Moderate forgetting benefits model performance
Categories:
The GPT-5 series has a high recall rate and can achieve very rigorous logic, making it excellent for debugging. At the same time, it can promptly detect contradictions in the context, offering a unique experience among current models. But recently, in my personal reverse engineering project, I found that GPT-5 has limited reverse engineering capabilities, clearly inferior to the opus model. It often explicitly replies that something won’t work, while switching to opus can quickly achieve functional reverse engineering. Opus records its reverse engineering process as an operation attempt history. After several days of my intensive use, the history document contains many contradictions. GPT can avoid trusting an outdated instruction, but when there is too much outdated information, GPT-5 will struggle to handle it correctly. In other words, the GPT-5 series models will be limited in a low-quality project full of holes, while opus can still maintain relatively good performance.
