MIT Researchers Cracked the Code | PDDL-Instruct Research Deep-Dive скачать в хорошем качестве

MIT Researchers Cracked the Code | PDDL-Instruct Research Deep-Dive

In this video, I do a comprehensive analysis into PDDL-Instruct, a new multi-step fine-tuning methodology that allows Large Language Models (LLMs) to generate fully logically consistent plans ~3x more frequently. Researchers tested their new method using fine-tuned variants of GPT-4 and Llama-3-8B across a variety of problem domains. I also discuss the implications of this research in multi-agent systems, embodied AI (robotics), and its limitations. Timestamps: 0:00 - Intro 0:19 - Table of Contents 00:45 - Overview of the Research 2:29 - How Chain-of-Thought Works 3:08 - Limitations of Chain-of-Thought 8:41 - Methodology: PDDL-Instruct 11:08 - How Researchers Did It 17:04 - How State Spaces Work 22:43 - Results & Benchmarks 26:09 - Limitations & Error Rates 29:26 - Implications for Agentic Systems 31:23 - Outro #openai #meta #agi #llm #gpt #finetuning #research #agi #artificialintelligence #ainews #largelanguagemodels