qeternity

8 ms·

121.
by qeternity 1y ago
> There's no reason for someone to use this over Anthropic, OpenAI, etc Uh except that you can host this yourself? You just have to license the model. What is the issue?
122.
by qeternity 1y ago
> where the crews are making well below union wages and can work long hours without regulation. This is what is so ridiculous about unions in 2025: there are people in dozens of other tier 2 countries that will work harder than you for l
123.
by qeternity 1y ago
A minimum production setup for V3/R1 is 16x H100s…I guess that’s up to you whether that qualifies as cheap.
124.
by qeternity 2y ago
> Would this even be possible to do, given that the latency between nodes is very high and the bandwidth limited? Yes, it's possible. But no, it would not be remotely sensible given the performance implications. There is a reason wh
125.
by qeternity 2y ago
DeepSeek probably would have done this anyway, but they did release a Llama 8B distillation and the Meta terms of use require any derivative works to have Llama in the name. So it also might have just made sense to do for all of them. Otoh,
126.
by qeternity 2y ago
Great, somebody reinvented control vectors. This industry needs to stop reinventing things every 6 months.
127.
by qeternity 2y ago
It's convenience. And people pay for convenience all the time.
128.
by qeternity 2y ago
Zero imperial gain? This is a rather incredible claim: why do you say so?
129.
by qeternity 2y ago
"Usable" is the key word here. Not all context is created equal. Have a look at the RULER benchmark for a bit more detail.
130.
by qeternity 2y ago
What? There are highly optimized Marlin kernels for W8A16 that function very well on a 3090 in both FP8 and INT8 formats.
131.
by qeternity 2y ago
Well, FP4 (mostly INT4 these days) means you are moving fewer bits per weight around, so the memory bandwidth that you have performs better.
132.
by qeternity 2y ago
That sounds correct except for that total parameter count. 110B per expert at 16 experts puts you just shy of 1.8T. Are you suggesting there are ca. 30B shared params between experts?
133.
by qeternity 2y ago
GPT-4 was rumored to be 1.8T params...not 1.2 And the successor model was called "Orion", not "Omni".
134.
by qeternity 2y ago
Well, it's not...it gets most details wrong.
135.
by qeternity 2y ago
GPT-4 was rumored to be 1.8T params...not 1.2 And the successor model was called "Orion", not "Omni".
136.
by qeternity 2y ago
If the world is smaller today then it means it was bigger at that time. What are you confused about?
137.
by qeternity 2y ago
We use TSDB and are pretty happy with it. But it is much less performant than CH.
138.
by qeternity 2y ago
Well, the most important US equity market is in Aurora…
139.
by qeternity 2y ago
How about you go study first, instead of just trying to crank out AI generated slop. Absolutely no point in helping you correct your articles when they should instead be left obviously bad, so that they can be flagged.
140.
by qeternity 2y ago
Was this written by DeepSeek? Aside from not being in depth, it’s also inaccurate (MoE details and MTP misunderstanding).
141.
by qeternity 2y ago
If they are the same underlying model, it’s unlikely the prices will be different on a per token basis. The high model will simply consume more tokens.
142.
by qeternity 2y ago
They don’t have poor the way other countries do. And that’s down to culture. I don’t think people appreciate just how homogenous Scandinavia is. (Although this is changing, especially in Sweden) I also don’t think most people realize how sm
143.
by qeternity 2y ago
> I still believe Sonnet is better, but I don't think it is 10 times better. Sonnet doesn't need to be 10 times better. It just needs to be better enough such that the downstream task improves more than the additional cost. Thi
144.
by qeternity 2y ago
Well for one, Libertarians are not far right in the way you are describing...
145.
by qeternity 2y ago
Yeah this simply wouldn't work. Models don't have any concept of "themselves". These are just large matrices of floating points that we multiply together to predict a new token. The context size would have to be in the t
146.
by qeternity 2y ago
This says more about benchmarks than R1, which I do believe is absolutely an impressive model. For instance, in coding tasks, Sonnet 3.5 has benchmarked below other models for some time now, but there is fairly prevalent view that Sonnet 3.
147.
by qeternity 2y ago
No. That’s not subsidization. Of course insurance is about pooling risk. But subsidizing implies you’re doing it below market rates. If I pay my market rate but still -EV insurance premium, I’m not subsidizing anyone. I’m just happy to pay
148.
by qeternity 2y ago
Why is there a specific list? Why don't we just let Medicare negotiate.
149.
by qeternity 2y ago
Which effectively means that anybody in a less risky area of California is just subsidizing those who live in the risky areas. Premia across the board will increase as a result. Typical California redistribution...but this is from the botto
150.
by qeternity 2y ago
Interesting that their pricing page is already 404'ing: https://www.sdf.com/pricing

More ›