Simons Institute Courses

The Inadequacy of Offline LLM Evaluations - A Need to Account for Personalization in Model Behavior

Discover how personalization fundamentally changes language model behavior, revealing why standard offline evaluations fail to capture real-world AI performance in user interactions.

• YouTube
• 32 minutes
• Self-Paced
• Free Video

Query Languages for Machine Learning Models

Explore query languages for ML models, focusing on recent proposals for discrete classification and real-valued models, their expressiveness, and evaluation complexity for trustworthy AI applications.

• YouTube
• 43 minutes
• Self-Paced
• Free Video

Synthesizing Pareto-optimal Interpretations of Black Box ML Models

Explore the synthesis of Pareto-optimal interpretations for black-box ML models, balancing explainability metrics with accuracy using MaxSAT solving techniques that provide PAC-style guarantees.

• YouTube
• 49 minutes
• Self-Paced
• Free Video

Alignment Problems in AI Governance

Explore critical challenges in AI governance, from technical limitations in model understanding to regulatory misalignment, with expert insights on adaptive oversight solutions.

• YouTube
• 1 hour 5 minutes
• Self-Paced
• Free Video

Linear Representations of Concepts in Modern AI Models

Discover how concepts are linearly encoded in neural networks using probes and Recursive Feature Machines to extract human knowledge from Large Language Models.

• YouTube
• 37 minutes
• Self-Paced
• Free Video

Self-Supervised Learning in Transformer Models for Birdsong

Explore how transformer models use self-supervised learning to decode and analyze birdsong communication patterns in nonhuman species research.

• YouTube
• 53 minutes
• Self-Paced
• Free Video

Probabilistic Safety Guarantees Using Model Internals

Explore probabilistic safety guarantees for language models through analysis of model internals with Jacob Hilton from Alignment Research Center.

• YouTube
• 46 minutes
• Self-Paced
• Free Video

Generalization in the Representations and Computations of Frontier Language Models

Delve into the geometry of language model representations, exploring superposition strategies, sparse autoencoders, and their implications for AI reliability and computational advancement.

• YouTube
• 52 minutes
• Self-Paced
• Free Video

The Curious Incident of Developing Artificial General Intelligence - From Theory to Future Challenges

Explore the evolving landscape of AI development, from human-like capabilities to superintelligence, examining key theories, challenges, and future implications for artificial general intelligence.

• YouTube
• 45 minutes
• Self-Paced
• Free Video

Understanding the Abilities of AI Systems - Memorization, Generalization, and Points in Between

Explore how Large Language Models develop abilities through training data, examining their capacity for generalization versus reliance on memorized patterns and heuristics.

• YouTube
• 44 minutes
• Self-Paced
• Free Video

Weak-to-Strong Generalization: Exploring Supervision Methods for Advanced Language Models

Explore how weak model supervision can elicit capabilities from stronger AI models, examining implications for alignment and scaling of techniques like RLHF through empirical studies.

• YouTube
• 43 minutes
• Self-Paced
• Free Video

Transformers Can Learn Compositional Function

Explore how transformers learn compositional functions with Jason Lee from Princeton University in this Simons Institute talk on the future of language models.

• YouTube
• 1 hour 4 minutes
• Self-Paced
• Free Video

Advancing Diffusion Models for Text Generation

Explore how diffusion models can be adapted for text generation, combining the fluency of autoregression with plug-and-play control to overcome limitations of transformer-based language models.

• YouTube
• 1 hour 2 minutes
• Self-Paced
• Free Video

Reducing the Dimension of Language: A Spectral Perspective on Transformers

Explore a novel approach to sequence modeling that leverages dynamical systems principles to achieve long-range memory, fast inference, and provable robustness beyond traditional Transformers.

• YouTube
• 1 hour
• Self-Paced
• Free Video

Transfer Learning for Weak-to-Strong Generalization in Language Model Alignment

Explore how to effectively transfer latent knowledge from weak to strong language models, focusing on alignment techniques and practical applications in LLM development.

• YouTube
• 42 minutes
• Self-Paced
• Free Video