A serious danger in the use of large language models like ChatGPT
Large Language Models, like embedded in ChatGPT, will prove to be really useful tools. I'm impressed, I didn't think there'd be much progress this fast.
However, there are unexpected challenges, many of which are explored elsewhere.
Here's one big unexplored problem, where existing chat systems intentionally misinform people.?
If a large language model is trained on reliable sources, then it's reasonably trustworthy. For?example, a chat system trained on Wikipedia, the Free Encyclopedia would be reasonably trustworthy, since they've gotten really good at identifying trustworthy and untrustworthy sources. (Yes, Wikipedia isn't perfect, much like democracy, but it's the best we got.
Maybe I should fund the development of a large language model based on Wikipedia, which would be much more trustworthy than most, but it would have to address the matter of corrections or other changes.
Big problem: the most popular chat system uses a large language model that includes some sources created specifically to misinform people.
The management of that system knows?about the problem,?but they don't?seem to be taking?action,?even in light of potential liability issues.
I'd like to find a way to address that, but don't know how to get started.
Any help out there?
Strategy for Software Teams and Startups. Ideas at Cooperation.org and LinkedTrust.us
1 年Does anyone else have examples of disinformation type sources being returned by chatgpt currently or recently?
Strategy for Software Teams and Startups. Ideas at Cooperation.org and LinkedTrust.us
1 年Craig Newmark do you have any examples of a "bad" response from chatgpt that uses disinformation sources? I believe we might be able to start with ui wrappers to get people interested in a new UI around it, while developing from open sources. I'd be interested to see if I can fix a training set of examples with a wrapper and see how well it would hold up against testing set. (A wrapper in the sense of a modified prompt using the api, that simply tells it what to prioritize) Longer term the Hugging Face/BLOOM models look quite promising and might be fast to ramp up.
Fractional CTO. Collaborate ? Deliver ? Iterate. ??
1 年Craig Newmark — If you're serious about getting more involved in the evolution of this technology, I highly recommend that you undertake a crash course. Start by reading the marvelous book, The Information, by James Gleick. You might need to trust me on that, but it's the correct place to start your journey. Then, watch all the video interviews with Rob Miles on the Computerphile channel on YouTube. https://youtu.be/gQddtTdmG_8
Hi Craig, sorry haven't checked through your comments in here, but thought you'd like to take a look at this article, in case you haven't seen it. It looks like it may be some help! From 1/17/23 on NPR: "This 22-year-old is trying to save us from ChatGPT before it changes writing forever." https://www.npr.org/sections/money/2023/01/17/1149206188/this-22-year-old-is-trying-to-save-us-from-chatgpt-before-it-changes-writing-for A friend of mine published something on Facebook by ChatGPT about Macworld Expo events and it was SO wrong and full of misinformation that it was worrying!
Founder/Director of Product Design & Strategic Markets - Deaf-Tek Studio
1 年Couldn't agree more, anything "programmed to respond" is going to have a bias to the source and content based on the individuals doing the code work, explicit or implicit, there are no un-biased perfect answers, even in this "magical" content regurgitation, the source is selected by who? the machine or the code written to direct the machine, and what if I asked for an opinion from a site that denies a living God, where does the bias step in when the question becomes "what happens when we die?", and will we see it as bias or fact, another human frailty, no brain... so let's ask the computer, after all, it has a brain... remember the Wizard of OZ, "Ignore the man behind the curtain" and that is why we have no center of focus today...