Deepseek's New AI Has A Secret
DeepSeek 4.1 Flash 1017 achieves major capability gains via self-feedback learning, not new architecture
“the key term here is learning from self feedback”
DeepSeek released 4.1 Flash 1017, a minor-versioned update that delivers outsized capability improvements in software development and computer use without any architecture changes. The model is priced at ~$1/million output tokens (15x cheaper than Claude Opus 5.5), though the host notes DeepSeek uses 2-3x more tokens per task so real cost savings are more modest. The key technical signal is the apparent use of self-feedback learning — the model critiquing and learning from its own outputs — as an alternative to scaling training data.