Beware of smart people: they seem to know the answer to every question, but they can tell the biggest lies convincingly. This is what the Future News Portal found To its editors About the latest research findings related to artificial intelligence (AI).
Each new version of AI gets smarter, but research has shown that it is becoming more and more unreliable at the same time. More recent iterations have taken on the most annoying characteristic of educated people: they would rather hide some sort of answer to the questions put to them, than admit that they don't know the answer.
It was published in the journal Nature He studies Its authors investigated the large language models of OpenAI GPT, Meta LLaMA, and BigScience open source BLOOM. On the one hand, they found that in most cases they gave more accurate answers than their previous versions, and on the other hand, the rate of incorrect answers increased.
Until now, intelligence has not been far behind AI, but it has gotten worse, according to Mike Hicks, a professor at the University of Glasgow. A specialist in philosophical issues of science and technology made a stern statement: “They seem to be doing what we call nonsense in everyday life. They are pretending to know the answer to something they do not know.
According to the study, the largest deaf people are OpenAi's GTP-4 and o1 programs, which “know the answer” to almost every question they receive. However, this is the case in all AI models. For example, none of the LLaMA programs achieved 60 percent accuracy even for the simplest questions.
Image: Deposit images
AI models have a stubborn weakness that goes beyond just bullshit. Namely, their answers to the simplest questions can be inaccurate. For example, in the case of chess puzzles, it regularly happens that the task is “solved” with irregular moves.
Bullshit propaganda hides the deafness of artificial intelligence
Worryingly, the majority of users are so impressed by the AI's knowledge that they tend to forget its mistakes. A survey revealed that people estimate the rate of wrong answers in AI models to range from 10% to 40%, much lower than the real answer. Experts believe that the solution may be to change the adaptation of chatbots.
They simply need to be trained not to “want” to answer the questions they receive at any cost. Don't “feel the problem” if they have to say sorry I don't know the answer to that question. The problem is that this is not in the interest of companies developing artificial intelligence. If their products admitted they could only “answer” half the nonsense questions they were asked, the hype around AI would decline. This may turn out to be not the wonderfully mysterious thing presented in nonsense trailers.
