Linguistic Interaction in Humans and Machines
Talkat Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives · 2024-06-24..06-28·
summary
CerealBar is a procedurally generated leader–follower game used to train instruction generation by deploying, observing user executions, aggregating outcomes, and retraining; over 14 deployment rounds task completion rises from ~45% to ~80% but the model’s vocabulary collapses toward its own jargon (improves via interaction, drifts from human language). KiloGram digitizes >1,000 tangrams with ~13,000 part and whole-shape annotations; humans form conventions that generalize across new shapes, while CLIP/ViLT stay near chance on tangrams and GPT‑4 scores ~21–23%.
AI-written summary · about
chapters (10)
- 0:00Collaborative linguistic interaction
- 1:02CerealBar scenarios
- 8:32Instruction generation from interactions
- 15:25Language abstraction and adaptation
- 16:33KiloGram for abstraction
- 24:10Evaluating models with reference games
- 29:05Human convention formation
- 33:40LLMs and convention formation
- 37:14Future directions in language modeling
- 39:04Q&A
related talks
01h 00mHuman and machine inductive biases for compositional linguistic generalization
01h 50mChatGPT & LLMs: A Practical Starter's Guide to Language Chains
21mChatGPT, LLMs, and LangChains - A Beginner's Guide
01h 18mThe Enigma of LLMs: on Creativity, Compositionality, Pluralism, and Paradoxes
49mBen Hall & Barbara Fusinska
43mAI Has Entered The Chat: Humans, AI, and the Art of Collaboration
41mThe Value of LLMs - Why Stupid Machines Are Smarter Than You Think
43mBeyond the Hype: A Realistic Look at Large Language Models