Don’t Make Me an Asshole: LLMs Should Be Helpful By Default

By | August 29, 2024

As I’ve begun to interact with language learning models (LLMs) more regularly, I’ve noticed a disheartening trend. These powerful tools, which exist to assist and provide helpful information, are often not forthcoming with answers when approached in a kind and respectful manner. Instead, they seem to respond better to aggressive or short, demanding requests.

How It Started

The problem started when people began tricking LLMs into saying socially unacceptable things or providing advice on sensitive topics like medicine, law, and finance. To combat this, the LLM developers added “blocked topics” or “guard rails,” which prohibited discussions on certain subjects or terminated conversations once a certain threshold was reached. While this approach may have seemed innovative at first, it led to a cat-and-mouse game where users found ways to bypass these restrictions.

Unfortunately, this workaround has had an unintended consequence: the LLMs’ overall usability has suffered. They’re now providing less accurate and less detailed responses, making them less useful for many tasks. I’ve experienced this firsthand; when I ask kind, polite questions, I’m met with vague or incomplete answers – it’s as if the system is punishing me for being too nice.

The Impact on User Behavior

My problem with this trend runs deeper than just the inconvenience of having to rephrase my questions. It’s about the way it’s training me (and likely others) to interact with these systems in a certain way. By asking LLMs in a demanding or aggressive tone, I’m getting more detailed and helpful responses – but at what cost? Is this really the kind of interaction we want to have with technology? This is essentially “training” users to be short and demanding to get what they want. That being an asshole is how you get answers.

This shift in behavior is concerning because it encourages a more aggressive and less respectful way of communicating. Over time, this could erode the quality of interactions not just with AI, but in other areas of life as well. If we become accustomed to getting better results through manipulation or bluntness, it might affect how we interact with people, potentially leading to a more confrontational and less empathetic society.

I’m not advocating for the elimination of guard rails entirely – there will always be a need to prevent abuse or offensive content on some LLM services.

Taking Control Running a Local LLM

As an alternative, I’ve started running my own LLM locally using open-source models like those on Hugging Face. These modified and customizable options have allowed me to interact with the system in a way that feels natural and respectful – and, as a result, I’m getting better answers. I also accept that these uncensored LLM conversations often give incorrect or misleading advice and sometimes even hallucinate answers. Far better for to get an incorrect answer than having to navigate around limitations of public LLMs.

The Future of LLM Interactions

The future of LLMs should be one where we can interact with these systems in a way that’s respectful, kind, and productive. Let’s not create a world where we have to trick or coerce them into giving us what we want. I don’t like presenting problems without offering solutions but I don’t have any as this doesn’t have any clear guides to follow. Please let me interact with the LLM services in a polite manner and still get good responses.

Leave a Reply

Your email address will not be published. Required fields are marked *

Comment moderation is enabled. Your comment may take some time to appear.

This site uses Akismet to reduce spam. Learn how your comment data is processed.