Generative AI (genAI) is becoming increasingly popular among the public and various businesses. However, its adoption is often hindered by errors, copyright infringement, and hallucinations, which can undermine trust in its accuracy.
According to a study from Stanford University, genAI makes mistakes when answering legal questions 75% of the time. The study found that most large language models (LLMs) behind genAI technology, such as OpenAI’s GPT-4, Meta’s Llama 2, and Google’s PaLM 2, are not only amorphous with nonspecific parameters but are also trained by fallible human beings with innate biases.
These large language models (LLMs) have been described as stochastic parrots, becoming more random in their conjectural or random answers as they get larger. One method of reducing genAI-related errors is Retrieval Augmented Generation or “RAG,” which creates a more customized genAI model for more accurate and specific responses to queries.
However, genAI’s natural language processing lacks transparent rules of inference for reliable conclusions. Some argue that a “formal language” or a sequence of statements is needed to ensure reliable conclusions at each step of the way toward the final answer genAI provides.
With monitoring and evaluation, genAI can produce vastly more accurate responses. David Ferrucci, founder and CEO of Elemental Cognition, compared it to the straightforward agreement that 2+2 equals 4, emphasizing the need for unambiguous final answers.
GenAI has faced issues, such as Google’s new Gemini tool, which created biased images based on user text prompts. To address these problems, Elemental Cognition developed a ”neuro-symbolic reasoner.”
2024-03-15 15:00:03
Original from www.computerworld.com