You are viewing a single comment's thread from:

RE: A talk with ChatGPT about Steem... or: A lot of text! / Ein Gespräch mit ChatGPT über Steem... oder: Eine Menge Text!

in Suggestions Club • last month

It's definitely possible at small scale. From skimming, I think this implementation creates a matrix that corresponds to the text. It actually is sort-of LLM like.

Pure speculation, but I sort of doubt whether it would be feasible for the rewards algorithm to compare every post at blockchain speeds, though. Especially if usage ever gets back to previous levels. Every post might have to look at some sort of summary of every other post in the window every 3 seconds(?). That seems intractable. Maybe a "good enough" approximation through the use of random sampling?

Or, this might be a good concept for an open source voting bot so it could be decentralized and focus on just tens or hundreds of posts per day. Stakeholders who prioritize uniqueness could invest and implement their own checks for it.

After all, the main difference between the rewards algorithm and a voting bot's algorithm is that one is centralized and the other one isn't.

Sort:  

Yes, that does seem to be one approach. However, on first reading, this does not appear to be an option for Steem. Obviously, all texts are compared against the others simultaneously. As you have already pointed out, this is not possible within the block time.
As we do not know the total number of words (across all posts) at the time of comparison, it is not possible to create a matrix or an array of the words used in equal measure. And even if we could, the array would probably be too large to store in such a way that we could access it within a reasonable time...

Maybe a "good enough" approximation through the use of random sampling?

Not if the reward curve is to be based on that, anyway.

this might be a good concept for an open source voting bot

Yes, but here too, suitable candidates for comparison need to be found first. Even if we assume, for example, that a 7-day window is sufficient for a contest, that’s still quite a lot of posts that would need to be compared.

 25 days ago 

Yes, but here too, suitable candidates for comparison need to be found first. Even if we assume, for example, that a 7-day window is sufficient for a contest, that’s still quite a lot of posts that would need to be compared.

True. It might be technically feasible if, for example, the bot looks at 6 day old posts each day and compares against 7-days worth of posts until it finds 10 to vote for. It could even cut the pool size down to a day or two. But, that would probably still require a lot of horsepower. Even if it's technically feasible, it's likely not economically feasible.

Coincidentally, a few minutes before I read this I was talking to an LLM about "surprise" and "perplexity" in language. Basically (if I understand correctly), perplexity is the degree to which tokens in a text are different from the most likely tokens. How surprising the text is.

I wonder if it's possible to consider the posts as a simple stream of text and maintain the current state of token likelihood within some window, and then calculate a perplexity value against the current expected values for each post that is published. I'm out of my depth here, but it almost seems doable - maybe also cost-prohibitive, though.

This vote is yours and deserved! Let's read, comment and connect.
❤️🍀

Boomin Three vote by wakeupkitty.jpg

This vote is yours and deserved! Let's read, comment and connect.
❤️🍀

Boomin Three vote by wakeupkitty.jpg