4 ms·

And where the training data for your AI advice comes from?

FirmwareBurner 1y ago
From Reddit and user forums of course.
thedevilslawyer 1y ago
Factual data mostly came from sources like wikipedia, news and social forums. I see these as continuing. Opinion data - from all other SEO sites. These were low signal mostly, and AI seems to have got the general gist of how to structure content for consumption. So rather information reaching our mind through a SEO writer, it's now being AI written, which is atleast more standardized, and can be grounded more easily, and personalized to boot.
- Aldipower 1y ago
  No SEO writer reaches my mind. You're leaving some things out. I am able to collect my information wisely. But if Google suppresses organic sites, they are not all SEO, then I just won't use Google anymore. My training data question is still valid. Social forums are also organic btw..
- pmontra 1y ago
  But if nobody write SEO content anymore because questions to AI remove all human traffic to those sites, where will AI find up to date opinions? Facebook ? Instagram? Meta will block Google from scraping them or ask money. Booking? Same for Microsoft. Tripadvisor? They better have to buy it soon. Or AI will invent opinions based on the better bid. That would be an extension of the ad business and sponsored contents too.
HamsterDan 1y ago
That cat is out of the bag, don't you think? Even if the US passes a strict law protecting copyright owners from having their works used to train AI, China is not going to obey that law, so US companies will just fall behind.
- floppyd 1y ago
  This is becoming tangential, but to me someone at this point saying "ban AI company from accessing content without explicit agreement" screams virtue signaling to me, because it directly translates to "I refuse to acknowledge the reality of the situation and will just continue saying that we need to do the 'right' thing".
  - throw10920 1y ago
    This is emotionally manipulative and fallacious. That's an invalid use of "virtue signaling" that also fails to refute the argument, and intentionally conflates the positive and the normative to push an agenda. Saying "I refuse to acknowledge the reality of the situation and will just continue saying that we need to do the 'right' thing" like it's a bad thing is also implying that morals don't matter. People who don't have morals don't belong in society.
  - ipaddr 1y ago
    They didn't drop copyright on music because of napster they sued everyone and anyone. Why give up now?
  - const_cast 1y ago
    The difference between you and those other people is those other people don't care if AI fails. They have the opinion that if you do something illegal or morally wrong, you should fail. That's not a side-effect to them, that's the whole point. So when you, and other's, stroll up and say "but but if we do this then AI companies won't be able to make money!!1! And also China!" that doesn't mean anything. Because then the response is, "well yeah, illegal actions shouldn't make you money."
- Kon-Peki 1y ago
  That’s not at all how it would work. The cat is out of the bag in terms of people passing around models that are “illegal” in your scenario, though such models would disappear from places like Huggingface. Running a commercial service that touches one of those models is off the table and will be blocked, at the IP level if need be, in essentially all the countries that matter economically to Chinese “exporters”.
jasode 1y ago
>And where the training data for your AI advice comes from? It seems like the unstated assumption in that question assumes that the world totally depends on the information from small independent blogs like this thread's article. I.e. all other information sources would be derivatives of the independent blogs. There are many other sources of organic info to feed AI training. Examples: - transcripts of Youtube videos. E.g. somebody (maybe a travel agent or a well-traveled vacationer) records a video giving advice and uploads it to Youtube. Google auto-transcribes the audio and feeds the text to the training algorithm. - AI assistants used by normal people to "analyze/summarize information" can feed that same data to the AI cloud. E.g. a travel agent types out an email giving advice to a customer. That customer then submits that same email content (or the AI autoscans the customer's email inbox) to enable the customer to ask the AI assistant, "Is this travel advice good? Is there anything this travel agent overlooked?" Of course, the travel advisor would want to limit his "proprietary and valuable travel knowledge" to only his direct clients in that private email but they stop the customer from exposing it to AI assistants. The common theme is that AI engines can insert themselves in between many types of communication between people. Those are the scenarios where you can think creatively about where all the new training data will come from. If AI assistants are used as mediators in private communication, information (including "travel advice") can "leak out" into the public. Independent blogs are a good source -- but they're not the only source.
- Aldipower 1y ago
  Thanks, that wants me to use AI even less. Both examples has nothing to do with an open and free internet. Meaning I cannot trust AI at all. All those examples of data source here, also in the other replies, using mainly highly biased sources. Wikipedia (biased by a small group of mods), YouTube filtered by Google itself, pasting customer travel advice email heavily violates GDPR, social forums also funneled. If we loose organic sites, we loose freedom. Fair enough, organic sites does mean the information there is correct, but still it is open and free, so organic sites can be treated as Gaussian distributed.
  - jasode 1y ago
    >, pasting customer travel advice email heavily violates GDPR Not seeing how pasting text with no personal identity information would violate GDPR. E.g. someone sends an email saying "For your career prospects, I think you should learn Rust instead of COBOL." Copy&pasting that into AI or an AI scanning that sentence with no identity information isn't going to violate GDPR. There's no personal data to violate. (If the AI companies deliberately want to ignore privacy laws and want to secretly attach personal data to that "Rust/COBOL" sentence, then yes, that violates GDPR.) EDIT reply to : >or the AI autoscans the customer's email inbox That auto-scan scenario still doesn't require the AI to save the personal identifiers attached to text fragments. Many ways to do that without violating GDPR. Consider how today's global spam filters "auto scan" customers' incoming emails to automatically categorize some of them for the customers' "Junk folder" without any intervention or violation of GDPR.

