The recent spectacular advances in the field of artificial intelligence, in particular with generative models such as GPT-4 or DALL-E, are raising a lot of enthusiasm but also questions about the future of this technology. We took a step back from the hustle and bustle of the moment to think in depth about the different scenarios for the evolution of AI in the coming years and decades.
We offer you a chapter from a course on AI safety that we developed. This prospective document seeks to provide an overview of recent advances in AI and to explore some potential trajectories towards the development of highly efficient AI systems.
In summary:
Although relatively technical in places, this document is intended to be accessible. Its objective is to provide factual and prospective insights on a subject that raises a lot of questions and fantasies. This chapter provides the basis for solid and well-reasoned thinking on the perspectives of generative AI.
While there are still many uncertainties about the future progress of AI, there does not seem to be an insurmountable theoretical obstacle to the development of ever more powerful and autonomous AIs. Some methodologies presented in this chapter even suggest that it is possible to see the emergence by the end of the decade of AIs capable of matching, or even surpassing, the best human performances in a very wide range of cognitive tasks.
Of course, these projections should be taken with care. Numerous factors, technical, but also economic, political and societal, could influence the pace and modalities of progress in AI. However, given the current dynamics, the hypothesis of “human-level” AI within 10 years cannot be ruled out.
This observation highlights the importance and urgency of continuing research on the alignment and safety of AI, in order to ensure that these powerful systems remain under human control and work for the common good. It is a major challenge that concerns us all.
