> I asked chatGPT to recite a poem by Emily Dickinson. It recited it correctly but then added its own made up lines. The poem actually ends with the line “till

> I asked chatGPT to recite a poem by Emily Dickinson. It recited it correctly but then added its own made up lines. The poem actually ends with the line “till acclimated-to”. So why would it then tack on additional invented stanzas after correctly reciting the full poem? The core lowest-level ability of large language models (LLMs) is to generate the probability distribution of the next token in a sequence. Early ones were so dumb that they could not even form complete sentences. Now, they are so advanced that they can not only make grammatically correct sentences, but also amazing capabilities in natural language processing have emerged, allowing them for example to do well on various standardized tests. Some extra reinforcement learning or other processes may be stacked or wrapped on top of the basic LLMs. But at their very core they are "bullshitters" that say whatever they think might come next. With that understanding it's not so weird at all that it would try to continue the poem. Maybe a more advanced one (GPT-5?) would say that the poem is over and say that another poem is starting. > Seems strange and misleading. That description may be the best description yet that I've heard for the emergent abilities of large language models, and it will probably remain valid even as they reach and surpass human levels of cognition.