Because you have a small batch size, my feeling is that you probably want a very small dropout rate on the important items, if only to decrease the chance the network overfits to them. Maybe 1 / 100 batches, excludes the important item and the rest include it. But perhaps it doesn't matter.

You can't self host?

Biggest issue with XetHub is current lack of open source. They have a way better storage design than DVC, but open source is absolutely necessary. If I can't give someone else zero-cost instructions to reproduce my research then my research becomes far less useful.

Perhaps industry will pay for this, but non-free software is a deal breaker for someone who wants to advance science.

I'm convinced. Brilliantly explained. Constructing the problem with respect to a bound makes a ton of sense.

Now I suppose we just code it up and throw it at CPLEX.

Isn't training a neural network NP hard? What's the polynomial time verifier?

Are you not signaling something that you value?

• You think that it's bad character to express a believe that demonstrates one's own good character or moral superiority.

• Someone expresses a belief that they believe demonstrates good character.

• You expressed an opinion and deride them by calling what they did virtue signaling.

• This serves to signal to other like-minded folks of your own good character or moral superiority.

Oh wait, does it only count if it's an opinion you disagree with? Oh ok... well you probably don't think of it like that... you wouldn't be able to feel superior and like you could call someone stupid if it was like that...

Let me sum it up for you with an example:

• Hypocritical: "Alice expresses to your Bob that people who hang pride flags are assholes because they virtue signal"
• Consistent: "Alice expresses to your Bob that people who hang pride flags are assholes because it makes her feel uncomfortable" or she thinks it sexualizes the children or some shit like that.

Say what you mean: don't hide behind that "virtue signaling is wrong" crap. If you are doing something that is morally correct, then you should be promoting it. Say why you think whatever they are "virtue signaling" is wrong.

Yeah basically, because criticizing virtue signaling IS dumb.

You can't criticize someone for expressing what they believe. That's ridiculous.

You criticize someone for the belief. You tell them why what they believe is wrong, not that the expression itself is wrong.

Note, I wasn't saying you were wrong to virtue signal. I was just telling you that criticizing it is literally self-contradictory. It doesn't leave any room open for discussion, you're basically just telling someone to shut up. While you can do that, don't trick yourself into thinking that you're saying anything of value.

The thing that pushed me over the edge was banning of accounts that promoted other social media platforms.

This is a clear violation of what Musk has expressed his view of free speech absolutism is: maximum free speech within constraints of the law and safety.

There's no safety concern here, there's no law concern here. That policy was pure censorship. Of course he has the right to do that, it's his platform, but I found it wildly hypocritical.

The value proposition of Mastodon is that it's distributed and censorship can only exist at a federated level. If you don't like it, you can move servers.

I absolutely made a mastodon account. I think Musk is off his rocker right now. It's made the weaknesses of a centralized forum clear. Mastodon isn't perfect, lots of improvements to make, but it lies on a much more solid decentralized foundation. But I'm also hedging my bets. I still have my Twitter account. Life's too complex to be absolutist about anything. Life, like computer science, is all about finding the right trade-offs.

Voicing that you're upset about virtue signaling is virtue signaling. Just want to make that hypocrisy clear.

But while I'm here, I'll acknowledge where you are correct. The authoritarian government in China is committing genocide. Not really sure what your point is.

I see we're playing this game.

I minimize my fossil fuel usage where possible. For the rest, I estimate my yearly footprint and buy carbon offsets to put myself at net zero. I currently need to only offset 10 more years to put my entire life at net zero.

I do limit my support of unethically produced products. Unfortunately to have access to the field of computing, some concessions need to be made. It's not ideal, but the trade-off is that AI research can help reduce future need for slave labor.

And for what's worth I haven't left Twitter. I've just also joined Mastodon. The vulnerabilities of the centralized nature of Twitter have been made clear recently. I think it's worth shifting to a distributed system.

So here's a question for you: just because life is full of unavoidable gray areas, does that mean you make no effort to be ethical in any area where you have significant degrees of agency?

Sounds like you can reach the right 0.01% of the audience then. The ones who still have their ethics in tact.

And for what it's worth, the applications for Zuck and Musk do have social good, even if it is a gray area. Remember, Facebook brought us torch as open source software.

Did you though? What was your method? Making observations? Did you make measurements? Did you make an effort to find your blind spots and search those as well?

You can say that you believe the anecdotal evidence you've observed is sufficient to come to a conclusion. But please, do not claim that you've designed an experiment. You (probably) haven't.

For what it's worth, I think you'd probably find that conservatives have more irrational protest behaviors than liberals, but then again, the trick is quantifying and measuring it.

That feels true, but I wouldn't trust my gut on that without solid measurable evidence. I can design an experiment that would falsify my instinct. That's what it means to check your bias. Personally, I feel like not enough liberals (more generally, not enough people) do that.

They're offended at their own idea of what a liberal is.

But to be fair, I may be projecting my own idea of what I think a conservative is. I don't think you'll see much of a difference when you compare the viewpoints as whole entities; at the risk of a platitude - there are outrageous and frankly unintelligent people on both sides, but what gives me confidence that I'm closer to the truth is that I will shine this critical light on my own viewpoint and that I'm actually trying to be fair in the first place.

Why doesn't it scale past 1TB currently? I have DVC repos that are indexing on the order of 1-2 TB of image data. The data hardly ever changes, and if it does there isn't a big problem in just storing both copies of the image (in XetHub it would be the same probably because most of the image pixels would be different, depending on the processing level). All that we really care about is that it's content addressable, it has access controls (otherwise we would use IPFS), and you can distribute subsets of the data.

If I tried XetHub on a 2TB dataset would it simply fail?