Muehe @ Muehe @lemmy.ml

Posts

9

Comments

243

Joined

2 yr. ago

Myth #1: Israel is guilty of “genocide” in Gaza.

The term “genocide” has a clear meaning—it’s the destruction or attempted destruction of a whole people.

Yes , 95% of people starving in the world were in Gaza when Israel chose to withhold aid to them. That is the attempted destruction of a whole people

Not sure why you are debating semantics here, as that statement is just straight up wrong, which can be easily confirmed by taking a single look at article 2 of the Genocide Convention (emphasis mine):

In the present Convention, genocide means any of the following acts committed with intent to destroy, in whole or in part, a national, ethnical, racial or religious group, as such

I don't doubt that a lot of well established companies are trying to scam their employees out of what they are owed. I mean that's just reiterating the OP basically.

“leadership is a lateral move”

Ahahaha, what the fuck? This is simply perfect. Peak oxymoron. A work of art. chef's kiss

Well I guess that depends very much on what you mean by being on life support. Like financially speaking? Oh yeah, they are more or less entirely dependent on Google. Regarding user numbers? Sure, Statcounter says 3.3% currently. Technologically speaking? Not really, quite the opposite actually. Besides Apples WebKit and Googles fork of it called Blink there is but one game in browser engine town, and its name is Gecko.

Now? Want?!

I sometimes forget that this picture exists, and then I happen upon it in places like here and it just smacks me in the face how perfectly it encapsulates the total and utter loss of decorum in politics. I mean it was never perfect obviously, but in past times there was a somewhat reasonable expectation of politicians being civil and them losing their office if they were publicly caught out not to be. It was rare, but it happened. Yet here you have the supposedly "most powerful man in the world" just dropping every pretence and hustling for some company in a flagrant abuse of his office. It's so brazenly corrupt. And the worst thing is this was just another Tuesday for Trump, mild shit-storm, on to the next fucked up thing he did. Society never even had time to realise what a historic moment this was. It was just dropped on the pile.

Sir, this is a Wendy's shitpost community.

If you have a mouse wheel try pressing it down while hovering over a link. Should open in new tab.

Biochemically adjusting to a brighter environment takes the eyes a fraction of a second, but adapting to a darker environment takes them 20-30 minutes. This is the reason pirates wore eyepatches despite both eyes being intact, you can enter the belly of a ship and have at least one eye already adapted to darkness.

The highest GDPR fine was 1.2 billion.

This isn't the GDPR but the DMA. That said, fines there are even steeper, 10% of global revenue for the first offence, 20% for repeated offences.

Absolutely, what we have on the Discord in the way of documentation is a straight forward install guide with one screenshot and a download link (of the GitHub release) and a FAQ channel, which is basically just links to the GitHub for a good part of the answers. We also automatically mirror our changelog there. But that's it, and it's all on GitHub as well.

What gets sadly lost on GitHub sometimes is "emerging events" like a new release of ours or the game we mod breaking something, where we will get yelled at on the Discord immediately and might have a hotfix release out before anybody even managed to create a proper GitHub issue.

Edit: Oh and temporary workarounds. If we figured something out on the Discord it doesn't get posted to Github necessarily even if there is already an issue. Hence why I'm looking into having a bot for that instead of literally having to copy and paste a message.

I hate how devs use Discord for documentation. All the info on there is fleeting.

Guilty as charged, but in our defence we mirror most of the info from/to GitHub best we can. Also you can make the information somewhat less fleeting by pinning comments to a channel, using forum channels, or creating channels where users only have read access. Of course this doesn't prevent the data from going away if Discord does, but to be fair the same can be said about almost all other services as well. GitHub servers get ransomwared and they don't pay? Yeah your changes until their last uncorrupted backup are gone now unless you had backups of your own.

The reason why we use Discord in the first place though is network effect. The amount of reports and questions we get on Discord is simply no comparison to GitHub. It's more simply because more users already have Discord than do GitHub leading to a lesser barrier of entry (account creation/program installation), especially for gaming related projects like ours. Of course this creates some added bureaucracy for keeping track of important reports from Discord. It's kind of manageable to do manually, but I have been looking into ways of having a bot transfer messages/threads to GitHub by simply replying with an !issue 4321 command or something. Sadly I'm pretty sure we wouldn't get half the reports we do on Matrix/IRC/XMPP/whatever, same diff if we were to switch from GitHub to GitLab basically.

