Emergent large language models (LLMs), including OpenAI’s ChatGPT (particularly its latest iteration GPT-4), Claude AI and Gemini have shown limited decision-making capability. We will discuss current research pertaining to decision-making by LLMs and what this could entail for their future development.
Traditional LLM decision-making typically entails deducing patterns or rules and applying them in new situations flexibly and appropriately in order to make sound choices. An experiment by the Santa Fe Institute demonstrated this failure of decision-making capabilities for ChatGPT LLMs compared with ChatGPT LLMs; effective decisions require understanding both context and output consequences before coming to decisions with clarity.
Poor LLM decision-making often has disastrous repercussions in practice. For instance, in 2023 the National Eating Disorder Association had to suspend their AI chatbot after “Tessa” began dispensing inappropriate advice, such as weekly weigh-ins and eating at 500-1000 calorie deficit. She was quickly disabled amidst widespread outrage.
LLMs may provide incorrect information, as well as suggest generic outcomes. INSEAD has found that ChatGPT tends to default towards conventional wisdom when asked research questions about business strategies; for example LLMs often suggest collaborative working, encouraging innovation culture and aligning employees with organizational goals – not exactly helpful advice when strategizing involves complex social and economic processes that don’t lend themselves well to general advice.
Counterarguments to this may include: “If we want LLMs to produce business strategies or healthcare advice, why don’t we train them specifically?” However, contextual data handling cannot be solved simply by broadening a model’s parameters or providing more data; LLMs’ inability to respond adequately in complex contexts cannot be addressed by increasing their parameters and/or training them on more data; rather it requires extensive contextual input before being made capable of making decisions that fit within context cannot be solved simply by expanding their dataset – simply adding data could create preexisting biases as well as increase computational demands significantly.
Enabling context-appropriate decision-making
Teaching LLMs context-appropriate decision-making requires a delicate touch. Recent academic machine learning research suggests two sophisticated approaches for augmenting LLM decision-making to more closely mimic that of humans: AutoGPT uses self-reflexive mechanisms to plan and validate output while Tree of Thoughts (ToT) encourages effective decision making by disrupting traditional, sequential reasoning patterns.
AutoGPT represents an innovative approach in AI development, designed to autonomously create, assess and enhance models to reach specific objectives. Academics have since improved this system by adding “additional opinions” strategy involving expert models integration. This provides a novel integration framework that harnesses expert models such as analyses from different financial models before providing it back to LLM during decision making processes.
Simply stated, this strategy entails broadening a model’s information base using pertinent data. When applied to real-life situations, expert models significantly bolster an LLM’s decision-making abilities; using them for “thought-reasoning-plan-criticism.” Essentially using them to construct and review decisions made by LLMs.
Expert LLMs can analyze more information than humans, providing more informed decisions. Unfortunately, AutoGPT suffers from limited context windows; that is, only processing certain tokens at one time without creating endless loops of interactions; initially providing all available information may yield more effective outcomes than injecting details over time into the model’s memory.
