Re:Microsoft and OpenAI Working on ChatGPT-Powered Bing in C...
by jillesvangurp on Wednesday January 04, 2023 @06:09AM
(#34244268)
This is not Google's fault but the result of people trying to game the system to promote whatever it is they are promoting.
Before Google the web was a lot less searchable. You had to go to Yahoo, Altavista, or similar websites and wade through pages and pages of nonsense before accidentally stumbling on a good link. Very hard to find stuff back then. The effect search has had on the internet is that we can now actually find stuff pretty easily.
Of course the new thing here is that we are now mixing search and content generation. Before you had to manually create content, put it somewhere and then hope people could find their way to that content. Now people ask a question and the AI generates the content that contains the answer. AI generated content has the potential to vastly outnumber any human generated content.
Trash content, is what we have already. E.g. reddit is full of it. Lots of people are addicted to spreading their nonsense there and trying to be as offensive as they can possibly be. There's also some good content there of course. With AI there will just be more of it.
This is an issue both for training and for users. A model trained on nonsense will produce nonsensical content. Not good. So there's an incentive for separating the low value garbage from the high value content. Another issue is that you get feedback loops once we start training models on AI generated content.
In other words, there's a need to tell what is what for effective training of models as well. A good strategy is to simply figure out reputability and nature of sources of content. You could use an AI for that. Digital signatures would help too. Trusting certain websites more than others is helpful. Some websites helpfully score and moderate their content or have people liking things to indicate it is important. So, plenty of signals to work with here. That's why Google is still pretty effective. And that's why chatgpt is mostly pretty impressive and not rehashing a lot of alternate reality nonsense from some substack/reddit. But it's definitely an arms race.
The next level in this is going to be people getting more protective of their content. IMHO it's a problem that most content on the web is not signed by their authors. We have no way of doing that in a sane way currently. We have SSL certificates for domains but there's no such thing that guarantees a blob of text comes from a specific person or entity (an AI model could be signing it's generated content as well). As soon as you have this, you can start figuring out reputability and relevance to a particular question much more effectively.
Before Google the web was a lot less searchable. You had to go to Yahoo, Altavista, or similar websites and wade through pages and pages of nonsense before accidentally stumbling on a good link. Very hard to find stuff back then. The effect search has had on the internet is that we can now actually find stuff pretty easily.
Of course the new thing here is that we are now mixing search and content generation. Before you had to manually create content, put it somewhere and then hope people could find their way to that content. Now people ask a question and the AI generates the content that contains the answer. AI generated content has the potential to vastly outnumber any human generated content.
Trash content, is what we have already. E.g. reddit is full of it. Lots of people are addicted to spreading their nonsense there and trying to be as offensive as they can possibly be. There's also some good content there of course. With AI there will just be more of it.
This is an issue both for training and for users. A model trained on nonsense will produce nonsensical content. Not good. So there's an incentive for separating the low value garbage from the high value content. Another issue is that you get feedback loops once we start training models on AI generated content.
In other words, there's a need to tell what is what for effective training of models as well. A good strategy is to simply figure out reputability and nature of sources of content. You could use an AI for that. Digital signatures would help too. Trusting certain websites more than others is helpful. Some websites helpfully score and moderate their content or have people liking things to indicate it is important. So, plenty of signals to work with here. That's why Google is still pretty effective. And that's why chatgpt is mostly pretty impressive and not rehashing a lot of alternate reality nonsense from some substack/reddit. But it's definitely an arms race.
The next level in this is going to be people getting more protective of their content. IMHO it's a problem that most content on the web is not signed by their authors. We have no way of doing that in a sane way currently. We have SSL certificates for domains but there's no such thing that guarantees a blob of text comes from a specific person or entity (an AI model could be signing it's generated content as well). As soon as you have this, you can start figuring out reputability and relevance to a particular question much more effectively.