Recently, I noticed that Google DeepMind has developed something quite interesting called SIMA 2. Simply put, it is an AI system capable of learning and playing games by itself within a virtual environment. This latest progress is indeed a bit different.

Compared to the previous generation, SIMA 2's task completion rate jumped from 31% directly to 65%, which is a pretty significant improvement. More importantly, it can now understand more complex high-level goals, meaning you don't have to give it very detailed instructions; it can interpret what you want it to do on its own. In games, it can even collaborate with virtual characters and transfer what it has learned in one environment to another.

On the technical side, SIMA 2 is powered by Gemini and supports text, speech, and image inputs, making interaction more flexible. Interestingly, it can also generate tasks itself for iterative learning, a self-driven learning approach that is still relatively new in the AI field.

However, the paper also honestly admits its limitations. SIMA still struggles with complex tasks that require many steps, and it faces challenges in visual understanding within 3D environments. These are areas that need breakthroughs in the future. Overall, this iteration of SIMA 2 marks a step toward general AI, although there are still many hurdles to overcome, but the direction is correct.