declaredapple

12 ms·

31.
by declaredapple 3y ago
Do these completely replace the nvidia drivers? Do they support DLSS 3 by chance? For some reason nvidia never released support for dlss 3/3.5 unless I missed it.
32.
by declaredapple 3y ago
Github's profile pictures are 40x40px in issues/pr's/commits/etc, they're very rarely seen above that, and I think sdxl lightning creates acceptable 1024x1024 images in many cases - downscaling to 512x512 hides
33.
by declaredapple 3y ago
A lot of use cases are cost-limited. Dalle3 makes great images but costs $0.12 per image so it would get extremely expensive at scale (1k generations is already 120$). The cost is by gpu time, the faster you can generate it the cheaper it i
34.
by declaredapple 3y ago
The best way to do it right now is controlnets. I'm not sure about re-doing the text of old images - you could try img2img but coherence is an issue, more controlnets might help
35.
by declaredapple 3y ago
Many people are annoyed by the recent influx of calling everything "AI". Machine learning, statistical models, procedural generation, literally an usage of heuristics are all being called "AI" nowadays which obfuscates t
36.
by declaredapple 3y ago
Well it was released today Very likely a coincidence. https://opus-codec.org/release/stable/2024/03/04/libopus-1_5...
37.
by declaredapple 3y ago
I'm more curious about the input/output token discrepancy Their pricing suggests that either output tokens are more expensive for some technical reason, or they're trying to encourage a specific type of usage pattern, etc.
38.
by declaredapple 3y ago
> No GPT5 = users go elsewhere. You're not wrong, but most of the big players will take a while to switch, at least in my experience you have to put more effort into making sure your prompts result in what you want, and that's
39.
by declaredapple 3y ago
Yeah the output pricing I think is really interesting, 150% more expensive input tokens 250% more expensive output tokens, I wonder what's behind that? That suggests the inference time is more expensive then the memory needed to load i
40.
by declaredapple 3y ago
GPT-4 is also cheaper, 10/30 $/m vs 15/75 $/m for claude 3 opus
41.
by declaredapple 3y ago
I'm actually surprised they did it without prepay. The reason is the card transaction fees with most providers are usually a flat rate + %. $0.3+some% is typical I think? And my last few month's invoices were 1.95, 3.81, 1.44, 2.0
42.
by declaredapple 3y ago
> LIKE WHAT? Well it being a stuttery mess for one. I could care less any of the features you mentioned when games or even the desktop is unusable. Signal couldn't even render correctly. I had a myriad of other issues, switching to
43.
by declaredapple 3y ago
> We get it, you don't care about mixed DPI, fractional scaling, HDR, non-tearing performance, VRR, a viable driver model for the 2020s, etc. FWIW my previous experience a year ago was a stuttery mess making it a complete non-starte
44.
by declaredapple 3y ago
My friend I know you just discovered nix but "don't have processes we don't need using cpu cycles" is low hanging fruit. If you really care then you'd use a bare linux kernel. Of course mining is done on ASICs. If y
45.
by declaredapple 3y ago
That's not their first "closed-source" model. "mistral-medium" was. Depending on your definition, none of mistral's models are "open source", AFAIK they've given no information on what training d
46.
by declaredapple 3y ago
How is nvidia on wayland these days? Is running an nvidia gpu + gnome/kde + steam viable these days? Do most games work well?
47.
by declaredapple 3y ago
I'll agree with you, and add that inference speed is a big factor too. SDXL-ligtning/cascade can generate images in 200ms which is fast enough to fit in a web request, and paradoxically makes it even cheaper to generate. And using
48.
by declaredapple 3y ago
It's not really. And 8x7B is not a 7B model, it's a MoE that's closer to 60B that has to be kept in memory, and uses 2 experts per token so it runs at 15B speeds. All of the current frameworks support MoE and sharding among G
49.
by declaredapple 3y ago
It didn't take long for perplexity, anyscale, together.ai, groq, deepinfra, or lepton to all host mistral's 8x7B model, both faster and cheaper then Mistral's own api. https://artificialanalysis.ai/models&#x2F
50.
by declaredapple 3y ago
I don't think it was too silent > On February 27, this feature will be start to be enabled automatically for all free accounts across GitHub. https://github.blog/changelog/2024-02-14-secret-scannings-pu...
51.
by declaredapple 3y ago
> The change in endpoint name is a strong suggestion I don't think the naming really suggests that. The new naming suggests they'll have two sets, the "open" models and their commercial ones. I do agree with your skep
52.
by declaredapple 3y ago
> We can already improve ourselves without identifying as disordered. The purpose of the "disorder" is to group common pathologies and their respective treatments (methods to improve). Let's take high functioning autism as
53.
by declaredapple 3y ago
> Spotify Premium is already hifi Hard disagree. I do agree that for the majority of people, the 320kbit/s is absolutely enough for most people (arguably 160kbps is). But it's not "hifi". And I don't mean "
54.
by declaredapple 3y ago
> I've never seen them fail due to lacking features or capability, and I haven't heard stories of it either. I've seen a mix of both for years. For the last 6 years the story has usually been some combination of "X p
55.
by declaredapple 3y ago
I'm pretty confident they haven't until at least 6 months ago when they freaked out about api usage. It's possible they've sold it and we haven't heard about it since then. I would think they'd want to brag abo
56.
by declaredapple 3y ago
> Arguably Reddit's value is it's data Reddit has a chicken and egg problem - nobody knows what the value of it's data even is. They've never sold it before. You can sell pet rocks for 1 billion dollars, but nobody wi
57.
by declaredapple 3y ago
Any chances of an API? And are there plans to release any more weights? Perhaps one or two revisions behind your latest ones?
58.
by declaredapple 3y ago
How many tokens/s are we talking for a 70B model? Last I saw they performed really poorly, like lower single digits t/s. Don't get me wrong they're probably a decent value for experimenting with it, but is flat out pathe
59.
by declaredapple 3y ago
They've been very happy selling shovels at a steep margin to literally endless customers. The reason is because they instantly get a risk free guaranteed VERY healthy margin on every card they sell, and there's endless customers l
60.
by declaredapple 3y ago
> At what point does it become more feasible to rewrite your architecture or use less GPU They are but it takes a lot of time. Most of the big players - Google, Meta, OpenAI, Amazon, and Microsoft all are actively developing TPU/NPU

More ›