Lastly, a server owner (or someone given the rights by them) can get an API key that enables them to dump the full server logs to disk. So if you really want your Discord server content to be indexed by search engines the possibility to just host a copy of your logs as a static website is technically there (we admittedly don't do this yet, not sure if there are existing projects for this).

Know what data source isn’t fleeting? Forums.

Guess you never were a member of a forum with private sub-forums that went out of maintenance? That info is just as gone as our Discord logs if the company croaks tomorrow. And the public part is only available if it was mirrored to web.archive.org or something, which isn't guaranteed either.

In summary, yes Discord isn't the shit, it's just shit, but the people are there. If the mountain won't come to you, then you must go to the mountain. ¯(ツ)/¯

https://knowyourmeme.com/memes/loss

TL;DR: Somebody made a really weird episode of a web comic and the configuration of the figures in the panel has become a meme named "Loss" after the name of that episode. Pic related:

The joke here is that evolution gave you pattern recognition to avoid predators, but now you are using it for useless things like recognising this comic is spatially organised in the same way the Loss comic is.

a neural network with a series of layers (W in this case would be a single layer)

I understood this differently. W is a whole model, not a single layer of a model. W is a layer of the Transformer architecture, not of a model. So it is a single feed forward or attention model, which is a layer in the Transformer. As the paper says, a LoRA:

injects trainable rank decomposition matrices into each layer of the Transformer architecture

It basically learns shifting the output of each Transformer layer. But the original Transformer stays intact, which is the whole point, as it lets you quickly train a LoRA when you need this extra bias, and you can switch to another for a different task easily, without re-training your Transformer. So if the source of the bias you want to get rid off is already in these original models in the Transformer, you are just fighting fire with fire.

Which is a good approach for specific situations, but not for general ones. In the context of OP you would need one LoRA for fighting it sexualising Asian women, then you would need another one for the next bias you find, and before you know it you have hundreds and your output quality has degraded irrecoverably.

Depressingly, the message that GHG emissions are heating up the planet has been passed down for over a hundred years now. People just aren't very good with passed down messages in general.

Yeah but that's my point, right?

That

1. you do not "replace data until your desired objective".
2. the original model stays intact (the W in the picture you embedded).

Meaning that when you change or remove the LoRA (A and B), the same types of biases will just resurface from the original model (W). Hence "less biased" W being the preferable solution, where possible.

Don't get me wrong, LoRAs seem quite interesting, they just don't seem like a good general approach to fighting model bias.

First, there is no thing as a “de-biased” training set, only sets with whatever target series of biases you define for them to reflect.

Yes, I obviously meant "de-biased" by definition of whoever makes the set. Didn't think it worth mentioning, as it seems self evident. But again, in concrete terms regarding the OP this just means not having your dataset skewed towards sexualised depictions of certain groups.

1. either you replace data until your desired objective, which will reduce the model’s quality for any of the alternatives

[...]
For reference, LoRAs are a sledgehammer approach to apply the first way.

The paper introducing LoRA seems to disagree (emphasis mine):

We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks.

There is no data replaced, the model is not changed at all. In fact if I'm not misunderstanding it adds an additional neural network on top of the pre-trained one, i.e. it's adding data instead of replacing any. Fighting bias with bias if you will.

And I think this is relevant to a discussion of all models, as reproduction of training set biases is something common to all neural networks.

Just here to leave the daily reminder that API reimplementation may constitute fair use under certain circumstances.

“Inclusive models” would need to be larger.

[citation needed]

To my understanding the problem is that the models reproduce biases in the training material, not model size. Alignment is currently a manual process after the initial unsupervised learning phase, often done by click-workers (Reinforcement Learning from Human Feedback, RLHF), and aimed at coaxing the model towards more "politically correct" outputs; But ultimately at that time the damage is already done since the bias is encoded in the model weights and will resurface in the outputs just randomly or if you "jailbreak" enough.

In the context of the OP, if your training material has a high volume of sexualised depictions of Asian women the model will reproduce that in its outputs. Which is also the argument the article makes. So what you need for more inclusive models is essentially a de-biased training set designed with that specific purpose in mind.

I'm glad to be corrected here, especially if you have any sources to look at.

The video was also mirrored on the Mastodon link in the OP. Here is an embed of that: