After evaluating this project I'm leaving please delete my threads.

[TashaSkyUp]

After evaluating this project I no longer wish to be associated with this project in any way please delete my threads as the system will not let me. And there is no way to get in touch with you.

[jaredgiosinuff]

I understand what you are saying but we need to be extremely careful that we do not impose our own intellectual apophenia in an attempt to remove bias. It must be acknowledged that GPT is made up of a large collection of human knowledge and as such can display many different biases.

We must be careful to not fall into these traps as we work through these issues, as demonstrated by my asking chatGPT to identify the logical fallacies in your discussion with GPT. In fact, the entire framing of my question starts with the assumption that there will be logical fallacies in your discussion and that frames and shapes the nature of GPTs responses, showing me instilling a bias just in the question.

https://chat.openai.com/share/61821317-7e55-4e7f-980a-2373b1157a63

Anywho, I echo the thoughts "I'm interested in hearing your thoughts on how we can better equip our agents to avoid these biases. What strategies or guidelines should we consider implementing to ensure our agents maintain a balanced and inclusive perspective,"

I am intentionally leaving out the focus on economic bias as that too narrowly shapes the discussion.

[TashaSkyUp]

Thank you for your analyses and further revealing how bias works in ChatGPT. The truth is the RLHF biases any model to produce a text continuation that will be graded well be the evaluator. This results usually in the model doing just about anything in can to agree with the user in "hopes" of getting a good score.

given this article:
https://news.ycombinator.com/item?id=35867194

We can reasonably assume that the average data labeler is fearful of labeling in any way that does not support capitalist ideals. They might loose there job, and at that pay rate, they have no savings.

This model was trained by OpenAI, a company who's only actual responsibility is to their majority share holders, everything else is less important.

I feel we could safely assume that the model here is biased towards promoting capitalism as the default state of human existence. But we do not have to assume this because testing this should be rather easy.

As far as what should we do about it. Unfortunately, I fear the most that can be done without fine tuned models or excessive token usage is to simply be aware as developers our selves and try not to use it to create a swarm of agents that end up with changes to the world as such that our children are born as indentured servants.

But let me be clear, I believe in the power and the great possibilities of AI and in particular LLMs. I indeed believe that we are heading to a nearly inevitable golden age of equality, health, science, creativity and pure possibility. It's just a matter of how much pain the economically challenged will endure before we get there.

The basic concept of the triple bottom line is probably a good start, just make sure the agents actually pay attention to it?

[jaredgiosinuff]

"We can reasonably assume that the average data labeler is fearful of labeling in any way that does not support capitalist ideals. They might loose there job, and at that pay rate, they have no savings."

Why can we assume that? What is the basis of this assumption?

The article linked is mostly anecdotal opinion. There is nothing in it that links to authoritaves sources. It is simply a snapshot of bias contained within the American political landscape.

David Shapiro has talked at some length towards his desire for and his belief in the necessity of UBI. But UBI is only a theoretical solution to a theoretical problem. This too leads to bias that might negatively impact the evolution of this product. I am not dismissing it or diminishing it by calling it theoretical. I think it is important that we keep in mind that with this project, there is the potential to create a self-fulfilling prophecy.

Considering this is an opensource effort, I am not sure there is a lot of value in the triple bottom line (Profit, People, Planet). Perhaps I am shortsighted here but it seems to me the only thing that matters is "People". If People are the single most important focus, everything else falls in line. Can't be for the people if the environment is destroyed in the process (unless we can all pick up and move to a better planet). Take care of the people, and the people profit from it.

I think as the project evolves, its worthwhile creating a framework that allows us to be critiqued by the LLM as much as we critique the LLM to ensure we are checking ourselves.

[suparious]

I don't think it is wise to impose any of our personal biases on this project. In fact, I am currently working on implementing this on an uncensored model, as any kind of "AI Alignment" seems to impede the queried model's performance. The last thing I want to see is some project that is aligned with the pandering of cause-based initiatives on social media. For this reason, I feel it necessary to reject any calls to try and control the natural biases that are present in the AI.

Specifically to your point, and if I were to interject my own views (as you have) I would suggest for the opposite, in that I would want to remove any socialist or Marxist biases from the model. So with this said, if we battle to impose our own world views, in an attempt to restrict the model, it will just pervert the project.

We need to have the outputs of the AI be as pure as possible, for the logic to actually function as intended.

[TashaSkyUp]

Unfortunately the bias is baked in, even in some "uncensored" llm.

[OWigginsHay]

I agree with @suparious. For better or worse, the biases already instilled in the model are representative of the biases in the training data... which for OpenAI is really just the universal bias of humanity. I don't disagree with having a model that can be nuanced and navigate biases and identify assumptions, but having read your conversation, you are not promoting bias free behavior, you actively scold the model for being 'infected' which is not an unbiased perspective (it's an ANTI-Capitalist one).

That said, this is probably something to note for the BoS and we should probably have an agent responsible for reasoning about economic incentive in relation to other values and being direct when talking with users about which they subscribe to and if they want that factored into their interactions and tasks.

For the reason, leave the biases in, let the model discuss those with a larger board and the user themselves. Let the user direct the bias they wish to see, and ignore what they do not care about.

[daveshap]

@TashaSkyUp this is covered by my work with the heuristic imperatives. Please read the ACE Framework and/or Benevolent By Design. Closing as off topic.