“Don’t shoot! I’m with the science team!”

If everyone has access to the model it becomes much easier to find obfuscation methods and validate them. It becomes an uphill battle. It’s unfortunate but it’s an inherent limitation of most safeguards.

Of course it was political retribution and not the whole unregistered securities and gambling market thing.

The model does have a lot of advantages over sdxl with the right prompting, but it seems to fall apart in prompts with more complex anatomy. Hopefully the community can fix it up once we have working trainers.

The “why would they make this” people don’t understand how important this type of research is. It’s important to show what’s possible so that we can be ready for it. There are many bad actors already pursuing similar tools if they don’t have them already. The worst case is being blindsided by something not seen before.

The 8B is incredible for it’s size and they’ve managed to do sane refusal training this time for the official instruct.

I’m sure the machine running it was quite warm actually.

Do another 2 day blackout. That’ll show 'em.

This article is grossly overstating the findings of the paper. It’s true that bad generated data hurts model performance, but that’s true of bad human data as well. The paper used opt125M as their generator model, a very small research model with fairly low quality and often incoherent outputs. The higher quality generated data which makes up a majority of the generated text online is far less of an issue. The use of generated data to improve output consistency is a common practice for both text and image models.

It’s size makes it basically useless. It underperforms models even in it’s active weight class. It’s nice that it’s available but Grok-0 would have been far more interesting.

I don’t think they care about the images being used, just the disruption of service. It’s pretty clear that this wasn’t a coordinated thing from Stability and was at most a lone individual acting in bad faith.

It’s pretty ironic though that the company that practices mass scraping has no rate limits to prevent outages due to mass scraping.

There should be no difference because the video track hasn’t been touched. Some software will display the length of the longest track rather than the length of the main video track. It’s likely that the the audio track was originally longer than the video track and because of the offset it’s now shorter.

You can use tools like ffmpeg and mediainfo to count the actual frames in each to verify.

I doubt any platform could be more volitile than Twitter with Musk at the helm.

Who’s dumb enough to pay for that? Everyone else is just scraping it for free.

Tun0 is the interface that most vpns are using so I assume proton is the same.

This isn’t necessarily about just hardware. Current ML architectures and inference engines are far from being at peak efficiency. Just last year we saw 20x speedups for llm inference on some hardware. “a million times” is obviously hyperpole though.

This is why you should always selfhost your AI girlfriend.

This is an article about another article, some top tier journalism. They’re right about the external display though. I’ve yet to see a positive comment about it, seems like just a weird gimmick that drains the already short battery life.

Seems kind of like phi but for writing, the smaller ones are trained with 50B tokens and the largest is only trained with 18B.

Anywhere speculative investment is involved there are cult like patterns. If your investors don’t believe that your product is going to revolutionize its field you’re not going to get the kind of funding these startups want.