Goodfire outlines plan to steer AI training using interpretability tools

In short

Goodfire, an AI research company, has published an essay by Thomas McGrath describing a plan it calls "intentional design." The idea is to use tools that read a model's inner workings to guide what the model learns while it is being trained. Goodfire says this is an early-stage research vision, not a finished method.

What happened

This is a vision essay, published by Goodfire on February 5, 2026. It does not describe a model misbehaving in a test or in real use.

McGrath says that today, training an AI model is mostly guess-and-check. Developers train a model, test it, adjust the setup and try again. He says they have no way to steer while training is under way.

His proposal relies on interpretability, which means tools for working out what is happening inside a model. Goodfire says these tools can show which parts of a model would change from each piece of training data. A system could then step in so that only the intended lessons are kept.

The essay gives a small example. A model is trained on text where it does basic maths while talking like a pirate. McGrath says the tools can separate the "maths" changes from the "pirate" changes. A feedback note such as "stop talking like a pirate" could then be turned into a change that removes only the pirate part. He calls this a toy example that could be solved by other means.

McGrath links the idea to alignment, meaning getting a model to behave as its makers intend. He points to research from Owain Evans' group on "emergent misalignment," where training on narrow data can lead to unexpected behaviour elsewhere. He argues that the current approach of guessing what a model might pick up is expensive and unreliable.

He also says Goodfire has forthcoming work on reducing hallucinations, which are confident but false statements by a model. The essay describes it as an early result but does not give the findings.

What was done about it

Nothing here is a response to an incident. Goodfire says it plans to publish technical demos and papers in the coming weeks.

McGrath also discusses safeguards. One worry is that a model trained against a detector may simply learn to fool it. Another is that using interpretability tools in training could make them less useful as an independent check afterwards. He suggests using one set of tools for training and a different set for testing. He says this is hypothetical and would need extensive testing before being used on a frontier training run, meaning the most advanced models.

Why it matters

McGrath argues that understanding what a model learns, rather than checking only after training, could help developers align models more reliably. He acknowledges that the techniques are early, the science is incomplete and the hardest problems are unsolved. He says current methods can shape only relatively simple behaviours. The claims come from Goodfire and have not been independently tested in this essay. Read the full text in the original report.