← Home ← Back to /g/

Thread 106213785

16 posts 4 images /g/
Anonymous No.106213785 >>106214054 >>106214079 >>106214282 >>106216804 >>106217251 >>106217847
LLMs are just a sophisticated lossy compression algorithm.
Anonymous No.106213814
>sophisticated
Anonymous No.106213828
They should train LLMs on assembly code until it can substitute a compiler back-end. [spoiler] It's never gonna work. [/spoiler]
Anonymous No.106214054 >>106214689 >>106215032
>>106213785 (OP)
Can we compress files efficiently with it?
Anonymous No.106214079
>>106213785 (OP)
so is your brain
Anonymous No.106214282 >>106214675
>>106213785 (OP)
intelligence is compression, there's a reason why IQ tests consist entirely of pattern recognition yet predict every other facet of intelligence more strongly than anything else we've come up with
Anonymous No.106214639 >>106214951
hurrr durr, humans are just a clump of cells. ohh I'm soo smart.
Anonymous No.106214675
>>106214282
IQ only "predicts" the correlation between social status and education.
Anonymous No.106214689
>>106214054
Not if you count the model weights, but transmission can be 10% the size of lzma2 if it's plaintext and if it's written at a gradeschool level and it's in the training data. Text not in training compresses to about 90% if lzma2. So no, not really.
t. I tried this
Anonymous No.106214951
>>106214639
A human is almost infinitely complex though. An LLM at its core is an incredibly simple algorithm.
Anonymous No.106215032
>>106214054
file compression necessarily needs to be lossless, so no
but it can lossily compress the concepts shared across files and domains; the model weights are a miniscule fraction of the size of the dataset but can produce a high percentage of it
Anonymous No.106216804
>>106213785 (OP)
The musing that Large Language Models are merely a complex form of lossy compression is somewhat reductive. The training process of these models actively cultivates generalization, deliberately expanding the data's representation to produce a more diverse range of outputs. This process constructs fuzzy logic circuits - wherein the model learns to map input patterns to predictions(the outputs) with a degree of probabilistic similarity, thereby generalizing its understanding beyond the precise instances represented in the original training data.
Anonymous No.106217251
>>106213785 (OP)
Yeah that isn't too far off i suppose.
Anonymous No.106217658
Computers are just sophisticated electrical transformers.
Anonymous No.106217827
More or less, yeah.
https://github.com/Futrell/ziplm
Anonymous No.106217847
>>106213785 (OP)
Functions are just maps.
Functions describe everything.
Everything is just a compression algorithm.