According to METR's recent analysis, AI systems are beginning to close the performance gap with human experts in completing tasks across various fields like coding and cybersecurity. Though current AI models like GPT-2 struggle with tasks exceeding one minute for human experts, models like Claude 3.7 have shown a marked improvement. The study observes that AI's task-completion efficiency has doubled approximately every seven months, accelerating to every three months in 2024. Predictions suggest by 2029, AI may tackle tasks requiring a month of human input at a 50% success rate, although experts warn against overly optimistic extrapolations due to uncertainties in practical AI applications.
Today's AI systems, while currently behind humans on longer tasks, are accelerating in development and may close the gap sooner than expected.
The analysis indicates that the time horizon for AI models' task completion has been improving exponentially, suggesting future capabilities might reach human expert levels.
The findings suggest AI models could complete tasks that currently take humans a month working at 50% reliability by as early as 2029.
Experts like Joshua Gans caution against extrapolating current progress, emphasizing the uncertainties surrounding the actual deployment and utilization of AI.
Collection
[
|
...
]