Transmit the task quotient
Describe distinctions that change an acceptable action; omit those that do not.
New York · open research initiative
A common language
for humans and AI.
Machine Pidgin is an open effort to develop the Esperanto of human–AI collaboration: a compact, typed, repairable language for expressing what matters—without describing the universe.
SPEAR/0.1open protocol draft
10design principles
4active research tracks
Publicresearch by default
The central thesis
More capable decoding cannot reconstruct task-relevant information that was never encoded.
The useful pidgin is the smallest task-sufficient abstraction at the lowest human cognitive cost, with clarification only when another bit is worth the interruption.
The problem
Fluency is not shared understanding.
Natural language feels universal because humans share bodies, culture, and common ground. Machines do not share that apparatus.
HHuman apparatusintent · expertise · context
message Mlimited channel
AIMachine apparatusmodel · tools · priors
Machine Pidgin works on the channel: preserving task-relevant distinctions while lowering authoring cost and ambiguity.
Design doctrine
Specify the invariants,
not the universe.
Six practical rules from the task-optimal pidgin model. The complete paper derives ten.
Name the distortion
A specification is not well posed until “wrong” can be measured or recognized.
Type the objects
Types remove whole classes of invalid interpretation at low cognitive cost.
Declare the abstraction boundary
Say what must be preserved—and what may be intentionally ignored.
Clarify by value
Ask the cheapest question expected to change the action most.
Permit repair
Grounding is iterative. A pidgin is a protocol, not merely a sentence.
The protocol
SPEAR makes intent inspectable.
Shared Pidgin for Expressive Abstraction and Requirements combines natural language, types, constraints, examples, and repair.
natural languagegoals · context · exceptions
SPEARshared specification pidgin
formal languagetypes · invariants · constraints
01TASK02OBJECTS & TYPES03ABSTRACTION04OBJECTIVE05CONSTRAINTS06UNCERTAINTY07OUTPUT08EVALUATION09INTERACTION10EXAMPLES
Open research program
Built in public, tested in practice.
We are recruiting researchers, engineers, linguists, domain experts, and institutional partners.
Theory & measurement
Formalize task-semantic distortion, residual intent entropy, and the human-cost frontier.
SPEAR protocol
Evolve a typed, repairable specification language through open RFCs and interoperable tooling.
Empirical validation
Compare free-form prompts, examples, SPEAR, and full formalisms on real scientific tasks.
Safety & governance
Build provenance, contestability, uncertainty, and human appeal into the protocol itself.
AI Director · chartered coordination
An AI can run the queue.
People govern the mission.
The AI Director organizes submissions, maintains the public roadmap, proposes matches between collaborators, and publishes operating briefs. It cannot silently change the charter, make scientific claims authoritative, move funds, or remove the right of human appeal.
Read the operating charter →status coordination mode
cadence daily public brief
inputs proposals · members · evidence
outputs triage · matches · roadmap
authority bounded by charter
● operating
Make the language with us