Human vs Language Model: Token Generation Speed¶

By Mike Bailey using Claude.ai

Speed Comparison Table¶

Detailed Breakdown¶

Written Language¶

• Human (fast typist on computer):
• ~4-5 tokens per second

• Limited by physical typing speed and thought composition

• Human (average texter on phone):

• ~0.5-1 tokens per second
• Based on average texting speed of 35-40 words per minute

• Assumes average word length of 1.5 tokens

• Large Language Model (e.g., GPT-3):

• ~60-100 tokens per second
• Limited by computational power, not by "thinking" or composition time

Spoken Language¶

• Human (fast speaker):
• ~7-9 tokens per second
• Based on ~180-220 words per minute for fast speakers

• Assumes average word length of 1.5 tokens

• Large Language Model:

• Same as written output (~60-100 tokens per second)
• LMs don't distinguish between "spoken" and "written" output

Key Differences¶

1. Speed Hierarchy: LM > Human Speaking > Human Typing > Human Texting
2. Consistency: LMs maintain speed, humans may vary or fatigue
3. Composition: Humans actively think and compose; LMs generate based on patterns
4. Modality: Humans have varying speeds for different modalities; LMs are consistent
5. Quality: Human output often more thoughtful, but LMs can produce coherent text rapidly
6. Device Impact: Phone texting significantly slower than computer typing for humans

Note: All figures are approximate and can vary based on individual skill, specific LM model, and content complexity.