From Reddit and user forums of course.

Factual data mostly came from sources like wikipedia, news and social forums. I see these as continuing. Opinion data - from all other SEO sites. These were low signal mostly, and AI seems to have got the general gist of how to structure content for consumption. So rather information reaching our mind through a SEO writer, it's now being AI written, which is atleast more standardized, and can be grounded more easily, and personalized to boot.

Aldipower 1y ago
No SEO writer reaches my mind. You're leaving some things out. I am able to collect my information wisely. But if Google suppresses organic sites, they are not all SEO, then I just won't use Google anymore. My training data question is still valid. Social forums are also organic btw..
pmontra 1y ago
But if nobody write SEO content anymore because questions to AI remove all human traffic to those sites, where will AI find up to date opinions? Facebook ? Instagram? Meta will block Google from scraping them or ask money. Booking? Same for Microsoft. Tripadvisor? They better have to buy it soon. Or AI will invent opinions based on the better bid. That would be an extension of the ad business and sponsored contents too.

HamsterDan 1y ago

That cat is out of the bag, don't you think? Even if the US passes a strict law protecting copyright owners from having their works used to train AI, China is not going to obey that law, so US companies will just fall behind.

floppyd 1y ago
This is becoming tangential, but to me someone at this point saying "ban AI company from accessing content without explicit agreement" screams virtue signaling to me, because it directly translates to "I refuse to acknowledge the reality of the situation and will just continue saying that we need to do the 'right' thing".
- throw10920 1y ago
  This is emotionally manipulative and fallacious. That's an invalid use of "virtue signaling" that also fails to refute the argument, and intentionally conflates the positive and the normative to push an agenda. Saying "I refuse to acknowledge the reality of the situation and will just continue saying that we need to do the 'right' thing" like it's a bad thing is also implying that morals don't matter. People who don't have morals don't belong in society.
- ipaddr 1y ago
  They didn't drop copyright on music because of napster they sued everyone and anyone. Why give up now?
- const_cast 1y ago
  The difference between you and those other people is those other people don't care if AI fails. They have the opinion that if you do something illegal or morally wrong, you should fail. That's not a side-effect to them, that's the whole point. So when you, and other's, stroll up and say "but but if we do this then AI companies won't be able to make money!!1! And also China!" that doesn't mean anything. Because then the response is, "well yeah, illegal actions shouldn't make you money."
Kon-Peki 1y ago
That’s not at all how it would work. The cat is out of the bag in terms of people passing around models that are “illegal” in your scenario, though such models would disappear from places like Huggingface. Running a commercial service that touches one of those models is off the table and will be blocked, at the IP level if need be, in essentially all the countries that matter economically to Chinese “exporters”.

jasode 1y ago

>And where the training data for your AI advice comes from? It seems like the unstated assumption in that question assumes that the world totally depends on the information from small independent blogs like this thread's article. I.e. all other information sources would be derivatives of the independent blogs. There are many other sources of organic info to feed AI training. Examples: - transcripts of Youtube videos. E.g. somebody (maybe a travel agent or a well-traveled vacationer) records a video giving advice and uploads it to Youtube. Google auto-transcribes the audio and feeds the text to the training algorithm. - AI assistants used by normal people to "analyze/summarize information" can feed that same data to the AI cloud. E.g. a travel agent types out an email giving advice to a customer. That customer then submits that same email content (or the AI autoscans the customer's email inbox) to enable the customer to ask the AI assistant, "Is this travel advice good? Is there anything this travel agent overlooked?" Of course, the travel advisor would want to limit his "proprietary and valuable travel knowledge" to only his direct clients in that private email but they stop the customer from exposing it to AI assistants. The common theme is that AI engines can insert themselves in between many types of communication between people. Those are the scenarios where you can think creatively about where all the new training data will come from. If AI assistants are used as mediators in private communication, information (including "travel advice") can "leak out" into the public. Independent blogs are a good source -- but they're not the only source.

Aldipower 1y ago
Thanks, that wants me to use AI even less. Both examples has nothing to do with an open and free internet. Meaning I cannot trust AI at all. All those examples of data source here, also in the other replies, using mainly highly biased sources. Wikipedia (biased by a small group of mods), YouTube filtered by Google itself, pasting customer travel advice email heavily violates GDPR, social forums also funneled. If we loose organic sites, we loose freedom. Fair enough, organic sites does mean the information there is correct, but still it is open and free, so organic sites can be treated as Gaussian distributed.
- jasode 1y ago
  >, pasting customer travel advice email heavily violates GDPR Not seeing how pasting text with no personal identity information would violate GDPR. E.g. someone sends an email saying "For your career prospects, I think you should learn Rust instead of COBOL." Copy&pasting that into AI or an AI scanning that sentence with no identity information isn't going to violate GDPR. There's no personal data to violate. (If the AI companies deliberately want to ignore privacy laws and want to secretly attach personal data to that "Rust/COBOL" sentence, then yes, that violates GDPR.) EDIT reply to : >or the AI autoscans the customer's email inbox That auto-scan scenario still doesn't require the AI to save the personal identifiers attached to text fragments. Many ways to do that without violating GDPR. Consider how today's global spam filters "auto scan" customers' incoming emails to automatically categorize some of them for the customers' "Junk folder" without any intervention or violation of GDPR.