A Preliminary Study of o1 in Medicine: Are We Closer to an AI Doctor?

• Overview
  • Emergent abilities and scaling effects
  • Understanding LLMs and their architecture
  • Training procedures and inference methods
  • Evaluating LLMs and exploring multimodal models
• Language Models as Generalists
  • LMs can perform multiple tasks by predicting the next token
• Paradigm Shift with LLMs
  • The transition from task-specific models to general-purpose ones like BERT (Oct 2018) and GPT (Jun 2018)
• GPT-2: Generalizing Without Fine-Tuning
  • Issues with fine-tuning pre-trained models:
    • Proliferation of model copies
    • Need for large labeled datasets
    • Limited to specific tasks
  • Aim to train models capable of zero-shot learning
• From Single-Task to General Models

  • Shift in modeling assumptions:

    • Single Task: P(output ∣ input)

      P(output ∣ input)P(\text{output} \mid \text{input})

    • General Model: P(output∣input,task)

      P(output∣input,task)P(\text{output} \mid \text{input}, \text{task})

  • Use text prompts to specify tasks (e.g., "Translate this French text to English")

• Key Factors in GPT-2's Success
  • Diverse training data enables broad capabilities without additional training
  • Scaling up model size and dataset volume enhances performance