Blog
15 posts
Uncertain Results: the search for a confusion direction
Qwen represents the emotional state of the text it reads, and steering along those states changes how it answers.
FeaturedQuantisation and surrogate probes
Quantisation does not affect probe performance down to 4 bits. Below that, it depends...
FeaturedInterpretability and Efficiency
By reading a model's thoughts, interpretability lets us extract useful work from the cheap, fast prefill stage of inference rather than waiting on slow, expensive decode.
Routing Across the Model Frontier
The best model for a request depends on the request. How Telluvian's XPerf and XCost estimate quality and cost before each turn of a conversation.
From Predicting Model Performance to Better Model Routing
A research log on query-aware expected performance: predicting how well each model will answer a given prompt. Available in Chinese and English.
Introducing Gallery
Send your request and we find the model that will answer it best at the lowest cost.
From Attention to Evidence: Research on Query-Aware Prompt Compression
Query-aware prompt compression research log. Available in Chinese and English.
Less is
morethe sameQuantisation costs our probes nothing, and cuts GPU time by 40% on Blackwell cards.
Is Your AI Hallucinating?
How our white-box proxy model gives you a per-token hallucination score, including exactly what it costs.
Manifold Destiny
Everything the model knows about a token is packed into the residual stream, and locally the model turns out to be using less than 1% of the stream for a single prompt.
The Two-Neighbour Method: How to Measure a Dataset’s “True” Dimensions
How many intrinsic dimensions does your data really have? An introduction to twoNN and what it reveals about a transformer's residual stream.
What is Mechanistic Interpretability?
Modern large language models are immense black boxes, yet they still represent internal features in a standard way we can begin to understand.
How We Detect AI Hallucinations in Real-Time
Learn about the technical approach behind Telluvian's real-time hallucination detection, including probing classifiers, activation analysis, and token-level confidence scoring.
AI 2030: Our Generation’s Platform for Global AI Action
Encode Justice’s platform for global AI action by 2030: five calls to world leaders, co-written by Lysander Mawby.
FeaturedAI Safety Needs Startups
Why the best way to deploy safety at scale might be to sell it.