Research memo

A language model that has never read a single sentence.

We are training a large language model on one input class only: raw video of human crowds moving through public space. No web crawl, no licensed text, no transfer from an existing model. The goal is a population behavior model. It also produces a ground truth motion corpus for humanoid and logistics robots, built entirely from what cameras observe.

FRAME_T FRAME_T+15S
3,057
Live count, Times Sq node
 100%
Video-only training input
 0
External text or web sources
 1
Live node, N more planned
 06
Funding mechanism, section 6
01The problem

Every existing model inherits someone's edit

Population-scale intelligence today is downstream of an editor. News, search indices, social platforms, and government releases all pass through a filtering, framing, or reporting layer before a model ever sees them.

That layer is useful for language. It is a liability for a model whose job is to describe what a population is actually doing in physical space. A model trained on how an event was covered is not the same as a model trained on how a crowd moved. We believe that second dataset has never really been built at scale. Building it is the point of this research.

02The method

One input class, deliberately

The training corpus is built from one source type: continuous camera streams of public pedestrian activity. Nothing else enters the pipeline.

Frames are pulled at fixed intervals, pedestrians are detected and counted, and every observation is timestamped and stored. No interpretation is added beyond what the video itself shows. The corpus is the record of movement, not a description of it.

Web crawl data News or media text Social platforms Search indices Government or census feeds Any existing LLM's weights Synthetic or simulated data
01

No language input, of any kind. The corpus contains no text, transcripts, captions, or metadata sourced from language.

02

No transfer learning. The model is not fine-tuned or distilled from GPT, Gemini, Claude, or any other pretrained system.

03

Multi-source verification. Where possible, multiple camera angles cross-reference a single location and counts are reconciled before they enter the corpus, to limit single-point manipulation.

03Architecture

From a public sidewalk to a training corpus

The pipeline is simple by design. Every stage is a well-understood, independently auditable component, not a black box.

Camera feed Fixed FOV, 15s pull Detection YOLOv8, person class Oracle read 20s cycle, on-chain Storage layer Counts + frames archived Training corpus Video-only, growing
FFmpeg YOLOv8n OpenCV Upstash Redis Supabase Postgres Cloudflare R2 Switchboard oracle Solana / Anchor
04Applications

Why a robotics team would want this corpus?

Simulated pedestrian environments are cheap to generate and easy to overfit to. Real, longitudinal crowd footage is neither, and that scarcity is the point.

Humanoid motion planning

How people pause, turn, yield, and re-route around each other, observed at scale rather than scripted, gives a humanoid model a reference for socially plausible movement.

Logistics pathfinding

Dense, high-traffic urban footage teaches a navigation model how to route through unpredictable human flow. This is a real crosswalk, not a warehouse aisle.

Longitudinal signal

The same location, recorded continuously across seasons and conditions, lets the model learn how a baseline crowd differs from an anomalous one. It learns this without ever being told what an anomaly is.

No institutional prior

Because the corpus has no text or curated-dataset component, its outputs aren't shaped by whatever bias already lives inside another lab's model.

05Data provenance

One live node today, a network by design

Each new location is a new camera, a new node, and a new slice of the corpus. Locations are chosen for pedestrian density and, where possible, existing third-party traffic data to cross-validate against.

Node
Status
Notes
Signal
Times Square, NYC
Live
First node. Fixed camera, continuous 20s frame capture, oracle-verified counts.
~2,800 / hr
Shibuya Crossing, Japan
Next
One of the busiest pedestrian scramble crossings in the world, with dense, well-documented crowd flow patterns.
In setup
Additional high-density nodes
Planned
Selection ranked by pedestrian density and availability of a fixed, self-owned or public camera source.
Roadmap
06Funding mechanism

Compute has to come from somewhere

Camera infrastructure, storage, and eventual training compute are funded through a token protocol licensed from Pedestrian Language Research. The funding flows in that direction, not the other way around. The research is the product, and the token is the funding rail.

The short version

Each location node can be paired with a location-anchored token whose activity helps fund that node's camera and data infrastructure. PLR holds the underlying research IP and licenses protocol deployment to a separate treasury entity, keeping research and token operations at arm's length.

Location node goes live
Token protocol licensed for that node
Proceeds fund camera + storage costs
Node output feeds the training corpus
07Structure

Who holds what?

Research

P.L.R. - Pedestrian Language Research

Pedestrian Language Research developed the protocol and holds the underlying IP, including the training methodology and corpus.

Funding

Licensing partner

A separate treasury entity licenses the protocol to mint and hold location-based tokens, keeping funding operations at arm's length from the research IP.

Governance

Holding structure

A holding and trust structure governs downstream capital flows across entities, with formal transfer pricing between the research and funding layers.

Diligence

Request the full research memo

A longer technical writeup covers the detection pipeline, data integrity safeguards, corpus growth model, and entity structure in more detail.