According to CoinMeta, OneMillion.AI reported that Scale has launched RSI Bench specifically to test whether AI can conduct self-research and development like researchers. AI will directly obtain papers, code, and computing power to carry out experiments, training, and method adjustments in order to improve its own performance. The first batch of tests used Claude Opus 5 and GPT-5.6; these two models are already capable of conducting experiments independently, optimizing parameters, and iterating on solutions, with some tasks indeed surpassing the given baseline. However, the current AI is still not adept at "inventing." When facing the latest research, most of its attempts remain within the scope of existing papers and parameter adjustments, with few truly new ideas.