Limitations of Large Language Models and Proposed Alternatives
Tom Dietterich discusses the limitations of large language models and proposes alternative approaches to improve their performance and reliability.
00:00:02 Large language models have impressive capabilities but also have problems. They produce inconsistent, incorrect, and biased answers. They can generate dangerous or socially unacceptable content. They are expensive to train and lack the ability to update knowledge. Lack of attribution is another issue.
📚 Large language models have impressive capabilities but also many problems.
❌ They often produce incorrect and contradictory answers, known as hallucination.
🚫 They can generate dangerous or socially unacceptable answers, reflecting biases.
💰 They are extremely expensive to train and lack the ability to update knowledge.
❓ There is no easy way to determine the sources responsible for the answers they give.
00:07:12 In his talk at #VSCF2023, Tom Dietterich discusses the limitations of large language models as knowledge bases and proposes the use of retrieval augmented language models to enhance question answering.
🔑 Large language models are statistical models of knowledge bases, not actual knowledge bases.
🌐 Retrieval augmented language models retrieve relevant sections of documents to answer questions.
💡 Using external documents in language models reduces hallucination and provides source attributions.
00:14:20 The video discusses the limitations of large language models (LLMs) and proposes alternative approaches to improve their performance and reliability. It covers issues such as contamination of statistical knowledge, lack of attribution, consistency, dangerous outputs, bias, and the need for inspectable models. It also mentions attempts to train models to recognize inappropriate content and the integration of non-linguistic knowledge.
📊 48% of sentences generated by language models are not fully supported by retrieved documents.
💡 Improving consistency by asking language models a set of questions and refining the answers.
🔒 Challenges in reducing dangerous or socially inappropriate outputs and defining what is appropriate.
🤖 Efforts to combine linguistic and non-linguistic knowledge in language models.
00:21:33 Tom Dietterich discusses the limitations of large language models and proposes building modular systems with separated language, common sense, and factual world knowledge. He suggests integrating episodic memory, situation modeling, and formal reasoning to overcome these limitations.
🔑 The current large language models combine language understanding, common sense knowledge, and factual world knowledge into one component, making it difficult to update the factual world knowledge in real time.
💡 The large language models lack episodic memory and situation modeling, which are crucial for understanding narratives and conversations.
🧠 Building modular systems that separate factual world knowledge, common sense knowledge, and language understanding while integrating episodic memory, situation modeling, formal reasoning, and planning can overcome the shortcomings of large language models.
00:28:44 Tom Dietterich discusses the limitations of LLMs and proposes building a system that extracts facts from a knowledge graph and trains dialogue systems. He also explores the challenge of truthfulness in models and suggests training systems to provide arguments and justifications.
Instead of manually constructing knowledge graphs, we can extract relevant facts from paragraphs and add them to existing knowledge graphs.
Tom Mitchell's project, the never-ending Learning System, attempted to create a knowledge graph by searching the web and using natural language extraction tools.
Using large language models, such as chat GPT, can effectively extract simple facts from paragraphs, which can be used to build knowledge graphs.
There is ongoing research on extracting knowledge graphs from large language models and from documents, as well as developing dialogue systems that can generate responses based on conversation history and knowledge graphs.
Achieving truthfulness in AI systems is challenging due to differing beliefs, lack of conclusive evidence, and cultural variations. One approach is to train systems to provide arguments and justifications for their beliefs.
Building upon previous work in knowledge representation and artificial intelligence, training systems to output answers along with arguments and justifications can lead to more truthful AI.
00:35:54 Large language models have surprising capabilities but their flaws include the inability to distinguish between different types of uncertainty, the inability to update with new knowledge, and the production of socially unacceptable outputs. Instead, modular systems that separate linguistic skill from other components should be built.
The trustworthiness of websites is a challenge in search engine optimization.
Large language models lack the ability to reason about ongoing processes.
Large language models treat everything as random, leading to socially unacceptable outputs.
Building modular systems that separate linguistic skills from other components is important.
00:43:05 Tom Dietterich discusses the limitations of large language models (LLMs) and proposes alternative approaches. He highlights the challenge of miscalibration and the need for funding and open-source efforts to address these issues.
🤔 There are limitations with neural network technology, such as its inability to represent variations not in the training data.
💡 Government funding for large enough computing facilities is needed to address miscalibration and make progress on the problems with language models.
✅ Systems with a way to verify correctness, such as executing code or running program analysis, can overcome some of the limitations of large language models.
Want to deep dive into this video?
