Why the input problem is actually an output problem — and a general pattern from biology that explains it.
Human thought operates at ~10 bits per second. Typing operates at ~10 bits per second. The keyboard is not the bottleneck.
The real bottleneck: you're manually decompressing your compressed intent into explicit step-by-step instructions for a computer that has no context about your world.
This repo documents a pattern found across neuroscience, biochemistry, ecology, and information theory — Contextual Decompression — where a low-bandwidth signal triggers high-bandwidth coordinated action because the receiver carries the complexity, not the signal.
Low-bandwidth intent signal
↓
Receiver with rich pre-built context
↓
High-bandwidth coordinated action
The pattern appears in motor cortex, hormonal cascades, ant colony stigmergy, sparse neural coding, gene expression — and increasingly, in AI-powered interfaces like Copilot.
Contextual Decompression: Why the Input Problem Is Actually an Output Problem
- Conscious thought runs at ~10 bits/sec (Zheng & Meister, Caltech, 2024). Typing roughly matches this. The keyboard isn't the bottleneck — the lack of context on the receiving end is.
- Biology solved this billions of years ago: one adrenaline molecule triggers 100 million cellular responses. "Reach for cup" decompresses into thousands of muscle commands. The intelligence is in the receiver.
- Five properties predict whether an interface achieves high intent amplification: hierarchical context, adaptive learning, lossy-but-correctable execution, bidirectional coupling, and trigger-based signaling.
- The pattern correctly predicts why voice assistants fail (shallow context), why spreadsheets succeed (the grid IS shared context), why Copilot is transformative (deep learned context), and why BCIs deliver only modest gains (faster channel, same dumb receiver).
- The history of HCI is a gradual shift of decompression labor from human to machine: punch cards → CLI → GUI → autocomplete → AI completion → LLM agents.
The paper now includes Part 7, an empirical test of the pattern using AI as the receiver. A 3×3 experiment (prompt length × connection density, 9 conditions, blind evaluation) found that connection density correlates with output quality at r = 0.953, while prompt length correlates at r = 0.058. A 13-word high-density prompt outperformed an 86-word low-density prompt by 2.15×.
The most striking finding: surprise (non-obvious ideas in the output) correlated with density at r = 0.961. Cross-domain conceptual triggers force the model to compute structural parallels between distant knowledge regions — and the novel ideas live at those intersections. We call this the "Hamming Door Effect," after Richard Hamming's observation that open doors at Bell Labs produced more important work than closed ones.
Full experiment data: notes/06-connection-space-experiment.md
The notes/ directory contains the raw phase-by-phase research notes, including the information theory analysis, biological systems survey, HCI survey, pattern synthesis, validation tests, and the connection space experiment.
This is a living document. If you see connections to other fields, have counterexamples that break the pattern, or want to formalize the math further — open an issue or PR.
Areas especially welcome:
- Formal information-theoretic treatment (the I(X:M|Z) formulation needs rigor)
- More biological examples or counterexamples
- Quantitative measurement of "amplification ratios" in real interfaces
- Connections to active inference, free energy principle, or predictive processing
- Historical examples from other engineering domains (telecommunications, control theory)
If you find this useful in your own work:
@misc{contextual-decompression-2026,
title={Contextual Decompression: Why the Input Problem Is Actually an Output Problem},
author={Tim Welch and Claude (Anthropic)},
year={2026},
url={https://github.com/tijwelch/contextual-decompression}
}
This research started from a simple observation: "I want cake" appears instantly in your head, but telling a computer to get you cake takes forever. The hypothesis was that there's a clean, general pattern — maybe found in nature — that could fix this. Turns out there is, and it's been hiding in plain sight for billions of years.
Research conducted collaboratively between a human and an AI in March 2026.