AI artificial intelligence will reach human level: experts' predictions for 2025.

According to Vox: Two prominent experts in the field of artificial intelligence recently discussed the development of this area. One of them, François Chollet, the author of the popular Keras library and the developer of the ARC-AGI benchmark, assesses whether AI has reached 'general' intelligence. Chollet is known for his cautious approach to optimistic predictions, but during the conversation, he noted that the timelines for achieving significant results have shortened. He believes researchers have made notable progress in overcoming key barriers to artificial general intelligence.

The second interlocutor was Dvarkesh Patel, whose podcast has become an important source for tracking the thoughts of leading scientists in the field of AI. However, Patel has become more pessimistic about improving models of experience-based learning. 'People learn from their mistakes. They get small improvements and efficiencies during work,' he noted. 'It seems there is no easy way to implement this important ability into these models.'

The War of Predictions, Three Years Later

Encouraging solutions to resolve these disputes are offered by a small group known as the Risk Prediction Institute. In the summer of 2022, the institute launched a tournament called 'The Existential Risk Prediction Tournament' (XPT). Its goal was to 'create quality risk predictions threatening humanity over the next century.' Researchers surveyed experts in various fields, including risks associated with AI, and reached out to 'super forecasters' from different professions who had reliable track records in forecasting.

There were significant discrepancies between specialists and general forecasters. Experts believed that the risks they analyzed could lead to the extinction of humanity, while general forecasters pointed out that it is the experts' responsibility to prove why technologies that do not yet exist could be threatening. A recent study compared the predictions of both groups over the past three years.

It turned out that both experts and super forecasters systematically underestimated the pace of AI progress. The real outcomes of the latest models in the summer of 2025 exceeded expectations, with experts being closer to the truth. Super forecasters thought that AI would be able to win the summer mathematics olympiad in 2035, while experts predicted 2030. This occurred this summer. 'Overall, super forecasters assigned an average probability of only 9.7% to the observed results,' the authors of the study stated.

There Are No Oracles

Perhaps this situation could have been anticipated. Economist Ezra Karger, co-author of both papers, noted that after the publication of the first work, discrepancies between groups with different views on these matters were not so significant. This means that predictions regarding the dangers of AI are essentially more complex and speculative.

Thus, both experts and super forecasters failed to adequately account for the pace of AI progress, which may suggest we all underestimate this technology, which will continue to develop faster than we expect. This not only challenges our understanding but also emphasizes the importance of individual learning and observation.