It's fun how complex the curves end up
The starting point is just min(1.0, 0.006 + max(0.0, (x - 72) * 0.06)), it's just a piecewise linear function with two pieces (technically three but the third part is kinda irrelevant)
But then you apply random chance a whole lot of times and you get these fun curvy graphs that go up and down all smooth like(edited)
There's obviously going to be a very high lower bound for hardware capable of running very good models
You can't run something actually smart on a potato(edited)
And the very best models will always be ludicrously expensive because well funded AI labs have the big bucks and they will use all the resources they're given
I don't say an ordinary person should do that. But what if Facebook decides to use their open source model which is free, for their own benefit? Instead of spending even more money to maintain their own?
exactly, consider this
OpenAI sents billions of dollars developing their own stuff
they sell their stuff for other companies for a fee
then a new free competitor comes up which tells the companies that theirs is free
suddenly, all other companies want to use that free version instead of yours
05:23
Boom. they just lost that much money. Billions
05:24
Not in an instant, this will accumulate over time
05:25
They lost the game, completely. Follows up all the other competitors which do the same, or similar thing
Oh, did I tell that the open source models have no restrictions?
05:29
This actually blew my mind when I read it from an article
risu
There's obviously going to be a very high lower bound for hardware capable of running very good models
You can't run something actually smart on a potato (edited)
Since their largest model is hosted by them on their website and the company is based on the Eastern region, it tend to have restrictions, but their open source models don't have that.
06:58
Cannot really verify tho myself, but I didn't hear anyone saying the opposite, and they themselves can check the open source project(edited)
DeepSeek's first-generation of reasoning models with comparable performance to OpenAI-o1, including six dense models distilled from DeepSeek-R1 based on Llama and Qwen.
10:02
Maybe this works
10:03
If you never used one, install ollama and run ollama run deepseek-r1:8b(edited)
10:04
But be warned, the smaller the model the dumber it is.
10:05
Even 32b will feel stupid compared to free chat gpt
BREAKING: President Trump announces the U.S. will be placing tariffs on all semi-conductors and pharmaceuticals imported from Taiwan in the very near future
Keep a list of states and their probabilities, and at every step go through each tuple and produce every possible resulting state and its corresponding probability
11:58
I checked, the number of states tracked grows linearly each step until it reaches a maximum at a bit over 1k states, then starts linearly decreasing until in the end all paths lead to the final C6 state
12:02
So I can claim that it runs in O(n) even though for a couple hundred pulls it's quadratic at best
12:06
"At best" I say because my "list of tuples" is actually a Python dict
but on first boot i managed to get image but only on second display
13:27
[0129/162149.937:ERROR:registration_protocol_win.cc(84)] TransactNamedPipe: The pipe has been ended. (0x6D)
[0129/162149.940:ERROR:settings.cc(319)] Settings magic is not 1129342067
theoretically you can also run a 32b on that. but it will have to constantly switch vram with ram, or force the GPU to read from your ram.
which makes the model much much slower
Keep a list of states and their probabilities, and at every step go through each tuple and produce every possible resulting state and its corresponding probability
Wait, there are only 1440 possible states (or 5760 with the new system that we don't have exact details on yet), so we could just store the probability of each state in a 1440 element vector, and to calculate next set of probabilities is just a matrix multiplication
15:28
Genshin gacha can be represented as a 1440x1440 matrix (5760x5760 actually)
Taiwan in the very near future
️
