I fucking loathe the term "compute". Every time one of these mealy-mouthed motherfuckers lets it slide through sphincter-like lips I want to kick some teeth in.
To be fair an argument can be made for the Lego block one, using a novel combination of existing technologies to get better results is how nearly all innovation happens in machine learning.
Proving a thing that's only known empirically is extremely valuable, too. We've an enormous amount of evidence that the Riemann hypothesis is correct - we can produce an infinite amount of points on the line, in fact - but proving it is a different matter.
Especially in ML too. It's currently easier to integrate multiple small specialised models than to train a big model for every use case. If I understand correctly, that was one of the main motivations for Anthropic developing the Model Context Protocol, including interacting with LLMs from front-end clients.
Depending on the application case and benchmark, being 0.1 to 0.3 % better than other SOTA approaches can still be statistically highly significant. Even though such a number does notmlook like much, it can mean a large leap forward in practise.
Anyway, I wouod add a machine learning paper type that was written by an LLM and nobody cared to call that out in the peer review.
I fucking loathe the term "compute". Every time one of these mealy-mouthed motherfuckers lets it slide through sphincter-like lips I want to kick some teeth in.
Your rage makes me feel seen. I share your feelings.
To be fair an argument can be made for the Lego block one, using a novel combination of existing technologies to get better results is how nearly all innovation happens in machine learning.
Proving a thing that's only known empirically is extremely valuable, too. We've an enormous amount of evidence that the Riemann hypothesis is correct - we can produce an infinite amount of points on the line, in fact - but proving it is a different matter.
And for the kid challenging the 0.1% result, that’s about as close to pure scientific method as you can get.
Especially in ML too. It's currently easier to integrate multiple small specialised models than to train a big model for every use case. If I understand correctly, that was one of the main motivations for Anthropic developing the Model Context Protocol, including interacting with LLMs from front-end clients.
"Our model has no sense of permanence or real understanding of what words even mean and we re-interpreted this as the ability to lie."
“We repeat the experiment with a newer dataset and act like we are the first doing this kind of experiment”
“We talk about possible applications in the future writing like your run-of-the-mill generalist newspaper”
“Another article resuming other articles”
Missing the Survey paper: Here’s the results of 5-10 other papers
Depending on the application case and benchmark, being 0.1 to 0.3 % better than other SOTA approaches can still be statistically highly significant. Even though such a number does notmlook like much, it can mean a large leap forward in practise.
Anyway, I wouod add a machine learning paper type that was written by an LLM and nobody cared to call that out in the peer review.