Communication is Hard
- Imagine a picture of an orange square, red circle, and yellow triangle
- Imagine the yellow triangle is above the orange square
- Now ask someone to draw what you're thinking
Communication is Hard
- You may get something like this:
- Did that match what you were thinking?
How Humans Communicate
graph LR
A[Person A
Blurry Intent] --> B[Medium
Words/Voice/Text/Art] B --> C[Person B
Reconstructed Intent] style A fill:#ffab40,stroke:#fff,color:#000 style C fill:#64b5f6,stroke:#fff,color:#000 style B fill:#81c784,stroke:#fff,color:#000
Blurry Intent] --> B[Medium
Words/Voice/Text/Art] B --> C[Person B
Reconstructed Intent] style A fill:#ffab40,stroke:#fff,color:#000 style C fill:#64b5f6,stroke:#fff,color:#000 style B fill:#81c784,stroke:#fff,color:#000
- Intent starts blurry and amorphous in our minds
- We use media to make it concrete and shareable
- Intent is reconstructed (differently) in the receiver's mind
Intent to Effort Speed Limit
- Higher quality media = more intent
- Lower quality media = less intent
- The receiver can correlate degree of intent to media quality
- We unconsciously have relied on this limit forever
LLMs Break Intent to Effort Speed Limit
- LLMs can compress and decompress intent
- LLMs generate high quality media
- LLMs muddle intent more than any other medium
- More in depth treatment: Large AI models are cultural and social technologies
LLMs Muddle Communication
graph LR
A[Person A
Blurry Intent] --> B[LLM
Translation] B --> C[Enhanced Media
Images/Videos/Text] C --> D[Person B
Muddled Intent] E[LLM Bias
& Interpretation] --> C style A fill:#ffab40,stroke:#fff,color:#000 style D fill:#f44336,stroke:#fff,color:#fff style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style E fill:#ff5722,stroke:#fff,color:#fff
Blurry Intent] --> B[LLM
Translation] B --> C[Enhanced Media
Images/Videos/Text] C --> D[Person B
Muddled Intent] E[LLM Bias
& Interpretation] --> C style A fill:#ffab40,stroke:#fff,color:#000 style D fill:#f44336,stroke:#fff,color:#fff style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style E fill:#ff5722,stroke:#fff,color:#fff
- LLMs drastically reduce communication effort
- Enable rich media creation from simple prompts
- Break the relationship between effort and intent
- Recipients can't distinguish human from AI intent
This is Completely Disorienting
- Human communication evolved around effort-based trust
- We can no longer gauge human effort behind communication
- We prize connection, but now swim in a totally different media landscape
- Some potential futures: human-only media, physical connection
How We Communicate with Computers
graph LR
A[Human Intent] --> B[Programming Language
Rigid Grammar] B --> C[Computer Program
Translated Intent] D[Years of Training] --> B style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#f44336,stroke:#fff,color:#fff style C fill:#64b5f6,stroke:#fff,color:#000 style D fill:#ff9800,stroke:#fff,color:#000
Rigid Grammar] B --> C[Computer Program
Translated Intent] D[Years of Training] --> B style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#f44336,stroke:#fff,color:#fff style C fill:#64b5f6,stroke:#fff,color:#000 style D fill:#ff9800,stroke:#fff,color:#000
- Programming is deterministic, with clear grammar
- Training needed to translate intent to coded intent
- Quality of translation based on experience
Computing Relied on Intent to Effort Speed Limit
- Traditional safety came from inherent human limits
- Code took time, attention, and peer review
- Human speed and attention provided natural safety bounds
- If it compiled, you probably meant it
Coding Agents Break This Speed Limit
graph LR
A[Human Intent] --> B[LLM
Translation] B --> C[Computer Program
Muddled Intent] D[LLM Bias
& Interpretation] --> C style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style D fill:#64b5f6,stroke:#fff,color:#000
Translation] B --> C[Computer Program
Muddled Intent] D[LLM Bias
& Interpretation] --> C style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style D fill:#64b5f6,stroke:#fff,color:#000
- Dramatically reduced barrier between intent and execution
- LLMs muddle intent and speed no longer a bottleneck
Shape Task Redux
- I asked the claude to draw the shapes 5 times with written intent
- Essence is there, but so is a bunch of other stuff
- Variation high = intent not well translated
No Speed Limit is Dangerous
- Small intent can unfurl into massive programs
- LLMs fill in gaps in our intent
- Like super-resolving images: looks right, but patches may be off
- Impact not intent -- and these things have impact
- BUT! Programs have a unique advantage
Programs are Quickly Verifiable
graph LR
A[Human Intent] --> B[LLM
Translation] B --> C[Computer Program
Muddled Intent] D[LLM Bias
& Interpretation] --> C C --> E[Verification
& Feedback Loop] E --> A style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style D fill:#64b5f6,stroke:#fff,color:#000
Translation] B --> C[Computer Program
Muddled Intent] D[LLM Bias
& Interpretation] --> C C --> E[Verification
& Feedback Loop] E --> A style A fill:#ffab40,stroke:#fff,color:#000 style B fill:#9c27b0,stroke:#fff,color:#fff style C fill:#81c784,stroke:#fff,color:#000 style D fill:#64b5f6,stroke:#fff,color:#000
- Verifications allow us to refine our intent
- If we get the cycle time down, we can iterate quickly
- The hardest part is expressing and recording our refinement
The Siren's Call
- Speed of intent-to-reality is incredibly alluring. This is the stuff of science fiction
- How do we maintain quality and safety?
- Programming language that captures true intent
- Preserve safety, reproducibility, and clarity
- Refinement as the core workflow
Intentc: An Experimental Approach
- Project exploring structured intent specification
- Language for clearly specifying what you want
- Maintains reproducibility and collaboration
- Bridges human intent and machine execution
Future Directions
- I am looking for 1 or 2 people to incubate this idea over time, talk to me if you're interested
- Create a first class grammar for intent
- Intuitive Refinement Loop
- Collaborative intent from multiple people / disciplines
- Remote execution and tracking built targets