Dave Winer, who developed the RSS protocol, is exploring ways to feed his archives dating back to 1994 to ChatGPT:
http://scripting.com/2023/03/29.html
The result is what he terms a "personal chapbot" that is well informed about his background and interests.
Would it be possible to use a similar strategy to feed the Snap! manual to ChatGPT to create a Snap!bot?
I don't see why not, but you know more about this sort of AI project than I do. :~)
I know nothing about AI (perhaps less than nothing), I just know some folks who do know something about it ... and consequently have probably learned just enough to have undoubtedly have picked up some significant misconceptions.
However, I have a lot of respect for Dave Winer, and if he is using it to create a personal assistant using the corpus of his work to train it, then I have to take that seriously. The prospect of having a Snap! co-pilot who could direct me to obscure footnotes on Page 6 of the manual is intriguing.
On the other hand, in a sense the Snap! forum provides the same function using real human intelligence. ... And since ChatGPT evidently has a tendency to just make things up when it doesn't know the answer, the collective (human) intelligence of the forum may be even better than artificial intelligence (at least for now).
Humans do that too. I've been known to ask a question at a Home Depot and get a very authoritative-sounding answer that would have electrocuted me if I'd believed it. :~(
I would actually really like to see a Snap!bot. It sounds like a really cool idea, and I would love to see that type of project on Snap! or Scratch!
I think it would be possible for someone who already has knowledge of creating a chat-bot outside of Snap! and of converting code to Snap! blocks, which is probably very rare because of how difficult it must be to create an advanced chat-bot, and then there's the extra challenge of figuring out how to convert the code into Snap! blocks. At least it'd be easier in Snap! since it has added extra blocks to the palettes which are more advanced, as doing this in Scratch is a whole other level of difficulty since the editor is restricted to skill levels of children ages 8 and older, so the code can't be too advanced like some Snap! projects.
Still, I'd like to see something like this on Snap!, mainly because it's a cool concept, and I'd like to know that the Snap! community is still alive, seeing as how the front page hasn't changed at all since I've started using Snap! again.
If any one has any ideas, please respond because It'd be nice to see people still talking about this topic!
If I could get a database of some forum posts posted on the forum I could feed it into the AI I'm working on.
The forums include a great deal of off-topic information (some of which I have posted). I think a good starting point would be:
Because the information presented at Snap!Con has been reviewed, the fidelity of information is higher.
P.S. I'll also share this with Professor Nguyen, to get his input.
It could be possible to train publicly released versions of GPT-2 or LLaMA on the Snap! manual and forum posts. Although I feel like the hardest part would be turning the raw data into something useful for training the language models.
How do they do that on their real data? Isn't it raw language too?
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.