Scientists are researching how and why AI systems are developing emergent skills.
As AI models become more sophisticated, they begin to perform tasks for which they were not designed. This caught the attention of specialists after an AI system managed to determine the name of a movie described with emojis.
Google Research Computer Scientist, Ethan Dyer, said that the emerging skills are surprising given that such models supposedly have a set directive. Meaning, they take a text string as input and predict what comes next based on statistics alone.
Recent research has revealed that large language models (LLMs) exhibit hundreds of emerging skills. Tasks ranging from multiplication to computer code production and beyond.
Theories on emerging AI skills
Scientists are racing to identify more emerging skills to discover why and how they emerge. Understanding this can solve deep problems about AI and machine learning in general. Additionally, help to know whether complex models are in fact doing something new or getting good at statistics. At the same time, it would aid researchers explore its potential benefits and minimize major risks.
Some recent theories suggest two possibilities for why new AI skills emerge. The first points out that language models function like biological systems. This means that more complex models acquire new skills spontaneously. “It’s quite possible that the model has learned something new and different that it didn’t have when it was younger,” Pavlick states.
The other theory suggests that what appears to be emergent is, instead, the culmination of an internal statistically driven process; it works through chain-of-thought reasoning. Larger LLMs could be learning heuristics, which would be beyond the reach of lower-quality models.
Lastly, the Universidad Internacional Iberoamericana (International Iberoamerican University) (UNIB) offers the Master in Strategic Management with a Specialty in Information Technology.
Source: Los sistemas de IA desarrollan habilidades impredecibles y los científicos no tienen explicación
