5 ms·

It's all going to go horribly wrong because people, but, yes: 1) Integration with voice assistants. Links/sources are irrelevant. 2) Models tuned against a pa

It's all going to go horribly wrong because people, but, yes: 1) Integration with voice assistants. Links/sources are irrelevant. 2) Models tuned against a particular body of work don't care if links go stale, or websites get SlashdottedHackerNewshuggedDDOSed, or etc. Links/sources are irrelevant. 3) Inbound "service requests" processed by something that can better understand the question and the available answers/solutions. Links don't matter much. 4) When "Okay what are some good websites to read more about this?" can be answered, too, bang. 5) Ever asked somebody a question and just rolled with their answer instead of demanding citations? I mean, you're doing it here. So, again, yes.

verdverm 4y ago
> 4) ... ChatGPT will already do this, when you get an answer you want to learn more about, ask "where can I learn more about this?" "can you provide links?" there are a number of prompts that work, you can even provide it with an AND clause in the first prompt
- ergonaught 4y ago
  I meant when they can do this reliably. ChatGPT will readily supply references and even summaries of things that are wholly fabricated.
  - verdverm 4y ago
    yes, the hallucination problem is problematic, but remember the pixel attacks on the first visual DNNs? They were made more resilient with GANs, which were the first hallucinators? I've been waiting for capsule networks to get a better algo, wonder if there is something similar for LLMs... I expect these LLMs to improve in the coming years, especially as they are made into products and wrapped with more tools into Q&A pipelines.
- PaulHoule 4y ago
  I've seen ChatGPT make up fake links and fake citations that look superficially real but don't really point to anything. It seems there are two kinds of people: those who see wrong results from ChatGPT and think they are wrong and those who see wrong results from ChatGPT and are amazed at how right they are. If I was going to use AI for something it might be to find the type II's and make sure they "never work in this town again" because if they can't call ChatGPT on its bullshit they probably can't call anybody on their bullshit.
  - mrtranscendence 4y ago
    Well, I mean, you can be both. I'm amazed by how right ChatGPT can be despite how often it's wrong. > I've seen ChatGPT make up fake links and fake citations that look superficially real but don't really point to anything. I tried asking it the other day for free software that would transcribe video. It gave me five pieces of software, four of which didn't exist and one of which was not free.
    - verdverm 4y ago
      I am amazed with both how right and wrong it can be. It can give impressive responses and fail spectacularly
      - PaulHoule 4y ago
        I am a little amazed at how people get amazed by it. As a software dev I am expected to get it right, not almost right. I just coded up something that works most of the time but failed in one case and the tester sent it back. In my case that defect takes somewhere between 1% and 500% of the effort to fix as the rest of that ticket but in the case of ChatGPT "close but no cigar" is not close at all in terms of having a strategy that is reliably truthful. If ChatGPT has a really "amazing" ability it is in getting people to think it is amazing. I think of that story https://en.wikipedia.org/wiki/The_Emperor%27s_New_Clothes https://en.wikipedia.org/wiki/The_Emperor%27s_New_Clothes and that it's not just cheaply dismissive to point out that ChatGPT's core competence is the same as that Emperor because not everybody can pull off what that Emperor pulled off. See also https://en.wikipedia.org/wiki/ELIZA_effect https://en.wikipedia.org/wiki/ELIZA_effect which is a anecdote from the beginning of AI research that seems forgotten today in how people react to ChatGPT.
        verdverm 4y ago
        It is not amazing all of the time, or perhaps amazing at all, but I can be amazed that LMs have reached a new threshold. I have specific examples where the interaction gave me a correct answer with rational faster than I could Google and sift through results. Something that took us weeks to figure out, ChatGPT gave us the two things we did, from just two prompts about what was going on. This was no trivial coding problem. It involved complications in a distributed build system. Sometimes it's not about getting the exact answer, but getting ideas about where to look. ChatGPT is not just for software either, it is quite useful as a co-writer. To think that ChatGPT is not impressive is to close yourself off from what is actually happening. I was there with you until I started trying it out regularly. To think that it is not providing value is to neglect reality. You do not get to judge others as inferior for finding ChatGPT useful. This is the impression you are giving. Nothing is perfect. Think of something that amazes you and then consider where it is not perfect. It is perfectly reasonable for people to find ChatGPT impressive without it being perfect. DAN and SAM are also impressive sessions people have had with ChatGPT that show somewhere it can be manipulated. TL;DR As devs, we use search because it saves us hours of time. ChatGPT is a tool you should add to your toolbox because it can also save you hours of time.
- JohnFen 4y ago
  How does it define what a "good" website is, though? It seems to me that this would be little different than relying on any other single source for information. That practice is always problematic.
  - verdverm 4y ago
    We can consider the following when thinking about how to answer this - how do current search engines rank results? - what features correlate with good? - how do features of sites get encoded, queried, and returned? - how might LMs be already involved in this process? - how might LLMs / ChatGPT like systems encode these things? - how might we enrich training data to facilitate improved LLM information retrieval? I would suspect, with current models and algos, that the more often certain statements are found across the training data, the more likely that information is to be "correct" A good site might be one that has many of these "correct" chunks. Though I doubt they are there yet. A next phase task will likely be along the lines we are discussing here. i.e. How do we return citations and references in responses
    - JohnFen 4y ago
      > how do current search engines rank results? I think there's a poor correlation between search engine rankings and website quality, so it might be better not to include this. > what features correlate with good? Right, that's the essential question. Google's original bit of genius was to to ignore that question entirely, and go with "popular" instead. You can measure popular, but "good" entirely depends on what it is you value. That makes it somewhat subjective.
      - verdverm 4y ago
        PageRank worked well at first by defining quality by the number of backlinks, or sites that linked to it. Essential if you had good content, it was likely that many others pointed to it. Ofcourse this was quickly gamed, but they were able to leverage their search engine usage to observe how users interacted with results, and later pages. I imagine that you could have a quality measurement by finding the chunk embeddings that are most popular, and then finding the sites that have these in the highest concentration. User interaction will still likely be a very important feature in the quality space.

