What’s the deal with Alpine not using GNU? Is it a technical or ideological thing? Or is it another “because we can” type distro?

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 names missing from the list say more about the board’s purpose than the names on it.

The issue is that they have no way of verifying that. We’d have to trust 2 other companies in addition to DDG.

All of Firefox’s ai initiatives including translation and chat are completely local. They have no impact on privacy.

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.

Cohere’s command-r models are trained for exactly this type of task. The real struggle is finding a way to feed relevant sources into the model. There are plenty of projects that have attempted it but few can do more than pulling the first few search results.

They’re already lying to get passed the 13 year requirement so I doubt it would make any difference.

I don’t think the term open-source can be applied to model weights. Even if you have the exact data, config, trainer and cluster it’s basically impossible to reproduce an exact model. Calling a model “open” sort of works but then there’s the distinction between open for research and open for commercial use. I think it’s kind of similar to the “free” software distinction. Maybe there’s some Latin word we could use.

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

Your best bet would probably be to get a used office PC to put the card in. You’ll likely have to replace the power supply and maybe swap the storage but with how much proper external enclosures go for the price might not be too different. Some frameworks don’t support direct GPU loading so make sure that you have more ram than vram.

An arm soc won’t work in most cases due to a lack of bandwidth and software support. The only board I know of that can do it is the rpi5 and that’s still mostly a poc.

In general I wouldn’t recomend a titan x unless you already have one because it’s been deprecated in cuda, so getting modern libraries to work will be a pain.

I really like the simplicity and formatting of stock pacman. It’s not super colorful but it’s fast and gives you all of the info you need. yay (or paru if you’re a hipster) is the icing on top.

Partnered with Adobe research so we’re never going to get the actual model.

This has more to do with how much chess data was fed into the model than any kind of reasoning ability. A 50M model can learn to play at 1500 elo with enough training: https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html

It does a little bit worse than v0.1 on all benchmarks which isn’t ideal. That doesn’t really say much about the finetuning potential though.

The “AI PC” specification requires a minimum of 40TOPs of AI compute which is over double the 18TOPs in the current M3s. Direct comparison doesn’t really work though.

What really matters is how it’s made available for development. The Neural engine is basically a black box. It can’t be incorporated into any low level projects because it’s only made available through a high-level swift api. Intel by comparison seems to be targeting pytorch acceleration with their libraries.

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.