Even the best LLMs are very bad at solving real problems because fundamentally their mechanism is to probabilistically select next tokens from a statistical, thus approximate, model of all the data they were trained on. We are developing technologies to go beyond that.