ChatGPT & Co. are less adaptive than previously assumed
AI models are less capable of learning on their own than previously thought, according to a study by the Technical University of Darmstadt. The study also found no evidence that large language models develop «intelligent» behavior or the ability to think complexly.
AI models such as ChatGPT, Gemini and Co. are less capable of learning independently than previously assumed. This is the conclusion of a study led by the Technical University (TU) of Darmstadt. As the TU writes, no evidence was found that large language models (LLMs) were beginning to develop general "intelligent" behavior
The research focuses on unforeseen and sudden leaps in the performance of language models, so-called "emergent capabilities". The models studied became more powerful with increasing size and the amount of training data, as the TU writes. The models can solve more language-based tasks such as recognizing fake news or drawing logical conclusions as they scale up.
This circumstance raised hopes and concerns, it is said. On the one hand, it was hoped that further scaling would make the models even better; on the other hand, there was concern that these capabilities could become dangerous. This led to the introduction of AI laws around the world.
LLMs still a long way from human skills
However, the study states that there is no evidence for this presumed development of differentiated thinking in the models. The LLMs can follow "relatively simple instructions", but they are still a long way from the skills of humans.
"However, our results do not mean that AI is not a threat at all," emphasizes Iryna Gurevych, professor at TU Darmstadt and co-leader of the study. "Rather, we show that the alleged emergence of complex thinking abilities associated with certain threats is not supported by evidence and that we can control the learning process of LLMs well after all. Therefore, future research should focus on other risks posed by the models, such as their potential to be used to generate fake news."
It is probably a mistake to rely on an AI model to interpret and execute complex tasks without help, says Gurevych. Users should explicitly state what the systems should do and, if possible, give examples. Confabulation, the tendency of these models to produce plausible-sounding but incorrect results, is likely to persist. (Alexandra Hüsler/jor)
