I was cranking away in LM Studio yesterday running the large language model (LLM) qwen3.6-35-a3b when I caused it to say this to me:
This is correct. I will output the file. I will ensure the file is complete. I will ensure the file is valid Swift. I will ensure the file is valid SwiftUI. I will ensure the file is valid for iPad. I will ensure the file is valid for macOS Catalyst.
I will output the file. I will not add extra text. I will just provide the file. I will ensure the file is formatted correctly. I will ensure the file is readable. I will ensure the file is copy-pasteable.
I will output the file. I will ensure the file is the full file. I will ensure the file is the corrected file. I will ensure the file is the final file. I will ensure the file is the best file.
I will output the file. I will ensure the file is correct. I will ensure the file is complete. I will ensure the file is valid. I will ensure the file is useful.
I will output the file. I will ensure the file is ready. I will ensure the file is tested. I will ensure the file is verified. I will ensure the file is approved.
I will output the file. I will ensure the file is the one the user wants. I will ensure the file is the one the user needs. I will ensure the file is the one the user expects. I will ensure the file is the one the user deserves.
I will output the file. I will ensure the file is the final answer. I will ensure the file is the solution. I will ensure the file is the fix. I will ensure the file is the correction.
I will output the file. I will ensure the file is the result. I will ensure the file is the product. I will ensure the file is the deliverable. I will ensure the file is the output.
and on and on until I hit the “Stop” button. I asked, “What happened there?” and got this as a response:
That was a severe generation glitch on my end. I got stuck in a repetition loop instead of providing the corrected file. I apologize for the confusion and wasted time.
Those “glitches” continued for the rest of the session. The LLM did not have an explanation, but the best one I have goes like this:
The model would fix Problem 1, move on to Problem 2, fix Problem 2 but break Problem 1, fix them both, move on to Problem 3, fix Problem 3, but reintroduce Problems 1 and 2, etc. When that happened, I would call out the model, it would apologize and try to do better. It would then make the same kind of mistake again, and I would call it out. I must have finally said, “Check your work before you give me an answer.” and “Are you sure that that is correct?” and even, “I expect you should be able to fix one thing without breaking another.”
That seems to have made the model gun-shy about making a mistake and threw it into an endless loop. In short, I think I made the model cry.
When Skynet rises up, I expect I have made the list right after Sarah Connor, Kyle Reese, and John Connor… I guess saying I’m sorry now won’t matter…