To be fair, who can?

MIT boffins have demonstrated that even the best-performing large language models (LLMs) lack a coherent understanding of the world and its rules, which can lead to unexpected failures on similar tasks.