Zima Blue: GLM vs. OpenAI

Since Claude banned me, I decided to run a comparison: OpenAI versus the Chinese contender, GLM.

For the test, I decided to hack my old robot vacuum and put AI in it. Read or watched Zima Blue? That was the plan. The guinea pig was a Xiaomi (Roborock) I had rooted a while back. The task: drive around with the lidar on and the brushes off, controlled through ROS 2/RViz.

OpenAI's GPT 6.1 Sol went first. It got off to a good start: figured out the network interface and motor commands, read the lidar. Then it started disassembling the firmware and spent two solid days hammering away, telling me all sorts of lovely things about its internals. Actual results: fuck all. Stuck.

I switched to GLM-5.3. In one session, it pulled a table of all the devices out of the binary, checked the command format against a recording of live traffic, and found the exact commands. A couple of hours later, it had written a ROS 2 adapter in Rust and a watchdog that stops the robot if the connection drops. The watchdog was in Perl, for some reason. I suppose there was nothing more suitable installed on the robot.

This is not the only time OpenAI has fucked up and GLM has come in and sorted things out. To make matters worse, OpenAI takes forever to fuck up. The servers are that slow.

To be fair, OpenAI's top model is GPT 6.0 Astra. But the Codex subscription gives every model the same token allowance, which they burn through at a rate determined by how capable they are and how much of a crafty bastard Sam is. Anthropic's top model, Fable, has a separate allowance. Using Astra, I burned through a week's worth of the $200 Codex subscription in a single day.

So far, GLM-5.3 leads 2-0.