The ChatGPT Paradox: Impressive Yet Incomplete

Prof. Thomas G. Dietterich discusses the current state of large language models like ChatGPT. He explains their capabilities and limitations, emphasizing their statistical nature and tendency to hallucinate. Dietterich explores the challenges in uncertainty quantification for these models and proposes integrating them with formal reasoning systems. He advocates for more robust knowledge representation methods, such as knowledge graphs, and discusses the importance of safety in AI development. Dietterich also touches on the changing landscape of academic AI research and the potential of open-source models to accelerate innovation in the field. MLST is sponsored by Tufa Labs: Are you interested in working on ARC and cutting-edge AI research with the MindsAI team (current ARC winners)? Focus: ARC, LLMs, test-time-compute, active inference, system2 reasoning, and more. Future plans: Expanding to complex environments like Warcraft 2 and Starcraft 2. Interested? Apply for an ML research position: [email protected] Thomas G. Dietterich Distinguished Professor (Emeritus), Computer Science, Oregon State University https://scholar.google.com/citations?user=09kJn28AAAAJ&hl=en https://medium.com/@tdietterich https://x.com/tdietterich TOC: 00:00:00 LLMs: Capabilities, Limitations, and Statistical Nature 00:09:25 Uncertainty Quantification in Large Language Models 00:15:49 Integrating Reasoning and Formal Systems with LLMs 00:19:39 Knowledge Structures and Future AI Architectures 00:20:11 Knowledge Extraction and Representation in AI 00:23:51 Challenges in LLM-based Knowledge Systems 00:33:12 Scientific Publishing and Knowledge Graphs 00:37:14 AI Safety and Truth in Knowledge Representation 00:39:56 AI Regulation and Safety Engineering 00:41:57 Challenges in AI Perception and Novelty Detection 00:50:04 Verification and Safety in Complex AI Systems 00:59:23 Open-Source Models and Academic AI Research 01:03:16 Limitations of Transformers and Future AI Architectures REFS: 00:00:16 Meta releases the biggest and best open-source AI model yet, https://www.theverge.com/2024/7/23/24204055/meta-ai-llama-3-1-open-source-assistant-openai-chatgpt 00:00:27 Tamper-Resistant Safeguards for Open-Weight LLMs, https://arxiv.org/pdf/2408.00761 00:00:41 Identifying latent disease factors, https://arxiv.org/pdf/2410.07890 00:00:53 EY position paper on AI, https://www.ey.com/en_ch/news/2023/12/new-ey-position-paper-on-artificial-intelligence 00:00:53 RAG for Knowledge-Intensive NLP Tasks, https://arxiv.org/pdf/2005.11401 00:00:55 AI's international research networks mapped, https://www.nature.com/articles/d41586-024-02986-2 00:01:15 A Survey of Large Language Models, https://arxiv.org/pdf/2303.18223 00:03:30 Embers of Autoregression, https://arxiv.org/pdf/2309.13638 00:07:33 GPT-4 Technical Report, https://cdn.openai.com/papers/gpt-4.pdf 00:10:52 LM-Polygraph, https://arxiv.org/pdf/2311.07383 00:14:16 Shifting Attention to Relevance, https://arxiv.org/pdf/2307.01379 00:16:00 Retrieval-Augmented Generation, https://arxiv.org/pdf/2005.11401 00:20:34 Baldur: Whole-Proof Generation, https://arxiv.org/pdf/2303.04910 00:22:11 Never-Ending Learning, https://www.cs.cmu.edu/~tom/pubs/NELL_aaai15.pdf 00:25:36 Machine Unlearning, https://arxiv.org/pdf/2404.01206 00:36:09 Papers with code or without code, https://www.sciencedirect.com/science/article/abs/pii/S0306457323002145 00:39:12 CYC: a large-scale investment in knowledge infrastructure, https://dl.acm.org/doi/10.1145/219717.219745 00:40:23 Community Notes, https://communitynotes.x.com/guide/en/about/introduction 00:41:39 REGULATION (EU), https://eur-lex.europa.eu/eli/reg/2024/1689/oj 00:43:44 Open Category Problem, https://futureoflife.org/ai-researcher-profile/ai-researcher-thomas-dietterich/ 00:50:38 BO as Assisstant, https://dl.acm.org/doi/pdf/10.1145/3526113.3545664 00:50:59 Testing of Autonomous Vehicles, https://research-assets.waabi.ai/GUARD/paper.pdf 00:54:02 Engineering a Safer World, https://mitpress.mit.edu/9780262533690/engineering-a-safer-world/ 00:54:50 The nature of the Internet, https://www.pnas.org/doi/epdf/10.1073/pnas.0501426102 00:58:43 Anomaly and Novelty Detection, https://dl.acm.org/doi/10.1145/3447548.3469453 00:59:06 Causality 2nd Edition, https://www.amazon.ca/Causality-Judea-Pearl/dp/052189560X 01:03:38 LLaMA: Open and Efficient Foundation Language Models, https://arxiv.org/pdf/2302.13971