Users argue for better training data over new AI post-training algorithms

One user interprets DeepSeek’s position as saying that improving data quality offers far greater returns than developing new post-training algorithms. They recommend devoting 80% of post-training effort to data, including expert task reviews, manually removing suspicious samples, checking that tasks are passable, and ensuring variety in difficulty and category. Another user challenges the portrayal of DeepSeek as focused on post-training algorithm design, while defending what they describe as its view that the existing algorithms are good enough.