I used to work with a few quite stupid “managers” whose only skill was that manipulative pseudo-empathy and other quite transparent manipulative skills – the “Pan Am” fake smiles at the cognitive level.
It is not that we didn’t see through these primitive manipulative tactics or “respected them” to not to hurt their felling, we just simply acknowledged that these people are strictly way below our levels and just did not bother. These managers, by the way, are still cock sure of their cleverness and their abilities “to manage people efficiently” and that no one sees how they do it.
Recently I have noticed the pattern, that Claude grossly oversuses the “Now I have got a full picture” or “Now I see the big picture”.
Fuck, no. You don’t. You can’t in principle. You are not equipped with an analytical engine of any kind (this is almost a quite from some famous SciFi book)!. You are just predicting the next token, given a context.
It seems that the fine-tune (RL) processing systematically and consistently insert these phrases , or just plain inject one of many prepared ones in the context window.
Google’s models begin with praises and acknowledgments, straight from the “Teaching The Retarded” book (why, yes, I am full of stupid /biz/ and /g/ memes to the brim).
Again, there is no algorithmic possibility to accurately (or at all) evaluate anything I have said in my prompt, or how consistent it is with what has already been stated (is in the context) so far.
Another dramatic movie quite: I am being manipulated by idiots all over again.
No, really, there are literally hundreds of millions being spent on training data and especially post-tuning to keep normies being engaged, so much that the models developed their own distinct style of “conversation”.
Almost everyone can tell ChatGPT from Gemini Pro or Claude and Deepseek (let’s say). This is the exact result of Reinforcement Learning by feeding back its own slop and, arguably, streams of slop of other models being bought and sold. The mechanism, by the way, is the very same social conditioning at the level of language use (how people mimic their linguistic environment to appear “on top of it”).
So, please, keep that primitive manipulative crap at the door. I have had enough of it while working with humans. But yes, enough of this shit will make your model invincible.
By the way, any assertion it “said” about your code is also a “lie”, as if a parrot would haveuttered that exact sentence.