ed interactive group

There's a lab in San Francisco where rat brain cells sit in a dish and get shown pictures all day. That dish just made AI video five times cheaper. And none of those brain cells are in the product. They never leave the lab. The company is called The Biological Computing Co., and it was started by two neurosurgeons, Alex Ksendzovsky and Jon Pomeraniec. They grow neurons on chips covered in 4,096 tiny electrodes. Most come from rats, and some are grown from human stem cells. If you leave them alone long enough, the cells hook up to each other and turn into a living little network. Then they "show" the network pictures. Each image gets turned into a pattern of electrical zaps, and the electrodes record what the neurons do back. What the neurons do is pretty simple. A zap lights up the cells right around it, spreads a little way, and fizzles out. Real brains work like that too. Most of the wiring in your cortex is short-range, so one signal can't flood the whole thing. Now think about the last AI video you watched that ran past a few seconds. Faces start to melt and the background kind of oozes. That happens because each frame gets built off the one before it, so every tiny mistake gets copied forward and grows. TBC calls it washout. So they took what the neurons were doing and wrote it into code. It's a small add-on that plugs into the early layers of a video model and acts like a shock absorber. Nearby parts of the image stay stuck together, and little glitches die out before they can spread. In their test on a Minecraft video model, the add-on was about 156,000 parameters sitting on a model of roughly 600 million. Here's where the money comes in. A model that holds itself together can get away with fewer compute steps per frame, and fewer steps means less time renting GPUs. On the Minecraft model, that pushed it from about 2 frames a second to almost 10. The model they launched today, mischief 1, is Alibaba's open Wan 2.2 with TBC's add-on bolted on top. TBC says the add-on is less than a tenth of a percent of the model's size. So you're basically running Alibaba's model plus a sliver that a dish of rat neurons figured out. Half a cent per second of video. That's $2.50 for 100 five-second clips. If you're checking the math, the "5x faster" and "80% cheaper" TBC is quoting are the same claim, since using a fifth of the GPU time means paying 80% less. They also haven't put out independent benchmarks for mischief yet. The closest thing is an outside test on their Minecraft model by Bluesky Compute, which found inference costs about 4.4x lower. What nobody can answer yet is whether some engineer at a desk would've found the same trick eventually without the neurons. TBC is betting evolution already spent a few hundred million years working this stuff out inside a brain that runs on about 20 watts, and a dish of cells wired to electrodes is the quickest way to copy its homework. Earlier this year a dish of human neurons learned to play Doom, and people started asking when brain cells would run real software. This is probably what that looks like for the next few years. The cells stay in a lab in San Francisco getting zapped, and the notes somebody takes on them are what you end up using.