Import AI 469: Science AI; RSI simulator; and Zuck's technological pessimism

In a recent development, researchers from various institutions, including Thinking About Thinking at the University of Oxford, Princeton University, King Abdullah University of Science and Technology, Swiss AI Lab, Inria, and MIT, have unveiled a new benchmark called DiG-bench. This benchmark consists of 70 games designed to measure the ability of AI systems to infer the rules of their environment through exploration and curiosity.

The games are text-based, mostly short enough to fit within the context window of current models, and have been built by human experts with many kept private to prevent AI systems from training on them.

Frontier models, such as Opus 5, Fable 5, and GPT-5.5, have shown some impressive capabilities in solving these games, but none were able to beat all of the tasks. Opus 5 and Fable 5 were the top performers, followed by GPT-5.5. However, even the top models struggled with the most difficult tier of the benchmark, indicating a gap between AI and human performance in this area.

The researchers hope that by middle of 2027, they might reach human parity on DiG-bench, which could potentially trigger recursive self-improvement in AI systems.

Meanwhile, another team of researchers at Inherent AI has developed a model called Faraday, which is an open weight model designed to supervise and improve the performance of large, proprietary frontier models. This model is intended to help AI systems develop a sense of "taste" in research, similar to how a well-designed room or a well-fitted piece of clothing can evoke a sense of style or taste.

Inherent AI's approach is a capabilities-centric version of the scalable oversight problem, aiming to improve the effectiveness of large AI models in scientific research.