PL. I · SPECIMEN
RESEARCH YOU CAN SEE THROUGH
Glasshouse Labs trains language models in the open. Weights, data, seeds, failed runs and the unedited lab notebook are all published, so you can check our work instead of trusting it.
PL. II · LATENT SPACE
CH. 1 · WHERE THE MODEL THINKS
INSIDEPANE-7
Before a model answers, every word you type becomes a point in a space with thousands of directions. We call it latent space. Most labs never show it. We fly you through ours.
FIG. 2 — 9,000 TOKENS FROM ONE CONVERSATION
CH. 2 · COORDINATES
EVERY WORDHAS AN ADDRESS
Inside Pane-7, each token is described by 4,096 numbers. Change one and the meaning shifts. We publish the full table for every layer.
FIG. 3 — 4,096 DIMENSIONS PER TOKEN · 32 LAYERS
CH. 3 · GEOMETRY
MEANING ISA DISTANCE
"Harbour" and "port" land close together. "Harbour" and "grudge" drift apart, until the sentence says "harbour a grudge". Watching that happen is most of our job.
FIG. 4 — TWO CLUSTERS, ONE IDIOM
CH. 4 · THE ATLAS
WE PUBLISHTHE MAP
The embedding atlas is open to anyone: search a word, see its neighbours, follow it through all 32 layers. No account, no waitlist.
Open the embedding atlasPL. III · PAPERS
PAPERS, WITH THE
WORKING SHOWN
Each paper ships with its code, data manifest, seeds and the notebook pages it came from. Drag or scroll to turn the ring.
FIG. 5 — THE BACK CATALOGUE, AS A RING
INDEX · HOVER A TITLE TO INSPECT ITS ATTENTION
Fig.TitleAreaYear
- 5.1Glass Box: reading circuits in a 7B model without guessingInterpretability2026Read
- 5.2Every Weight, Every Seed: a training run anyone can repeatReproducibility2026Read
- 5.3The Lens Test: when attention looks where it shouldn'tEvaluation2026Read
- 5.4Slow Thinking at Small ScaleReasoning2025Read
- 5.5Open Bench: 41 evaluations, all public, all re-runnableEvaluation2025Read
- 5.6Refraction: how fine-tuning bends what a model believesAlignment2025Read
- 5.7Notebook Zero: our whole 2025 lab log, uneditedMethods2025Read
PRINCIPLE I
PUBLISH THEFAILURES TOO.
A result without its failed attempts is an advert. Every run we start goes in the log, including the 212 that went nowhere.
212
FAILED RUNS LOGGED
100%
EVAL CODE PUBLIC
9 d
RESULT TO PREPRINT
PRINCIPLE II
IF NOBODY CANREPRODUCE IT,WE DIDN'T FIND IT.
Seeds are fixed, hardware is logged, data is versioned. Three outside labs re-ran our 2025 headline result before we called it one.
3
OUTSIDE RE-RUNS
0.4%
MAX SCORE DRIFT
1
PAPER RETRACTED
PRINCIPLE III
SMALL ENOUGHTO SEE INSIDE.
We size models to be understood, not to top a chart. Pane-7 has 32 layers and we have mapped every one of them.
7B
PARAMETERS
32/32
LAYERS MAPPED
1 GPU
TO RUN IT AT HOME
PL. V · PEOPLE
THE PEOPLE
AT THE BENCH
Forty people. One open notebook.
Every plot on this site is signed by the person who made it. Hover a portrait to resolve it; the dots are the same halftone we use to print our figures.
■ WHO SIGNS THE PLOTS
INTERPRETABILITY
TRAINING
EVALUATIONS
NOTEBOOKS
