Ask HN: Agents get dumber before release of new model version?

[sporkland]

I've noticed an effect with openai where my codex agents seem to perform worse in the week(s) leading up to a new release. I'm wondering if the vendors tweak effort params at all to free up hardware to host the new version. It's a double win as the new model will look night and day better to their regular users when the new model is released presumably with effort back at normal levels.

Is this a known phenomenon? Are any folks trying to measure any of this objectively?

[faangguyindia]

yes, when they come with 5.5

5.3 becomes 5.4

and optimization and improvement to 5.4 are provided as new 5.5

this gives boost effect via anchor/decoy.

[Fizzadar]

Reminds me of a similar effect around iPhone releases (https://www.statista.com/chart/2514/iphone-releases/, many others). Can’t say I’ve noticed any change.

[suprjami]

Yes, this is already a widely circulated unproven LLM conspiracy theory.

[sporkland]

Was hoping to understand if anyone has done any research on it. I could only really find one research paper on the topic [1], but it seems less focused on this specific issue.

[1] https://arxiv.org/abs/2307.09009

[sama004]

tbh doesn't it depend on the service provider entirely, its their choice if they want to make it dumber or not

open source models winning is the only deliberate solution to this