Stanford CS25: Transformers United V6 I Overview of Transformers

Key Concepts

• Transformers: Deep learning architecture utilizing self-attention mechanisms to process sequential data in parallel.
• Self-Supervised Learning: Training models by masking or corrupting input data to learn general representations without explicit labels.
• Retrieval-Augmented Generation (RAG): Augmenting LLMs with external domain-specific documents to improve output accuracy.
• Chain of Thought (CoT): A prompting technique where models generate intermediate reasoning steps before providing a final answer.
• Hallucination: Defined as a "world modeling error" where a model’s internal beliefs contradict a ground-truth reference world.
• Mechanistic Interpretability: Techniques to unpack the "black box" of neural networks by analyzing internal circuits and features.
• State Space Models (SSMs): An alternative to Transformers (e.g., Mamba) that uses linear-time scaling for long sequences.
• JEPA (Joint Embedding Predictive Architecture): A world model approach that predicts latent representations rather than raw tokens.

---

1. Course Overview and Logistics

• Objective: Explore the architecture, scaling, and cutting-edge applications of Transformers across domains like biology, neuroscience, and robotics.
• Structure: Weekly guest lectures from academia and industry.
• Requirements: Attendance is mandatory (in-person preferred). The course is supported by AI House and MongoDB, offering students networking opportunities via the "Frontier Lunch Club."

2. Evolution of Machine Learning Architectures

• Pre-2012: Hand-engineered features fed into shallow models.
• Supervised Deep Learning: Transitioned to raw data processing, removing the "middleman" of manual feature engineering.
• RNNs/LSTMs: Sequential processing with hidden states; limited by vanishing gradients and inability to parallelize.
• Transformers: Introduced Self-Attention (Query, Key, Value matrices) to learn relationships between all tokens in a sequence simultaneously, enabling massive parallelization on GPUs.

3. Pre-training and Data Strategy

• Scaling Laws: Performance scales with data and model size, but there are diminishing returns.
• BabyLM Project: Investigated if models could learn from the limited data a human child is exposed to (10–100 million words). Findings suggest that data quality, structure, and diversity are more critical than raw volume at small scales.
• Multilingual Learning: Research indicates that training on multiple languages simultaneously does not cause "confusion" or degrade performance in a primary language.
• Curriculum Guided Layer Scaling (CGLS): A methodology where model size and data complexity increase in tandem during training, outperforming static training approaches.

4. Post-training and Reasoning

• Chain of Thought (CoT): Encourages models to decompose complex problems into sub-steps.
• Reinforcement Learning (RL):
  • RLHF: Training reward models based on human preferences.
  • DPO (Direct Preference Optimization): An RL-free approach to align models with human preferences.
  • GRPO (Group Relative Policy Optimization): Ranks responses within a group for finer-grained feedback.
  • Process Supervision: Rewarding individual reasoning steps rather than just the final output to reduce "reward hacking."

5. Applications Beyond Language

• Vision Transformers (ViT): Divides images into patches to leverage Transformer flexibility, avoiding the locality assumptions of CNNs.
• Neuroscience (fMRI): Using Transformers to model brain networks. By masking entire functional networks (e.g., the "daydreaming" Default Mode Network), researchers can predict disease progression (e.g., Alzheimer’s) and map inter-network dependencies.

6. Challenges and Future Directions

• Hallucination: The instructors propose a unified framework: Hallucination = (Reference World Model) vs. (Model View) + (Conflict Resolution Policy).
• Memory and Continual Learning: Current models are largely stateless. True lifelong learning requires updating model weights (the "brain") rather than relying on inference-time context.
• Alignment: Ensuring models pursue intended goals via Constitutional AI and scalable oversight (using AI to supervise AI).
• Beyond Transformers: The field is shifting toward SSMs (Mamba) for linear-time efficiency and World Models (JEPA), which focus on predicting latent states rather than next-token probability.

Synthesis

The course emphasizes that the future of AI lies in moving beyond simple "next-token prediction" on massive datasets. Success requires a shift toward smarter data utilization, grounded world models, efficient architectures (SSMs), and robust alignment strategies. The instructors argue that while Transformers currently dominate, the path to AGI likely involves solving fundamental issues in memory, interpretability, and autonomous reasoning.