An autonomous artificial intelligence? This ‘scientific AI’ is capable of creating and carrying out its own experiments

A team from the University of British Columbia (UBC) in Vancouver has developed a ‘scientific AI’ capable of inventing and conducting its own experiments, an advance that could completely revolutionize artificial intelligence research. Although the initial results of this AI are not particularly innovative, researchers consider it an important first step towards an AI that could learn and create autonomously.

ChatGPT DOWNLOAD

Professor Jeff Clune, leader of the project at UBC, along with collaborators from the University of Oxford and the startup Sakana AI, acknowledges that the ideas generated by this AI are not yet ‘revolutionary.’ However, he emphasizes that this approach could unlock capabilities that surpass what humans have taught AI systems so far. ‘They are not tremendously creative, but they seem like pretty cool ideas that someone could try,’ says Clune.

The key to the project lies in the AI’s ability to learn in an open-ended manner, that is, by exploring ideas and experiments without the need for training data generated exclusively by humans. This approach is based on the use of large language models (LLM), which allow the AI to identify and explore what it finds most interesting. Previously, AI programs needed hand-coded instructions to define what was interesting; now, LLMs allow for greater autonomy.

Despite the criticism, such as that of Professor Tom Hope from the Hebrew University of Jerusalem, who questions the reliability and originality of current results, the UBC team is moving forward. Their latest breakthrough is an AI program capable of designing other AI agents, which already outperform humans in some tasks.

Clune acknowledges the potential risks of this technology and emphasizes the importance of developing secure systems. ‘It is potentially dangerous,’ he warns, ‘but I believe it is possible to do it right.’

Source: Wired

ChatGPT DOWNLOAD