verdverm 4y ago

> 4) ... ChatGPT will already do this, when you get an answer you want to learn more about, ask "where can I learn more about this?" "can you provide links?" there are a number of prompts that work, you can even provide it with an AND clause in the first prompt

ergonaught 4y ago
I meant when they can do this reliably. ChatGPT will readily supply references and even summaries of things that are wholly fabricated.
- verdverm 4y ago
  yes, the hallucination problem is problematic, but remember the pixel attacks on the first visual DNNs? They were made more resilient with GANs, which were the first hallucinators? I've been waiting for capsule networks to get a better algo, wonder if there is something similar for LLMs... I expect these LLMs to improve in the coming years, especially as they are made into products and wrapped with more tools into Q&A pipelines.
PaulHoule 4y ago
I've seen ChatGPT make up fake links and fake citations that look superficially real but don't really point to anything. It seems there are two kinds of people: those who see wrong results from ChatGPT and think they are wrong and those who see wrong results from ChatGPT and are amazed at how right they are. If I was going to use AI for something it might be to find the type II's and make sure they "never work in this town again" because if they can't call ChatGPT on its bullshit they probably can't call anybody on their bullshit.
- mrtranscendence 4y ago
  Well, I mean, you can be both. I'm amazed by how right ChatGPT can be despite how often it's wrong. > I've seen ChatGPT make up fake links and fake citations that look superficially real but don't really point to anything. I tried asking it the other day for free software that would transcribe video. It gave me five pieces of software, four of which didn't exist and one of which was not free.
  - verdverm 4y ago
    I am amazed with both how right and wrong it can be. It can give impressive responses and fail spectacularly
    - PaulHoule 4y ago
      I am a little amazed at how people get amazed by it. As a software dev I am expected to get it right, not almost right. I just coded up something that works most of the time but failed in one case and the tester sent it back. In my case that defect takes somewhere between 1% and 500% of the effort to fix as the rest of that ticket but in the case of ChatGPT "close but no cigar" is not close at all in terms of having a strategy that is reliably truthful. If ChatGPT has a really "amazing" ability it is in getting people to think it is amazing. I think of that story https://en.wikipedia.org/wiki/The_Emperor%27s_New_Clothes https://en.wikipedia.org/wiki/The_Emperor%27s_New_Clothes and that it's not just cheaply dismissive to point out that ChatGPT's core competence is the same as that Emperor because not everybody can pull off what that Emperor pulled off. See also https://en.wikipedia.org/wiki/ELIZA_effect https://en.wikipedia.org/wiki/ELIZA_effect which is a anecdote from the beginning of AI research that seems forgotten today in how people react to ChatGPT.
      - verdverm 4y ago
        It is not amazing all of the time, or perhaps amazing at all, but I can be amazed that LMs have reached a new threshold. I have specific examples where the interaction gave me a correct answer with rational faster than I could Google and sift through results. Something that took us weeks to figure out, ChatGPT gave us the two things we did, from just two prompts about what was going on. This was no trivial coding problem. It involved complications in a distributed build system. Sometimes it's not about getting the exact answer, but getting ideas about where to look. ChatGPT is not just for software either, it is quite useful as a co-writer. To think that ChatGPT is not impressive is to close yourself off from what is actually happening. I was there with you until I started trying it out regularly. To think that it is not providing value is to neglect reality. You do not get to judge others as inferior for finding ChatGPT useful. This is the impression you are giving. Nothing is perfect. Think of something that amazes you and then consider where it is not perfect. It is perfectly reasonable for people to find ChatGPT impressive without it being perfect. DAN and SAM are also impressive sessions people have had with ChatGPT that show somewhere it can be manipulated. TL;DR As devs, we use search because it saves us hours of time. ChatGPT is a tool you should add to your toolbox because it can also save you hours of time.
JohnFen 4y ago
How does it define what a "good" website is, though? It seems to me that this would be little different than relying on any other single source for information. That practice is always problematic.
- verdverm 4y ago
  We can consider the following when thinking about how to answer this - how do current search engines rank results? - what features correlate with good? - how do features of sites get encoded, queried, and returned? - how might LMs be already involved in this process? - how might LLMs / ChatGPT like systems encode these things? - how might we enrich training data to facilitate improved LLM information retrieval? I would suspect, with current models and algos, that the more often certain statements are found across the training data, the more likely that information is to be "correct" A good site might be one that has many of these "correct" chunks. Though I doubt they are there yet. A next phase task will likely be along the lines we are discussing here. i.e. How do we return citations and references in responses
  - JohnFen 4y ago
    > how do current search engines rank results? I think there's a poor correlation between search engine rankings and website quality, so it might be better not to include this. > what features correlate with good? Right, that's the essential question. Google's original bit of genius was to to ignore that question entirely, and go with "popular" instead. You can measure popular, but "good" entirely depends on what it is you value. That makes it somewhat subjective.
    - verdverm 4y ago
      PageRank worked well at first by defining quality by the number of backlinks, or sites that linked to it. Essential if you had good content, it was likely that many others pointed to it. Ofcourse this was quickly gamed, but they were able to leverage their search engine usage to observe how users interacted with results, and later pages. I imagine that you could have a quality measurement by finding the chunk embeddings that are most popular, and then finding the sites that have these in the highest concentration. User interaction will still likely be a very important feature in the quality space.