Before we start I wanted to point out two new things:

These newsletters will begin to be archived on the LA-AI site as blogs. You can access them at https://la-ai.io/blog
We recently launched the AI Partnership Program, allowing Builders to create profiles and be discovered by companies looking for talented people to help them. Be sure to check it out at https://la-ai.io/marketplace

How AI “Sees”

Ever pointed your phone at something and asked AI to identify it? Maybe you've used ChatGPT's camera feature or Google Lens to figure out what plant is growing in your yard. It feels like magic, right? The AI instantly knows it's a fiddle-leaf fig or that your mystery bug is just a harmless stink bug. AI doesn't actually "see" anything though. Not the way you and I do, anyway.

When you look at a photo of your dog, you see... well, your dog. Floppy ears, furry tail, the way they tilt their head when confused. AI sees something completely different. It sees millions of numbers arranged in a grid.

Remember those paint-by-number kits from when we were kids? Each little section had a number that told you which color to use. That's essentially what AI works with - except instead of "1 = blue" and "2 = red," it's dealing with values like "pixel 200,350 = red:142, green:56, blue:98." Every photo is just a massive spreadsheet of color values.

So how does a spreadsheet become "that's definitely a golden retriever"? It happens in layers, kind of like how you might recognize a friend from far away.

First, the AI looks for edges and basic shapes. Where does light suddenly become dark? Where do colors shift? This is like squinting at something from a distance - you can't see details, but you notice the outline.

Then it starts combining those edges into features. A curve here plus a texture there equals "probably fur." Those parallel lines with a certain spacing? Could be a fence. The AI builds up from simple to complex, just like a kid learning to draw - first stick figures, then adding details until suddenly it's recognizable.

By the time it reaches the final layers, it's comparing what it found against millions of examples it learned from. "These patterns match what I've seen labeled as 'cat' about 94% of the time, but 'small dog' only 6%."

Since AI doesn't truly "understand" what it's looking at, it can be fooled in ways that would never trick a human. Researchers have shown you can add nearly invisible changes to an image (stuff you'd never notice) and make AI think a turtle is a rifle. Or put a specific sticker on a stop sign that makes self-driving cars think it says "Speed Limit 45." The patterns changed just enough to trigger a different match in the AI's database.

Yet in other ways, AI vision is superhuman. It can analyze X-rays and spot tumors doctors miss. It can look at infrared images and find heat leaks in your house. Farmers use drones with AI vision to detect crop diseases weeks before the human eye would notice anything wrong.

AI doesn't get tired after looking at the 500th product on an assembly line. It doesn't miss defects because it was thinking about lunch. And now, with something called "edge computing," your Ring doorbell can identify packages without even sending video to the cloud - all that number-crunching happens right there on the device.

But we're also seeing the challenges. These systems can develop biases if they're not trained on diverse enough data. Facial recognition has shown accuracy problems across different ethnic groups. Medical AI trained mostly on one demographic might not work as well for others. It's a reminder that these tools are only as good as the data we feed them.

AI vision has moved from experimental tech to everyday reality in just a few years. Your Ring doorbell knowing it's a package delivery, not a person. That plant identifier app that actually works. The security camera that can tell the difference between your cat and an intruder. It's everywhere now, quietly converting light into numbers, finding patterns, making decisions.

As this technology spreads, the question isn't whether AI will see better than humans. In many ways, it already does. The real question is how we'll use this expanding vision. Will we catch diseases earlier? Make our roads safer? Find new ways to understand our world?

One thing's certain: when you point your phone at something today and AI tells you what it is, you're watching math pretend to see, and it's getting pretty good at it.

Know someone who’d enjoy this?

Don’t miss the next issue.