>What fundamental limitation of these large language models makes them unable to produce natural language?
LLMs can get indefinitely good at coding problems by training in a reinforcement learning loop on randomly generated coding problems with compiler/unit tests to verify correctness. On the other hand, there's no way to automatically generate a "human thinks this looks like slop" signal; it fundamentally requires human time, severely limiting throughput compared to fully automatable training signals.
LLMs can get indefinitely good at coding problems by training in a reinforcement learning loop on randomly generated coding problems with compiler/unit tests to verify correctness. On the other hand, there's no way to automatically generate a "human thinks this looks like slop" signal; it fundamentally requires human time, severely limiting throughput compared to fully automatable training signals.