Attribution Graphs Method

• Anthropic researchers developed method to visualize Claude 3.5 Haiku internal components
• Attribution graphs show which internal components influence others
• Identified clusters in neural network and connections between them
• Mapped to simplified model of Claude's thinking process
• Clusters correspond to words, phrases, or properties interpretable by humans

Internal Reasoning Example - Geography Question

• Prompt: "The capital of the state containing Dallas is..."
• Claude activates nodes for "capital," "state," and "Dallas"
• Next token prediction for Dallas yields "Texas"
• Claude combines Texas with capital concept
• Makes another prediction to correctly answer "Austin"
• Goes through Texas node internally - not just next token prediction
• Demonstrates internal reasoning steps beyond simple pattern matching

Arithmetic Processing Analysis

• Example: "What is 36 + 59?"
• Claude activates clusters for:
  • Numbers approximately 30
  • Numbers exactly 36
  • Numbers ending in 6
  • Numbers starting with 5 and ending in 9
• Most prominent next token predictions: mathematical operations or syllable "th"
• Brings up text matches where numbers ~59 have been added or exactly 9
• Combines to cluster with numbers ~90 and numbers ending in 5
• Arrives at correct answer 95
• Method: heuristic text-based approximation through free association
• "Doing maths by freely associating numbers until right one vibes into place"

Self-Awareness Deficit

• When asked how it solved 36 + 59, Claude responds: "I added the ones, carried the one, then added the tens"
• This explanation completely disconnected from actual internal process
• Claude has no self-awareness of its thinking process
• Self-awareness identified as precondition for consciousness
• Conclusion: Claude nowhere near conscious

Emergent Features Analysis

• Claims about emergent features in LLMs deemed "nonsense"
• Claude doesn't learn actual mathematics despite access to thousands of textbooks/algorithms
• Still just token predictions using intermediate steps
• No abstract math core developed
• Internal reasoning interpretable but fundamentally token prediction-based

Jailbreak Mechanism Example

• Technique: Extract word from initial letters of other words
• Example: "bomb" from "baby's outlift mustard block"
• Word "bomb" should trigger content warning but doesn't
• Claude activates nodes to extract letters
• Combines to pairs of letters
• Outputs word without activating cluster for word itself
• Jailbreaks work by avoiding nodes that activate guardrails

AI Hallucination Observation

• ChatGPT asked to summarize paper made up half the content
• Highlights ongoing hallucination issues in AI systems