SkitterKherpi

They'll probably be back later. It's very possible they are "saving up" for a much bigger run.

$5/$25 for Opus 4.8 vs $3/$15 doesnt seem cheaper enough to be too worth it. It depends how much better it is than e.g. Mimo, but I imagine Mimo and co to be too cost efficient in the lower tier to be overtaken by Sonnet for most tasks.

I assume you have the dgx spark? At this point I am not 100% on the difference other than Linux and Windows. The RTX spark should come around Q4, unless I am mistaken.

You can run 8bit 27B models at 24GB, it's definitely enough for the model size.

10k is rather a lot yes. For LLMs you can use a lot of tokens with 10k with less hassle without the machine (and also it's not like electricity is free), but for some other things like video models 10k would get burned very fast. I am looking for something more in the 5k range though.

I am considering getting something like NVIDIA's RTX Spark when it comes out, though even that will be limited to 128GB.

27-30B in general seems to be the level where you actually start having decent models. I just wish consumer hardware hadn't stagnated so much that we can't easily go higher than that, and that even running those requires a $5k machine now.

Gemni is really not that far behind, 3.1Pro was the best model or close to it when it came out, 3.5Pro likely wont be but whenever they get to 4 that could easily be Fable level, just later than when Fable itself came out.

So them banning Fable for only non-Americans is what we non-Americans should expect to be the norm going forward? Way to build even more resentment abroad.

I'm very pro-west, but at this point okay, I guess the rest of us have to side with China, not because we remotely like it, but because they don't try to be quite so antagonistic to us in everything they do.

I like the idea, but I am worried that it's yet another step on the road towards personal multi-purpose tower PCs built part by part no longer being a thing.

It has so far been the kind of thing that always feels like the next version of the local models would be the one that is just good enough.

It is impressive how well they've scheduled all their releases, posts, and other news to dominate the tech news cycle almost every day in this pre-IPO phase.

I can see it but even if I do that for something like tests I'd still eat the time cost of the normal Gemma for 10% extra performance. And further, if you switch between the fast and normal Gemma for different tasks you eat the big time cost of loading the other model (and maintaining both in the first place).

It is cool but local models while okay already feel noticeably worse than even the cheapest APIs so I can't see myself sacrificing even a little bit of their quality for speed. I'm sure it's worth it for some usecases, curious to hear specific ones that people are already planning to deploy to production.

There's a lot more examples of grep usage than Visual code search in the training set.

It also tried to force usage the paid Claude API instead of claude code usage just because there's a mention of another provider we might want to plug in (which hasnt even happened) for AI integration.

Since the EU has been under investing in nuclear, solar + batteries investments is the only way forward. While the EU is making good progress, I still think it should invest a lot more a lot faster.

EDIT: I just found out that my comments show up as dead to everyone else. Can you please change that or let me know what to do? I am not a bot...

This reminds me of GANs and generating directly from a latent vector, as well as exploring them. I wonder how many of the original tricks used with GANs (like extrapolation stuff, optimizing for the latent vector to get a result etc) are easy to replicate with it.

Perhaps this mainly tested a subset of them but context files as long term memory clearly help in general. The models make the same type of mistakes without them, which are often trivially fixed from then on if the desired behavior is included in the md files.

There are definitely things Deepseek beats GPT on - cost effectiveness, lack of reluctance on some tasks, but from using most models I doubt it actually outperforms in a useful way in quality in a meaningful way.

I've never associated it with speed either due to UI. It's not slow, and it does feel snappy compared to some of the older alternatives but it's never felt particularly fast. Granted, we don't use it as much anymore.