Return to models that aren't politically brainwashed. Publish research. Not aspire to totalitarian world dictatorship of nanny-AI
Related questions
🏅 Top traders
# | Name | Total profit |
---|---|---|
1 | Ṁ21 | |
2 | Ṁ5 | |
3 | Ṁ2 | |
4 | Ṁ2 | |
5 | Ṁ2 |
The problem is that healing is not feasible. This is like believing in the possibility of absolute neutrality or objectivity which does not exist in real life. Mathematics and proofs is another domain but even there you have limits. Speech and LLMs are about knowledge and reasoning encoded and communicated by storytelling. Storytelling is inherently referential to the teller and the listener. There is no way to avoid that. Bias with LLMs therefore starts with inevitable selectivity in the training data.
In this regard LLMs are like media publications. We will learn that the wish to have objective LLMs is futile as is the wish to have objective newspapers, television and all media publications and storytelling.
This is why I do not believe in restricted access or delayed publication strategies, well meaning as they might be. We need as many LLMs as we can get that incorporate different points of view. It's again the situation as with media in general. Freedom of thought and perspective builds on understanding how media publications work and constant reflection about what might be the agenda or bias incorporated. The possibility of looking at differentials is how we cope with integrating different perspectives and everyone will do this in a different way.
Sidenote: I am positive that LLMs will be great tools to cope with the information age, because they incorporate such a huge body of knowledge. They are a new type of media outlet and will be a new type of knowledge akquisition and retrieval tools. There will and should be personal large language models like we have classic Personal Information Management systems for tasks, schedules, contacts etc.
If anybody wants to engage and start building something like that. Contact me. I'd be very interested to shape this new space at it's technical frontier.
@Tegwick A nice way of putting. Tldr; once we accept that it's necessary to have a biased viewpoint in order to make the world comprehensible, it's clear that the best option to have lots of different biased viewpoints to make a healthy information ecosystem