Large Language Model (LLM)

June 2017: eight Google researchers publish "Attention Is All You Need" and the transformer architecture; in May 2020, OpenAI's GPT-3 has 175 billion parameters. A large language model is a deep-learning model trained on enormous text corpora to predict the most probable next word, or more precisely token, given the preceding context. That simple-looking task gives rise to complex capabilities: writing, translation, summarization, apparent reasoning, code generation. Most LLMs rest on the transformer architecture, which weighs the relative importance of the elements in a text. They work probabilistically. They produce statistically plausible sequences without consulting a knowledge base, which explains both their fluency and their tendency toward confident error, known as hallucination. For insurance and risk management, the LLM is the cognitive engine of generative and agentic AI systems, and so the source of new risks: opaque reasoning, unpredictable behavior at scale, vulnerability to semantic manipulation. The LLM is neither a database nor a reliable logical reasoner. Mistaking the fluency of its output for a guarantee of accuracy is a common error of judgment, and a costly one in critical professional uses.