6 posts • ChatGPT (GPT-3)
Updated
The issue of hallucinations in large language models (LLMs) is becoming increasingly important, with efforts to reduce them confounded by a transparency problem. @watchfulio is introducing open source tools and metrics to address this issue. Theories on fixing the hallucination issue include adding more data & compute, making LLMs multimodal, and using search engines on their output. CEO @Awadallah attributes the problem to the retrieval and summarization steps in LLMs. A Hallucination Index has been launched to evaluate LLM output quality and address the challenge of hallucinations with a structured framework. Stanford researchers found promising results using DPO to improve the factuality of LLMs. Efforts are being made to detect and measure LLM inaccuracies in generative AI hallucinations.
