Widely covered MIT paper saying AI boosts worker productivity is, in fact, complete bullshit it turns out.

https://www.wsj.com/tech/ai/mit-says-it-no-longer-stands-behind-students-ai-research-paper-11434092?st=sF3Wvo&reflink=desktopwebshare_permalink

@GossiTheDog

FWIW, here's my take.

0: "AI" means three things nowadays: neural nets, machine learning, and LLM stuff. They are different things.
1: There was a paper in Science last year in which Materials Science types were doing some seriously kewl work on systems with 5 different metals using "machine learning" (gradient descent search in high dimensional spaces). And calling it AI.
2: The Econ. grad student didn't understand this and thought they were doing LLM stuff. Oops.

@djl

How are neural nets and LLMs not machine learning?

@GossiTheDog

@snarkweek @GossiTheDog

Machine learning is a field that uses statistics to do its thing. Its tools include neural nets but not LLMs. (I dislike the term "machine learning", but to the best I can tell, they're smart sensible folks, statisticians doing gradient descent in insanely high-dimensional spaces.)

Dunno how LLMs could be called "machine learning", since they're exactly and only random text generators.

@djl

But the Wikipedia for LLM begins with:
"A large language model (LLM) is a type of machine learning model designed "

Why is that wrong in your view? I'm not trying to gotcha you, this field is new and quite incomprehensible, so it irks me a bit when people reshuffle the categories I'm just learning.

@GossiTheDog

@djl

So how I see it, LLMs apply machine learning to lingual token probability.

@GossiTheDog

@snarkweek @GossiTheDog

Blokes building LLMs use machine learning.

Blokes doing machine learning don't use LLMs.