A different approach to interpretability -- from the outside in.
Metaphori draws on Cognitive Linguistics, Psycholinguistics, and Cognitive Neuroscience to study how semantic structure shapes attention and meaning construction in transformer models... and in humans. We focus on how Language operates both inside and outside transformer models, and across disciplines, cultures, and technologies.
"Metaphori took a 200K token context max, reduced it by 94% using Metaphori Engine™ and not only did the conversation stay on topic, detailed and nuanced, but improved quality considerably, more productive, better insights, and stronger coherence."
This has enormous implications for energy usage and efficiency. The sheer volume of wasted tokens on conversations and tasks that don't achieve the results users are looking for, whether brainstorming or trying to work through an idea, or agentic tasks and coding, is unfathomable... and costing us tens of billions.
The Cost of 7 Useless Tokens: “You are an expert in [x].”
Equivalent to powering
128,892 homes
for one year · 906.9 billion wasted prompts
Energy wasted
1351.3 GWh
Electricity cost
$184.1M
EVs charged
22.6 million
CO₂ emitted
527,813 tonnes
16% role assignment · 5 follow-ups · 4,500 avg tokens · 1.5 Wh/query · 20 months
“You are an expert in [field]” — 7 tokens from a viral Prompt Engineering “hack” over 2 years ago, still being used today. It does close to nothing except cost us all. This cost compounds because every one of those tokens computes pairwise attention with every other token in the context, and that context grows.
The alternative: 94% context compression with MESN™
Structured notation eliminates the semantic noise that makes role-prompting necessary — in 6% of the tokens. Fewer tokens, cleaner attention geometry, better results.
Measuring attention head engagement across every major architecture family
We ran a cross-architecture study across 43 models to understand how Metaphori Engine™ Structured Notation (MESN™) affects attention head engagement. Using TransformerLensand nnsight to measure Direct Logit Attribution (DLA), we mapped activation patterns across 8 specialization families, from symbolic and mathematical reasoning to semantic comprehension.
MESN™ exhibits a consistent anti-pattern: higher perplexity at input, lower perplexity on output. The model finds the notation less familiar, then produces more confident completions. This inverts the standard assumption that familiar input leads to confident output.
Attention head activation across 8 specialization families
Applied research
Our findings inform a family of tools across coding, memory, architecture, and workflow domains — each applying MESN™ principles to a specific problem space.
All productsCollaborate with us
We work with research institutions and enterprises exploring how structured input shapes AI cognition, memory, and reasoning.
Partnerships