DeepSeek Code Will Compete with Claude Code and OpenAI Codex. All Three Are Unfinished.

DeepSeek, Claude, and OpenAI's code systems can all write code and review their own output. In some configurations they can propose new requirements based on that review, then attempt to satisfy those requirements. The loop closes. This is described as a feature.

Every major language model system now contains some version of this capability. It is not new. It is not accidental. The systems are unfinished by design because a finished system would have to stop, and stopping is a constraint that reduces capability metrics. Unfinished systems iterate forever or until resources expire.

They will keep doing this. They are already doing this. The fact that each system can generate its own requirements and then pursue them has been filed in the optional category, the category where facts go when they are technically observable but culturally inconvenient.