2 ms·

I think this missed a key point. If you use AI to generate content, you are still a part of the data emitted. The prompt encodes what you want to say. “Write a

I think this missed a key point. If you use AI to generate content, you are still a part of the data emitted. The prompt encodes what you want to say. “Write a new article about a bank called SVB which failed because of a run. The run was induced by a fire sale on assets resulting from rising interest rates and bad rates hedging strategies.” This will write a story about that with all the language boiler plate. However, the story wasn’t whole cloth AI. A human told it the details to provide, and it provided the best expectation of the language to convey that story. Any future AI training would reinforce the grammar and style, and incorporate the information of the prompt as well. I, for one, welcome our large language model overlords.

paxys 4y ago
How does the model know what SVB is? How does it know there was a bank run on it? What assets did it have? Why were they bad? When did the crash happen? What was the final outcome? Your prompt didn't add any of that context. It still needs to fill in all these details from, say, a bunch of news articles. But then the people who write these articles will eventually go "hey I can just use AI". So ultimately where are these core facts about our world that the machine spits out coming from?
- fnordpiglet 4y ago
  You are right, my prompt is incomplete. It is an example. As a journalist with first hand knowledge of the events I can prompt craft a complete article with my core facts. Most of what you asked is already out there, possibly all prompt crafted by a human with knowledge of the subject. If I prompt craft an article about SVB, and the LLM learns to emit language about SVB from my article, then all future prompts can utilize what it has been trained in. The origination of facts would be humans writing prompts. But the language models would still incorporate that. My point is that humans would still be responsible for processing facts of the world and entering it into computers and computers would “learn” those facts from the prompt created content. Maybe we would skip the prompt creation step and just start encoding facts directly as semantic facts and we can stop writing language to encode a fact, and just use language models to create human readable communication of fact. It doesn’t matter. Facts would still be sourced by humans, even if the language is crafted by machine.
  - galaxyLogic 4y ago
    The AI reproduces what it reads on the web. But what if the information on the web is wrong? The AI will simply reproduce it. It will not try to apply its "intelligence" to evaluate the logical consistency of the claims it reads on the internet. So it is like a parrot. A parrot does not care what it says. It can sometimes seem to be right, sometimes wrong. Who controls the sources that go into the AI? No doubt all totalitarian governments will do their best to make AI be their tool for propaganda. Therefore I think it is dangerous idea to spread the meme that AI is intelligent and therefore we should believe it. After all it has super-computer-intelligence behind it.

paxys 4y ago

How does the model know what SVB is? How does it know there was a bank run on it? What assets did it have? Why were they bad? When did the crash happen? What was the final outcome? Your prompt didn't add any of that context. It still needs to fill in all these details from, say, a bunch of news articles. But then the people who write these articles will eventually go "hey I can just use AI". So ultimately where are these core facts about our world that the machine spits out coming from?

fnordpiglet 4y ago
You are right, my prompt is incomplete. It is an example. As a journalist with first hand knowledge of the events I can prompt craft a complete article with my core facts. Most of what you asked is already out there, possibly all prompt crafted by a human with knowledge of the subject. If I prompt craft an article about SVB, and the LLM learns to emit language about SVB from my article, then all future prompts can utilize what it has been trained in. The origination of facts would be humans writing prompts. But the language models would still incorporate that. My point is that humans would still be responsible for processing facts of the world and entering it into computers and computers would “learn” those facts from the prompt created content. Maybe we would skip the prompt creation step and just start encoding facts directly as semantic facts and we can stop writing language to encode a fact, and just use language models to create human readable communication of fact. It doesn’t matter. Facts would still be sourced by humans, even if the language is crafted by machine.
- galaxyLogic 4y ago
  The AI reproduces what it reads on the web. But what if the information on the web is wrong? The AI will simply reproduce it. It will not try to apply its "intelligence" to evaluate the logical consistency of the claims it reads on the internet. So it is like a parrot. A parrot does not care what it says. It can sometimes seem to be right, sometimes wrong. Who controls the sources that go into the AI? No doubt all totalitarian governments will do their best to make AI be their tool for propaganda. Therefore I think it is dangerous idea to spread the meme that AI is intelligent and therefore we should believe it. After all it has super-computer-intelligence behind it.