AI Doesnt Think in Words. It Thinks in Shapes @GadgetHacks
AI Doesnt Think in Words. It Thinks in Shapes  @GadgetHacks
Uploaded August 2026 | Updated November 2026, 1 day ago
AI models may not “think” in words the way their text output makes it seem.

Goodfire CEO Eric Ho explains research suggesting that neural networks can represent concepts as complex, curved geometric structures inside their latent space.

Instead of a concept existing as one simple direction, models can encode higher-dimensional structures—manifolds that capture relationships, position, movement and meaning in surprisingly rich ways.

Eric walks us through neural geometry, the linear representation hypothesis, rabbit ears, the mountain-car steering problem, shape rotators, latent representations and what all of this could reveal about how AI models actually process information.

If we can understand the geometry inside AI, we may eventually be able to understand and deliberately shape how these systems think.

Learn more about Goodfire: goodfire.ai

Sponsored by SAS AI Governance: sas.com/en_us/solutions/ai/governance.html?utm_source=other&utm_medium=cpm&utm_campaign=sas-vpl-gbc-gps-global#

More from The Neuron: theneuron.ai
AI Doesnt Think in Words. It Thinks in ShapesiOS 19 Leaks Are Wild!Top 5 E-Readers for Book LoversiOS 13 Makes Bulk-Deleting Emails in Search Easier [Demo]Check If Your Phone Supports Project Treble [How-to]Run Powerful AI On Your Own PC: No Cloud Required!Customize the Nav Bar Buttons on Your Nexus Device [How-To]This Mod Helps You Save Battery in Every Way Imaginable [How-To]Install Googles Brand New Nexus Launcher on Other Devices [How-To]Should We Pause AI Development? A Rational DebateTurn Your Nexus 6P into a Pixel [How-To]Remap Androids Hardware Buttons Without Root [How-To]
Gadget Hacks |

AI Doesn't Think in Words. It Thinks in Shapes

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER