This is why I think it is important to pay a lot of attention to the training data being selected. There are ethical and legal issues in personalizing existing content, moreover, it shouldn’t be too easy and too fast because of quality and relevance. Choosing, preparing, and optimizing the right training data can be a huge undertaking, and it can also involve thousands of risks. But deleting a large group of Internet pages is also not complicated enough in my opinion.

Almost all human communication is seriously biased, with one strong and full-on sometimes. For example, the best science pieces that build on dozens of rigorously proven prior works. Or the best journalism is nothing but truth and honest framing. Much of Wikipedia can also be attributed to this, although it is much more diverse. These are the exceptions, so there is very little human material on which to train the model which is not a problem. For pragmatic purposes anyway; Art forms such as paintings and poetry have fewer problems besides copyright. But when it comes to politics, society or people, we make some assumptions and ignore the nuances, even with a lot of experience. Everyone has their own narrative point of view based on their own worldview, upbringing, culture, speech and their own preferences, fears, etc.

Overall, I think we’re going to experience a decade or two of the most bizarre ups and downs in AI technology, and how owners, other companies, governments, and individuals react to it. We will all have to learn what can and cannot be done, what is somewhat biased, whose actions or scripts are included in the training data, how this data is explained or formatted, what is the right degree of transparency and openness, and so on. , etc. Also questions and issues that we likely haven’t explored yet. What we now think will be better or safer can also cause problems that we don’t know much about yet. This will be trial and error.

Slow development would be fine, with a lot of attention to the process and implications of AI, not just building a smarter model as quickly as possible as the end result. But the race is on now and it’s all about profit again, so other considerations will again fade away.

[Reactie gewijzigd door geert1 op 30 maart 2023 12:13]