How do I stop gpt 5.6 sol from turning “yieldTime”/“yield_time” to “yield-time” automatically?
This affect both in chatgpt web and codex.
My prompt:
says
yieldTime
yield-time
yield_time
Welcome to the community!
Can confirm: 5.6 sol medium
neat find lol. It’s the new seahorse looks like.
From experience I’d say this is not something that can be fixed any time soon within this model. I would propose using some other key, and if you really need yieldTime or yield_time, to rewrite it with a script.Unfortunately.
Edit: nevermind, spoke too soon. Try this:
it looks like the model simply can’t read. it seems you have to specify camelCase or snake_case.
yup
yieldTime
yield-time
yield_time
yieldTime
yield-time
yield_time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
yieldTime
yield-time
yield-time
Looks like this one’s overcooked
It is worse than that: In ChatGPT the AI seems to observe the prior output incorrectly. I gave it a programming task where it was to leave both symbols defined at the end. Then it borks about its own code that was correct:
Then after an automatic retry at writing code - the AI again thinks what it wrote is wrong:
Then it does write in a retry an alternate variable name, and then we see hyphens as reported: the snake-case likely desired is wrong and we see a bad symbol instead of the instructed underscore:
Then even more AI frustration and self-correction at what it has output, and it can’t even report on the problem correctly:
In all, four tries at writing a correct code block and then further tune-ups that would have been about 15 lines.
There is nothing particularly odd about the encoding as input, as [“yield”, “_time”] is o200k [68396, 6425] - but OpenAI has disallowed logprobs with reasoning to see what is really going on with the AI’s production of output that might be a different sequence. I might try reasoning:“none” and capture a correct and wrong output.
Conclusion: there is either some serious blocking in embeddings layers preventing correct observation, or more likely, a backend fault in autoregressive tool or channel processing that treats “yield” as special.