The difference is that OpenAI’s competitors and open-source projects can also use fediverse posts.

That varies by subreddit, which might actually help in training LLMs to recognize the difference.

That will remove your account from public view, but will it remove it from the data they use for AI training?

If not, you’re just enhancing the value of their proprietary data.

At some point someone’s going to train an LLM on material from successful scams to autonomously generate new scams, then wire the money to server farms to run more copies of itself.

But how will we automate our trolley problems?

Would that make it a type of sapphire?

Marching up to the next non-empty key would skew the distribution—pages preceded by more empty keys would show up more often under “random”.

There’s plenty of things that turned out to be useful to me in spite of my not recognizing their names or taglines when I first encountered them—so I don’t just assume that anything I’m not already familiar with isn’t “for” me. A brief explanation for non-insiders (or even a mention of what field it’s relevant to) would have been helpful in establishing that.

Skimming through the linked paper, I noticed this:

Scaling beyond a certain point will deteriorate the compression performance since the model parameters need to be accounted for in the compressed output.

So it sounds like the model parameters needed to decompress the file are included in the file itself.

Seems awful weird to me that twitter, facebook, and Reddit have all had similar types of issues recently and resulted in dramatic user loss.

I think the “enshittification” theory is a more likely explanation.

Sure, the “consensus view of general quality” will depend on the opinions of your user base—but if that’s the source of your objection, your issue is with the user base and not vote manipulation per se.

In that situation, what function do the upvotes serve in the first place? If the potential audience already knows they’re going to read and enjoy more content from the same source, do they need to see upvotes to tell them what they already know?

(Remember that without effective permanent karma, upvotes only serve to call attention to particular posts or comments in the short term.)

If an account is upvoted because it’s posting high-quality content, we’d expect those votes to come from a variety of accounts that don’t otherwise have a tendency to vote for the same things.

Suppose you do regression analysis on voting patterns to identify the unknown parameters determining how accounts vote. These will mostly correlate with things like interests, political views, geography, etc.—and with bot groups—but the biggest parameter affecting votes will presumably correlate with a consensus view of the general quality of the content.

But accounts won’t get penalized if their votes can be predicted by this parameter: precisely because it’s the most common parameter, it can be ignored when identifying voting blocs.

There are legitimate reasons for creating a “low-usage” server to host your personal account, so you have full control over federating etc.

If we start assuming all small instances are spam by default, we’ll end up like email now—where it’s practically impossible for small sites to run their own mail servers without getting a corporate stamp of approval from Google.

Here’s an idea: adjust the weights of votes by how predictable they are.

If account A always upvotes account B, those upvotes don’t count as much—not just because A is potentially a bot, but because A’s upvotes don’t tell us anything new.

If account C upvotes a post by account B, but there was no a priori reason to expect it to based on C’s past history, that upvote is more significant.

This could take into account not just the direct interactions between two accounts, but how other accounts interact with each of them, whether they’re part of larger groups that tend to vote similarly, etc.

Is the reviewer just now encountering the internet for the first time?