2026-09-10 –, Unconference
As Large Language Models (LLMs) transition into core enterprise infrastructure, treating them as "black boxes" is no longer an acceptable engineering risk. Can we turn LLM agents into debuggable, predictable software systems?
The answer lies in an accelerating field of research made famous by an LLM obsessed with the Golden Gate Bridge: mechanistic interpretability (mechinterp). This discipline aims to reverse-engineer neural networks into human-understandable algorithms.
Aimed at data scientists and AI engineers familiar with transformer basics, this talk showcases what the current state of the art can actually achieve. Moving past theoretical toy examples, we will explore recent breakthroughs in extracting high-level semantic features, such as isolating abstract concepts like emotional subtext within model weights.
You will learn the intuitions behind how we reconstruct model features from LLM layers, how they can be actively manipulated via activation steering to change model behaviour without fine-tuning, and a realistic assessment of where current tooling hits a hard wall.
Mechanistic interpretability treats neural networks not as statistical black boxes, but as complex, reverse-engineerable source code. The ultimate goal of the field is to map the internal weights and activations of a transformer directly to discrete, human-understandable algorithms.
This session traces the mathematical and empirical evolution of the field, exploring how researchers spent recent years validating core working hypotheses, such as the Linear Representation Hypothesis (LRH) and feature superposition, within controlled toy settings before successfully scaling them to modern state-of-the-art architectures. We will deconstruct how to isolate monosemantic concepts from dense layers using techniques like sparse dictionary learning, examine Anthropic's milestones in mapping computational subnets, and evaluate how activation steering translates into a live, low-latency mechanism for model control that does not require fine-tuning.
To anchor these concepts, we will showcase how Ramp Labs deployed activation steering in practice, and examine Anthropic’s 2026 discovery of functional emotion circuits. We will then address current limitations exposed by these applications, such as the subjectivity of post-hoc automated feature labelling, before concluding with how the field is evolving beyond traditional dictionary learning toward newly proposed architectures like "Natural Language Autoencoders" designed specifically to bypass these fundamental limits.
Outline
- 0-5': Intro and LLM steering demo (Ramp Labs).
- 5-15': From polysemantic neurons to monosemantic features.
- 15-25': From dictionary learning to circuit tracing.
- 25-30': Shortcomings of dictionary learning.
- 30-35': LLM steering
- 35-40': Newest research trends (Anthropic, Nous Research) and conclusions.
Target audience
Aimed at data scientists and AI engineers. Attendees should have a basic understanding of core transformer primitives such as attention, residuals and activations. No prior interpretability background is required; necessary vector-space intuition will be built from scratch.
AI Engineer @xtream. Feel free to reach out!