OpenAI has released a research paper and accompanying blog post noting a potential step forward in the development of chatbots. Known as “instruction-following large language models” (such as OpenAI’s ChatGPT and rivals such as Google’s Bard and Anthropic’s Claude), these systems, according to the AI engineers at these companies, have the potential to transform businesses. However, they are often unreliable, prone to errors and can pose risks such as outputting toxic language or encouraging unsafe or illegal behaviour, leading many firms to search for ways to mitigate these risks.
OpenAI’s latest research centres around a process known as “reinforcement learning from human feedback” to tame the incorrect responses often created by these models. The idea is that humans select the responses generated by the model that best answer the inquiry, and the model focuses its attention on parameters that lead to the correct answers. OpenAI then improved this approach by asking machines to think step-by-step when approaching specific problems, allowing them to understand the process of logical problem-solving better, with engineers rating the results produced throughout each step of the process. Ultimately, machines will still only output what humans have taught them, leading to questions as to whether any such system can genuinely exhibit creativity.
The difficulties of applying these models to humans and their values, adding credibility to the argument that we are yet to see a truly intelligent machine capable of exhibiting creativity.
OpenAI’s findings could be a step backwards in the
