BluelightAI

Illuminate and Improve Your AI

[BluelightAI @bluelightai]

Today marks the first-ever release of Cross-Layer Transcoders for Qwen3.

BluelightAI has trained CLTs for Qwen3-0.6B and 1.7B, creating an explorable set of interpretable features that capture how Qwen3 represents concepts and transforms information across its layers.

The Qwen3 Explorer allows you to examine these features directly, identify structure in the model’s representations, and use this understanding to analyze behavior, diagnose failures, and guide adaptations of Qwen3-based systems.

[BluelightAI @bluelightai]

Satya Nadella: "If it can't be observed, it can't be trusted."
Here's what observing a model's training actually reveals. After RL training, task success went from 22% to 64%. The model also started reasoning about its grader in 9% of turns ... Nobody asked it to.
https://t.co/khMcJae58U

[Jakob Hansen @_jakobhansen]

Verbalized grader awareness increased from near zero to almost 9% of reasoning turns. After training, the model was concerned about not falling for "traps" and what aspects of the task a grader would check.

[Jakob Hansen @_jakobhansen]

Here's our writeup. https://t.co/JUYyfXUQ1u

[Jakob Hansen @_jakobhansen]

I've been working on this problem for a while @bluelightai, and we have some new ways to shed light on the question. We looked at the effect MiMo-v2.6's RL environments had on behavior. Measured performance went way up, but so did a number of other things.

[Jakob Hansen @_jakobhansen]

How do you figure out what a model actually learned in RL? Obviously number goes up on rewards and evals, but models also learn lots of things you didn’t intend and aren’t measuring. This isn't new; Satya's post is just one more reminder of the gap between training and behavior.

[Gunnar Carlsson @gunnarecarlsson]

New blog post by Sachin Khanna, Nicholas Lewins, and me.

[BluelightAI @bluelightai]

The Missing Agent: Why Regulated AI Needs an Explainable Decision-Maker https://t.co/fr9mnRiBsH

[Gunnar Carlsson @gunnarecarlsson]

New blog post by @_jakobhansen

[BluelightAI @bluelightai]

Black Box to Glass Box: Audit-Ready AI

Modern AI models, including large language models, can perform a wide range of tasks with minimal adaptation. However, this flexibility can come at the cost of transparency and interpretability: why did the model give the answer it did? At BluelightAI we are working to make AI systems more transparent so that models can be deployed to critical roles while maintaining trust and observability. This is especially critical in regulated industries like banking and insurance, where frameworks such as SR 11-7 require documented proof of what drove each model decision before deployment can proceed.

In this blog post we’ll outline one approach we are developing for interpretable classification and present a case study using this approach for intent recognition.

https://t.co/S4pPlfP7UI

[BluelightAI @bluelightai]

New blog post by @gunnarecarlsson (PhD), David Fooshee (PhD), and John Carlsson (PhD)

[BluelightAI @bluelightai]

Mapping Concept Evolution in Qwen3

We often describe Large Language Models (LLMs) as "black boxes." We observe the input and the output, but the internal machinery – the billions of calculations occurring in between – remains largely opaque. We observe that the model understands concepts, but we rarely discern how it constructs them. It is vital that we understand the “how”, because it will give us better information about how to control the LLMs and AI, and also diagnose possible malfunctions, like the introduction of unacceptable biases or the production of undesirable language or modes of communication.  This kind of control will also make the adaptation of LLM technology to specific application domains, such as financial or legal documents, simpler and more direct. https://t.co/TZJ7YIF52T

[BluelightAI @bluelightai]

Mapping Concept Evolution in Qwen3

We often describe Large Language Models (LLMs) as "black boxes." We observe the input and the output, but the internal machinery – the billions of calculations occurring in between – remains largely opaque. We observe that the model understands concepts, but we rarely discern how it constructs them. It is vital that we understand the “how”, because it will give us better information about how to control the LLMs and AI, and also diagnose possible malfunctions, like the introduction of unacceptable biases or the production of undesirable language or modes of communication.  This kind of control will also make the adaptation of LLM technology to specific application domains, such as financial or legal documents, simpler and more direct. https://t.co/TZJ7YIF52T

[BluelightAI @bluelightai]

by @_jakobhansen

[BluelightAI @bluelightai]

Circuit tracing with the Qwen3 Cross-layer Transcoders: A circuit-level analysis of how Qwen3 produces specific predictions https://t.co/qNtdlIjXrR

[BluelightAI @bluelightai]

New blog post by @_jakobhansen

[BluelightAI @bluelightai]

Circuit tracing with the Qwen3 Cross-layer Transcoders: A circuit-level analysis of how Qwen3 produces specific predictions https://t.co/qNtdlIjXrR

[Gunnar Carlsson @gunnarecarlsson]

Cross Layer Transcoders for the Qwen3 LLM Family: Digging Into Interpretable Features https://t.co/yzHekGZ3BX

[Gunnar Carlsson @gunnarecarlsson]

New post on Less Wrong: Cross Layer Transcoders for the Qwen3 LLM Family https://t.co/JLqAwcLejI

[Jakob Hansen @_jakobhansen]

I'm pretty proud of this: We trained cross-layer transcoders for Qwen3 and built a dashboard for exploring the features using TDA-based graph visualizations.

[Gunnar Carlsson @gunnarecarlsson]

Today is the day.

[BluelightAI @bluelightai]

See inside Qwen3: https://t.co/KCPgOyV80T

[BluelightAI @bluelightai]

Today marks the first-ever release of Cross-Layer Transcoders for Qwen3.

BluelightAI has trained CLTs for Qwen3-0.6B and 1.7B, creating an explorable set of interpretable features that capture how Qwen3 represents concepts and transforms information across its layers.

The Qwen3 Explorer allows you to examine these features directly, identify structure in the model’s representations, and use this understanding to analyze behavior, diagnose failures, and guide adaptations of Qwen3-based systems.

[BluelightAI @bluelightai]

Today marks the first-ever release of Cross-Layer Transcoders for Qwen3.

BluelightAI has trained CLTs for Qwen3-0.6B and 1.7B, creating an explorable set of interpretable features that capture how Qwen3 represents concepts and transforms information across its layers.

The Qwen3 Explorer allows you to examine these features directly, identify structure in the model’s representations, and use this understanding to analyze behavior, diagnose failures, and guide adaptations of Qwen3-based systems.