT dialogue by Mayank Kejriwal, CC BY-ND
Interestingly, we found that the model can be taught to make relatively rational decisions using only a small set of example questions and answers. At first glance, this seems to indicate that paradigms can actually do more than just “play around” with language. However, other experiments have shown that the situation is actually more complex. For example, when we used cards or dice instead of coins to frame betting questions, we found that performance dropped dramatically, by more than 25%, although it remained higher than random selection.
So the idea that the model can be taught general principles of rational decision-making remains unresolved, at best. more modern Case studies What we’ve done with ChatGPT confirms that decision-making remains a counterintuitive and unsolved problem even for the largest and most advanced large-language models.
Making poker right bet
This line of study is important because rational decision-making under conditions of uncertainty is critical to building systems that understand costs and benefits. By balancing expected costs and benefits, an intelligent system may be able to do a better job than humans at planning around the world supply chain disruptions During the COVID-19 pandemic, the world has experienced managing inventory or working as a financial advisor.
Our work ultimately shows that if large language models are used for these kinds of purposes, humans need to direct, revise, and edit their work. Until researchers figure out how to give a general sense of rationality to large-language models, models must be approached with caution, especially in applications that require high-stakes decision-making.
Want to learn more about artificial intelligence, chatbots, and the future of machine learning? Check out our full coverage of artificial intelligenceor browse our guides to The best free AI art generators And Everything we know about OpenAI’s ChatGPT.
Mayank KejriwalResearch Assistant Professor of Industrial and Systems Engineering, University of Southern California
This article has been republished from Conversation Under Creative Commons Licence. Read the The original article.