AI Still Sees Like a Toddler

Andrew Dai runs Elorian, a startup betting that the next real bottleneck in AI isn't language, it's sight. Ask a modern model what's in a photo and it'll nail it: a dog on a couch, a red Honda, a cluttered desk. Ask it to trace which cord plugs into which port, or to figure out if a robot arm will clip a shelf mid-swing, and the wheels come off fast.

Dai's point is that description and reasoning are two different skills, and AI vision systems mastered the first while barely touching the second. A toddler can point at a cat and say cat long before she can explain how gears mesh or predict where a ball will land after bouncing off a wall. Current models are stuck in that same phase: fluent at naming, clumsy at inferring structure, sequence, and cause and effect from a static image.

That gap matters more than a parlor trick failing. Engineering drawings, satellite imagery, product schematics, floor plans, the guts of a server rack, these are all visual puzzles that demand spatial logic, not just object recognition. A model that can identify a resistor but not trace the circuit it's part of isn't much use to an electrical engineer. Same problem shows up in robotics, where a machine has to understand not just what it sees but what will happen next if it moves.

Elorian's pitch, as Dai frames it, is that whoever cracks visual reasoning first unlocks a wave of automation in fields that have mostly sat outside the generative AI boom: industrial design, remote sensing, physical robotics, even architecture. Text and code got the flashy early wins because language is linear and well-labeled. Vision is messier, spatial, and full of implicit relationships that don't show up neatly in training data.

So the interesting part isn't that AI struggles with tangled cords. It's that the struggle points to a specific, fixable gap between recognition and reasoning, and whoever closes it stands to matter a lot more than the next chatbot upgrade.