vrighter

5 ms·

91.
by vrighter 2mo ago
the kitty graphics protocol supports graphics via a pipe too afaik. You're describing x11 all over again.
92.
by vrighter 2mo ago
User demands haven't evolved and changed. They stayed the same. They didn't need it before. They still don't need it now. They didn't ask for it then, they aren't asking for it now. They disliked it then, they still
93.
by vrighter 2mo ago
spacex the AI company, right?
94.
by vrighter 2mo ago
"trusted ground truth" Reads like "... and then draw the rest of the fucking owl"
95.
by vrighter 2mo ago
only if you bias your "random sampling of the distributions it gives" Fixing the seed is still intentional bias. Or you could force it to always take the one token with the highest probability, but that is still biased sampling.
96.
by vrighter 2mo ago
deterministic but completely unpredictable is not really any more useful
97.
by vrighter 2mo ago
this is not a belief. it is a fact. Doing the same thing in a higher dimensional space is still doing the same thing.
98.
by vrighter 2mo ago
except that it's really not. It's about finding a function to fit the data. And "generalizing" is only when the question is jsut barely outside their training data. Like a truncated taylor series, is still somewhat accur
99.
by vrighter 2mo ago
I think of it as more "providing you with the model you can use to decompress". You only really learn something when you do it yourself, not just listen to someone talk about it.
100.
by vrighter 2mo ago
While they were not monitoring the channel, it was not a distress call in the first place. Running out of fuel is considered "non-emergency marine assistance". And one is not obliged to respond to it. Only in a life threatening sc
101.
by vrighter 2mo ago
Some people don't have that.
102.
by vrighter 2mo ago
the limitations of an "emulated typewriter" keeps things from becoming too complicated. And also allows you to control it using just "echo". It has its warts, but I still think it's a good system.
103.
by vrighter 2mo ago
Because I don't want to order toilet paper and accidentaly have a monster-cock ordered instead
104.
by vrighter 2mo ago
this is overkill and way slower for most short commands. And you still need to scroll to wherever you need to edit. doesn't save anything, really.
105.
by vrighter 2mo ago
it absolutely does affect the semantics of the program. A properly tail called function can recurse indefinitely. One without TCO will stack overflow.
106.
by vrighter 2mo ago
I use a variation of it. "Typing the code in was never the hard part. The work is in figuring out what you want to describe"
107.
by vrighter 2mo ago
soon to be "had 15 years of top 1% knowledge"
108.
by vrighter 2mo ago
a prediction is not necessarily correct. What compressors actually store is how wrong the prediction was, in fact, a guess takes exactly 0 bits to encode. It builds a st atistical model at runtime based on past encountered data. Exactly the
109.
by vrighter 2mo ago
personally? 1: It's easier and faster to use IFF you spend the time to learn it. 2: Information density. It actually uses the whole screen, and isn't designed around applying (and taking to the extreme, which is the style at this
110.
by vrighter 2mo ago
not quite the same. Some of those survived, the llm model is unprofitable at its core. There will be no survivors.
111.
by vrighter 2mo ago
Statistical compressors are all about prediction. They can trivially be reversed to make a generator (not as good as an LLM, but only because it would be a much much much smaller model). And an llm can trivialyl be plopped in there to rep
112.
by vrighter 2mo ago
Actually, it's trivial. I did it for fun once when I was learning about the PPM algorithm. It took about 15 minutes to reverse the whole thing.
113.
by vrighter 2mo ago
Statistical compression works by learning statistics about the data as compression progresses. This is a(n admittedly very good) implementation of that, but the idea of learning about the data is as old as time (or computers, whichever came
114.
by vrighter 2mo ago
This is exactly why I think they are one and the same. It's relatively trivial to just plop a (lossy) machine learned markov chain instead of one learned (perfectly) from the data into PPM. With zero changes to the rest of the algorith
115.
by vrighter 2mo ago
there are dictionary compressors (decent compression, most common, fast), and statistical compressors (better compression, slower). Statistical compressors are much closer to LLMs in that an llm is learning statistics about the data too. An
116.
by vrighter 2mo ago
Once I was told to upgrade tls to 1.2 (latest version at the time) due to some compliance thing or other. So I did. Then I noticed something. The user never provided any certificates themselves. I asked why. "Because it would be bad ux
117.
by vrighter 2mo ago
If they take vc funding, it's only a matter of when, not if. They all say that.
118.
by vrighter 2mo ago
in our case, updates were done on a staggered schedule, automatically, and rebooted at a particular agreed upon downtime window. dev, then test, then live machines. I had to explain multiple times to them, for example, that I have updated t
119.
by vrighter 2mo ago
I embed lua when I need it. 200 to 400kb and we're off. And it's infinitely easier to interoperate with c than python is
120.
by vrighter 2mo ago
I also think they are one and the same. Because a network is a markov chain, with a huge, lossily compressed, state transition table. (that the chains weren't designed by a human is irrelevant to what they are). given that, you can ver

More ›