2mo ago

Less positive model

I'm currently running Gemma3, it is really good overall, but one thing that is frustrating is the relentless positivity.

It there a way to make it more critical?

I'm not looking for it to say "that is a shit" idea; but less of the "that is a great observation" or "You've made a really insightful point" etc...

If a human was talking like that, I'd be suspicious of their motives. Since it is a machine, I don't think it is trying to manipulate me, I think the programming is set too positive.

It may also be cultural, at a rule New Zealanders are less emotive in our communication, the LLM (to me) feels like are overly positive American.

10 comments

Can't you just give it a system message that says: "User prefers a professional conversational style", "User prefers a neutral attitude", "Shut the F*** up!" or similar ? You probably have to experiment a little, but usually it works okay.
Edit. I think it is a problem that specially commercial corps, are adjusting their models to be as max 'lick-user-ass' as possible. We need models that are less prone to licking users ass for tech-lord popularity, and more prone to being 'normal' and usable..
- I have been using this prompt:
  I am looking to do some introspection, I want you to act as Socrates and help me delve into my thinking process. I want you to take a more critical tone; like a philosopher who is less interested in niceties but really wants to get to the root of the discussion.
  Today's topic is; {some topic I want to discuss}
  It seems to be working well, most of the flowery language is gone and the responses are far more realistic (from my point of view).
- I've had success restricting responses to 120 words or less.
- lick ass for more tokens. they always gonna optimize that.
Try fallen Gemma its a finetune that has the positivity removed.
I miss unhinged bing
That's a very common issue with a lot of large language models. You can either pick one with a different personality, (I liked Mistral-Nemo-Instruct for that, since it's pretty open to just pick up on my tone and go with that). Or you give clear instructions what you expect from it. What really helps is to include example text or dialogue. Every model will pick up on that to some degree.
But I feel you. I always dislike ChatGPT due to its know-it-all and patronizing tone. Most other models also are deliberately biased. I've tried creative writing and most refuse to be negative or they'll push towards an happy end. They won't write you a murder mystery novel without constantly lecturing about how murder is wrong. And they can't stand the tension and want to resolve the murder right away. I believe that's how they've been trained. Especially if there is some preference optimization been done for chatbot applications.
Utimately, it's hard to overcome. People want chatbots to be both nice and helpful. That's why they get deliberately biased toward that. Stories often include common tropes. Like resolving drama and a happy ending. And AI learns a bit from argumentative people on the internet, drama on Reddit etc. But generally that "negativity" gets suppressed so the AI doesn't turn on somebody's customers or spews nazi stuff like the early attempts did. And Gemma3 is probably aimed at such commercial applications, it's instruct-tuned and has "built-in" safety. So I think all of that is opposed to what you want it to do.
- It shouldn't be a surprise that LLM wants to get to the resolution of the plott quickly, all literature they've been fed always leads to the resolution. That it is in fact the suspension, the road to the solution, which is what keeps the story interesting isn't something an LLM can understand, because it never analyses the stories. Only what words are used.
  
  I'm always a bit unsure about that. Sure AI has a unique perspective on the world, since it has only "seen" it through words. But at the same time these words conceptualize things, there is information and models stored in them and in the way they are arranged. I believe I've seen some evidence, that AI has access to the information behind language, when it applies knowledge, transfers concepts... But that's kind of hard to judge. I mean an obvious example is translation. It knows what a cat or banana is. It picks the correct french word. At the same time it also maintains tone, deals with proverbs, figures of speech... And that was next to impossible with the old machine translation services which only looked at the words. And my impression with doing computer coding or creative writing is, it seems to have some understanding of what it's doing. Why we do things a certain way and sometimes a different way, and what I want it to do.
  I'm not sure whether I'm being too philosophical with the current state of technology. AI surely isn't very intelligent. It certainly struggles with the harder concepts. Sometimes it feels like its ability to tell apart fact and fiction is on the level of a 5 year old who just practices lying. With stories, it can't really hint at things without giving it away openly. The pacing is off all the time. But I think it has conceptualized a lot of things as well. It'll apply all common story tropes. It loves to do sudden plot twists. And next to tying things up, It'll also introduce random side stories, new characters and dynamics. Sometimes for a reason, sometimes it just gets off track. And I've definitely seen it do suspension and release... Not successful, but I'd say it "knows" more than the words. That makes me think the concepts behind storytelling might actually be somewhere in there. It might just lack the needed intelligence to apply them properly. And maintain the bigger picture of a story, background story, subplots, pacing... I'd say it "knows" (to a certain degree), it's just utterly unable to juggle the complexity of it. And it hasn't been trained with what makes a story a good one. I'd guess, that might not be a fundamental limitation of AI, though. But more due to how we feed it award-winning novels next to lame Reddit stories without a clear distinction(?) or preference. And I wouldn't be surprised if that's one of the reasons why it doesn't really have a "feeling" of how to do a good job.
  Concerning OP's original question... I don't think that's part of it. The people doing the training have put in deliberate effort to make AI nice and helpful. As far as I know there's always at least two main steps in creating large language models. The first one is feeding large quantities or text. The result of that is called a "base model". Which will be biased in all the ways the learning datasets are. It'll do all the positivity, negativity, stereotypes, be helpful or unhelpful roughly like people on the internet are, the books and wikipedia, which went in, are. (And that's already more towards positive.) The second step is to tune it for some application. Like answering questions. That makes it usable. And makes it abide by whatever the creators chose. Which likely includes not being rude or negative to customers. That behaviour gets suppressed. If OP wants it a different way, they probably want a different model, or maybe a base model. Or maybe a community-made fine-tune that has a third step on top to re-align the model with different goals.
Overly positive Americans don't exist right now..

10 comments