Goodfire says the shapes inside AI models could help us understand and control them

In short

Goodfire, an AI research company, says the ideas inside AI models are often stored as curved shapes, not straight lines. In a small simulated test, the company reports that steering a model along the curve worked, while steering along a straight line produced garbled results. No model misbehaved.

What happened

Goodfire published a research post on May 7, 2026. It is the first in a series on what the company calls "neural geometry." The idea is that the numbers a model uses internally can form shapes, and those shapes reflect the structure of the real world.

The company points to other researchers' findings as examples. It says months of the year form a loop inside language models. It also says colors sit on smooth surfaces inside image models.

To show the idea, Goodfire ran a simple test in a simulation. It used "mountain car," a classic training exercise where a car in a valley must reach the top of a hill. Goodfire trained a small model to predict the next image, given the car's position and a random push left or right.

When the company plotted how the model stored the car's position, the points formed a curved, string-like shape. Nearby points on the string matched nearby car positions.

Goodfire then did what it calls "brain surgery." It changed the model's internal numbers to slide along the curve. The car moved smoothly up and down the hill in the predicted image.

The more common method is a "steering vector," a straight-line push applied to a model's internal numbers. Goodfire says taking a straight path between two positions crossed gaps where the model's output became garbled. Some straight paths landed on a different valid spot, so the car appeared to jump to a new location.

What was done about it

Nothing needed fixing, because no model behaved in an unintended way. Goodfire says it is developing new methods that respect these shapes, and that further posts in the series will cover them. It also says it has an unsupervised method, meaning one that works without human labels, for finding such shapes in real language models. It says it will describe the method in later work.

Goodfire also compares its approach with sparse autoencoders, a popular tool for breaking a model into parts. It says that tool can split one curved shape into many small pieces that look unrelated.

Why it matters

Goodfire says that if researchers understand these shapes, they could better understand and control AI models. It argues this could help them find when models will fail, fix problems, and check models with more confidence. These are the company's own views. The simple example comes from a toy simulation, not a real-world system. The source does not report results from steering a large deployed model this way.