4 ms·

> Demonstrating the problems of AI (what it has produced is dodgy in various ways). So you believe the LLM was prompted to produce faulty or imperfect licenses

by Topfi 3y ago

> Demonstrating the problems of AI (what it has produced is dodgy in various ways). So you believe the LLM was prompted to produce faulty or imperfect licenses on purpose with the intent of highlighting the deficiencies of the current tech, and they did not intend anyone to use these licenses? That'd actually be an interesting way of protesting, even if it ignores both how prompting can affect LLM output quality and that these models will only improve. Add to that, if one such as myself (and seemingly most people on HN) misses this part of ironic protesting and just takes them seriously (i.e. their intent is that people use these licenses), that harms their cause by making them seem unintentionally unprofessional, rather than intentionally unprofessional.

codetrotter 3y ago
> prompted to produce faulty or imperfect licenses on purpose I don’t read it that way. You could prompt it perfectly “honestly” for the following: > create a license for creative works, similar to the Creative Commons licenses, but which disallows using the content in any AI context OpenAI GPT4 tells me about some considerations, and proceeds to give me a template that I can build on, which is very very near in fact to the kind of content you see in OP link. No need to tell the AI to create a faulty or imperfect license. The LLM is perfectly able to do that on its own. And then instead of refining what it gave you, set up a basic website hosting the content. Hell, you can even prompt the LLM to make the HTML version of the document for you, if you want.
- Topfi 3y ago
  I have absolutely no doubt that you may have received such output from the prompt you provided, but I'd like to enter into the record what GPT-4-turbo via the Chat Website provided to me using your specific prompt[0], which was more generalized advice on the different aspects of how such a license could be written, rather than a license as seen on their homepage. These also, very specifically, pointed to the issues such an undertaking will inevitably have: > [...] Definitions: Clearly define what constitutes "AI context" to avoid ambiguity. This could include the use of content in machine learning, algorithm training, data scraping, AI-generated art, automated content creation, predictive modeling, etc. [...] LLMs, by their nature, are not deterministic, so getting differing outputs is no surprise. Considering, however, that, using your prompt, I got a rather differentiated output that points to the very issues that are inherent to these licenses, namely whether things such as data scraping would fall under "technologies designed to simulate human intelligence" (clearly define human intelligence and the simulation thereof, I'd argue LLMs may not fall under that at this point in time either), showcases that the creator most likely did not receive such a faulty license from the get-go, and if they did, the output was likely littered with severe warnings to employ legal advice and lean on experts rather than just post that first draft. To me, while a LLM may provide such a license draft with such casual prompting in limited cases (perhaps someone could prompt this 100 times and see how often they get a license over general advice using your prompt), the license on that website still seems more like the result of actively prompting for erroneous results to make LLMs appear incapable of such a task (or again and much worse for their cause, this is in fact a serious attempt at a license by someone inexperienced yet unwilling to contact any expert who'd explain things like search engine crawlers and the wider implications of such an undertaking). Regardless, even if these licenses were the first output the LLM provided and even if they are meant to be taken as a joke/way to start a conversation rather than something people should use for their projects in earnest, what remains is that LLMs are not static and will improve, so any point made concerning the current day's capabilities only undercuts the serious and important discussion on licensing and the future of copyright. LLMs will get better, pointing out temporary imperfections only distracts from actually important topics. So, whatever their intent and whether this license is serious or not, I feel this does not help their alleged cause. [0] https://chat.openai.com/share/bf36f182-ba20-4f66-a702-2b636bdffbdf https://chat.openai.com/share/bf36f182-ba20-4f66-a702-2b636b...
  - codetrotter 3y ago
    > I have absolutely no doubt that you may have received such output from the prompt you provided I didn’t want to spam the thread with copy-pasting the full response that GPT-4 gave me for my prompt. But for completeness, here is a link to the conversation: https://chat.openai.com/share/7295ca81-cff4-4f01-a932-35a3d6edf850 https://chat.openai.com/share/7295ca81-cff4-4f01-a932-35a3d6... Archived copy of link: https://archive.is/xmPci https://archive.is/xmPci There you can see the prompt I used (which I also included in my comment above), and the response it gave me. This was the first and only time I asked it. But as you say, LLMs are inherently using a bit of randomness. So other people using the same prompt will get different variations in the responses.

