Key Points:
- ChatGPT is a large language model (LLM) that operates based on an artificial neural network and has been trained using human-written texts scraped from the web and other sources.
- The performance and capabilities of LLMs are limited by the size of their network and the amount of data they are trained on, as well as the computational power required for training.
- Data availability is a major impediment to the progress of LLMs, as there is a vast amount of high-quality text stored in individual and corporate databases that is inaccessible at a reasonable cost or scale.
- LLMs like ChatGPT can provide plausible responses based on their training but cannot always verify or explain the reasoning behind their answers.
- LLMs are susceptible to inaccuracies, hallucinations, and misuse, which can have potentially harmful consequences.
- Fine-tuning pre-trained LLMs on curated, domain-specific data, implementing reinforcement learning from human feedback, and conducting “red teaming” exercises can help address some of the issues related to inaccuracies, accidents, and misuse.
- LLMs and AI/ML applications are essentially black boxes, and efforts to develop explainable or interpretable AI/ML methods have been slow and challenging.
- Regulations and comprehensive model risk governance policies are needed to ensure the safe and responsible deployment of LLMs and to guard against unforeseen risks and extreme scenarios.
- Model governance should focus on implementing use and safety standards, checking the accuracy and appropriateness of LLM output responses, and ensuring the presence of safeguards in the business process itself.
- Adoption of LLM technology requires new model risk governance standards to mitigate the inherent risks and limitations.
ChatGPT, like other large language models (LLMs), operates using an artificial neural network that has been trained on human-written texts scraped from the web and other sources. While LLMs like ChatGPT may seem magical due to their capabilities, they are ultimately just giant neural networks with a complex architecture consisting of about 400 core layers and 175 billion parameters or weights.
Training and tuning LLMs like ChatGPT require substantial amounts of data, computing power, and financial resources. For example, training ChatGPT-3 consumed about 1.3 gigawatt hours of electricity and cost OpenAI about $4.6 million. The larger ChatGPT-4 model is estimated to have cost $100 million or more to train.
However, the availability of data is a major obstacle to the progress of LLMs. While ChatGPT-4 has been trained on all the high-quality text available from the internet, there is a vast amount of high-quality text stored in individual and corporate databases that are inaccessible at a reasonable cost or scale. Access to such curated training data could significantly improve the performance of LLMs in domain-specific tasks and queries.
One of the limitations of LLMs like ChatGPT is their statistical nature. While they can provide plausible responses based on their training, they cannot always verify or explain the reasoning behind their answers. This becomes evident when LLMs are asked to solve mathematical problems, where they may provide incorrect responses or different answers when asked to describe the steps they took.
The statistical nature of LLMs also makes them susceptible to hallucinations and the spread of misinformation. They can be used for illegal or unethical purposes, such as writing fake news articles or obtaining sensitive personal information. To address these issues, pre-trained LLMs can be fine-tuned on curated, domain-specific data to improve the accuracy and appropriateness of their responses. Reinforcement learning from human feedback and “red teaming” exercises can also help reduce the potential for inaccuracies and misuse.
However, LLMs and AI/ML applications, in general, lack transparency and explainability. They are essentially black boxes, and even the programmers who developed them do not fully understand how they produce their output. Efforts to develop explainable or interpretable AI/ML methods have been slow and challenging, especially for large and complex models like ChatGPT.
To address the lack of transparency and potential risks associated with LLMs, regulations and comprehensive model risk governance policies are needed. These policies should prioritize the safe and responsible deployment of LLMs, with a focus on implementing use and safety standards, checking the accuracy and appropriateness of LLM output responses, and ensuring the presence of safeguards in the business process itself.
In conclusion, while ChatGPT and other LLMs represent significant advancements in AI/ML technology, their adoption comes with important limitations and risks. It is crucial for firms to adopt new model risk governance standards before deploying LLM technology in their businesses to ensure their safe and responsible use.
