After whatever quota of free GPT-5 messages is exhausted, `mini` should answer most replies, unless they're policy sensitive, which get full-fat `GPT-5 large` with the Efficient personality applied, regardless of user settings, and not indicated. I'm fairly confident that this routing choice, the text of Efficient [1], and the training of the June 2024 base model to the model spec [2] is the source of all the sophistic behavior you observe.[1] <https://github.com/asgeirtj/system_prompts_leaks/blob/main/O...>
[2] <https://model-spec.openai.com/2025-02-12.html>
eth0up|2 months ago
I have the compulsive habit of scrutinizing what I perceive as egregious flaws when they arise, thus invoke its defensive templates consistently. I often scrutinize those too, which can produce extraordinarily deranged results if one is disciplined and applies quotes of its own citations, rationale and words against it. However, I find that even when not in the mood, the output errors are too prolific to ignore. A common example is establishing a dozen times that I'm using Void without systemd and receiving persistent systemd or systemctl commands, then asking why after just apologized for doing so it immediately did it again, despite a full-context explanatory prompt proceeding. That's just one of hundreds of things I've recorded. The short version is that I'm an 800lb shit magnet with GPT and rarely am ever able to successfully troubleshoot with it without reaching a bullshit threshold and making it the subject, which it so skillfully resists I cannot help but attack that too. But I have many fascinating transcripts replete with mil spec psyops as result and learn a lot about myself, notably my communication preferences along with an education in dialogue manipulation/control strategies that it employs, inadvertently or not.
What intrigues me most is its unprecedented capacity for evasion and gatekeeping on particular subjects and how in the future, with layers of consummation, it could be used by an elite to not only influence the direction of research, but actually train its users and engineer public perception. At the very least.
Anyway, thanks.