by Topfi 3y ago

codetrotter 3y ago

> prompted to produce faulty or imperfect licenses on purpose I don’t read it that way. You could prompt it perfectly “honestly” for the following: > create a license for creative works, similar to the Creative Commons licenses, but which disallows using the content in any AI context OpenAI GPT4 tells me about some considerations, and proceeds to give me a template that I can build on, which is very very near in fact to the kind of content you see in OP link. No need to tell the AI to create a faulty or imperfect license. The LLM is perfectly able to do that on its own. And then instead of refining what it gave you, set up a basic website hosting the content. Hell, you can even prompt the LLM to make the HTML version of the document for you, if you want.

Topfi 3y ago
I have absolutely no doubt that you may have received such output from the prompt you provided, but I'd like to enter into the record what GPT-4-turbo via the Chat Website provided to me using your specific prompt[0], which was more generalized advice on the different aspects of how such a license could be written, rather than a license as seen on their homepage. These also, very specifically, pointed to the issues such an undertaking will inevitably have: > [...] Definitions: Clearly define what constitutes "AI context" to avoid ambiguity. This could include the use of content in machine learning, algorithm training, data scraping, AI-generated art, automated content creation, predictive modeling, etc. [...] LLMs, by their nature, are not deterministic, so getting differing outputs is no surprise. Considering, however, that, using your prompt, I got a rather differentiated output that points to the very issues that are inherent to these licenses, namely whether things such as data scraping would fall under "technologies designed to simulate human intelligence" (clearly define human intelligence and the simulation thereof, I'd argue LLMs may not fall under that at this point in time either), showcases that the creator most likely did not receive such a faulty license from the get-go, and if they did, the output was likely littered with severe warnings to employ legal advice and lean on experts rather than just post that first draft. To me, while a LLM may provide such a license draft with such casual prompting in limited cases (perhaps someone could prompt this 100 times and see how often they get a license over general advice using your prompt), the license on that website still seems more like the result of actively prompting for erroneous results to make LLMs appear incapable of such a task (or again and much worse for their cause, this is in fact a serious attempt at a license by someone inexperienced yet unwilling to contact any expert who'd explain things like search engine crawlers and the wider implications of such an undertaking). Regardless, even if these licenses were the first output the LLM provided and even if they are meant to be taken as a joke/way to start a conversation rather than something people should use for their projects in earnest, what remains is that LLMs are not static and will improve, so any point made concerning the current day's capabilities only undercuts the serious and important discussion on licensing and the future of copyright. LLMs will get better, pointing out temporary imperfections only distracts from actually important topics. So, whatever their intent and whether this license is serious or not, I feel this does not help their alleged cause. [0] https://chat.openai.com/share/bf36f182-ba20-4f66-a702-2b636bdffbdf https://chat.openai.com/share/bf36f182-ba20-4f66-a702-2b636b...
- codetrotter 3y ago
  > I have absolutely no doubt that you may have received such output from the prompt you provided I didn’t want to spam the thread with copy-pasting the full response that GPT-4 gave me for my prompt. But for completeness, here is a link to the conversation: https://chat.openai.com/share/7295ca81-cff4-4f01-a932-35a3d6edf850 https://chat.openai.com/share/7295ca81-cff4-4f01-a932-35a3d6... Archived copy of link: https://archive.is/xmPci https://archive.is/xmPci There you can see the prompt I used (which I also included in my comment above), and the response it gave me. This was the first and only time I asked it. But as you say, LLMs are inherently using a bit of randomness. So other people using the same prompt will get different variations in the responses.