As we approach mid-2027, the key factors in AI research will be the taste of AI agents and the amount of compute available for experiments, as these will determine the effectiveness of automated AI research.

The efficiency of AI agents in conducting research will depend on their ability to learn from experiments and adapt their strategies, which may lead to breakthroughs in various fields.

The emergence of AI agents allows models to autonomously retrieve information, run code, and perform multiple tasks without human intervention, leading to significant economic impacts.

In the coming years, we will see more experiments on how to effectively deploy and manage multiple AI agents for complex tasks.

AI 2027 is our scenario trying to forecast the next few years of AI progress, aiming to provide a concrete story that shows the transitional steps leading to AGI in 2027 and potentially superintelligence in 2028.

By 2025, we expect AIs to improve coding capabilities, but they won't be able to operate autonomously for long periods, still making basic errors.

By 2028, I believe we'll have advanced AI systems capable of solving complex problems that humans can handle, thanks to significant ongoing work in AI development.

The intelligence explosion is expected to occur in 2027, with AIs significantly accelerating research progress through a multiplier effect.

The potential for AI agents to perform tasks in virtual environments is promising, with a clear path for increasing their capabilities through reinforcement learning.