# What if the inside of an AI turns out to be readable after all? New research suggests it might be, and that could make AI mistakes easier to catch.

DevFeed: [What if the inside of an AI turns out to be readable after all? New research suggests it might be, and that could make AI mistakes easier to catch.](<https://devfeed.tech/articles/what-if-the-inside-of-an-ai-turns-out-to-be-readable-after-all-new-research-suggests-it-might-be-and-that-could-make-ai-mistakes-easier-to-catch-95731.md>)

Original publisher: [Read original article](<https://natesnewsletter.substack.com/p/mechanistic-interpretability-explained>)

Author: Nate

Published: 2026-10-22T13:03:36Z

Content type: article

Language: en

Sources: [Nate's Substack](<https://devfeed.tech/sources/nate-s-substack.md>)

Topics: [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Representation learning](<https://devfeed.tech/topics/representation-learning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [anthropic](<https://devfeed.tech/topics/anthropic.md>)

Tags: [ablation](<https://devfeed.tech/tags/ablation.md>), [ai](<https://devfeed.tech/tags/ai.md>), [circuits](<https://devfeed.tech/tags/circuits.md>), [features](<https://devfeed.tech/tags/features.md>), [research](<https://devfeed.tech/tags/research.md>)

## AI overview

The article argues that some model components may be more understandable than researchers assumed, which could help investigate AI errors, while stressing that a response label is not yet an explanation. It distinguishes fixed training weights from input-dependent activations, and neurons from the features and circuits they may support. Goodfire reports 99-100% interpretability for the final layer of an improved vision model and estimates that roughly 90% of neuron response poles in Qwen3-32B have one main job; these are not measures of AI safety. In its microscope, neuron 127 responds positively to many ladybugs and negatively to toy terriers, while attribution and ablation provide different evidence about its role. The article recommends checking what was observed, what was inferred, and what was experimentally tested. Feature visualizations can suggest hypotheses; unfamiliar inputs and interventions offer stronger checks. These methods require access to internal activity, and the broader claim remains a working hypothesis.

## Source excerpt

Weights, activations, features, and circuits, explained from the moment you hit enter to the answer you get back.