The worse the AI performance, the higher the bonus? Doctor of New York University offered a reward for the task of making the big model perform poorly

The bigger the model 、 The worse the performance, the better the prize ？

The total bonus is 25 Ten thousand dollars （ Renminbi conversion 167 ten thousand ）？

such “ Out of line ” It really happened , A man named Inverse Scaling Prize（ Anti scale effect Award ） The game of caused heated discussion on twitter .

The competition was organized by New York University 7 Jointly organized by researchers .

Originator Ethan Perez Express , The main purpose of this competition , It is hoped to find out which tasks will make the large model show anti scale effect , So as to find out some problems in the current large model pre training .

Now? , The competition is receiving contributions , The first round of submissions will end 2022 year 8 month 27 Japan .

 

Competition motivation

People seem to acquiesce , As the language model gets bigger , The operation effect will be better and better .

However , Large language models are not without flaws , For example, race 、 Gender and religious prejudice , And produce some fuzzy error messages .

The scale effect shows , With the number of parameters 、 The amount of computation used and the size of the data set increase , The language model will get better （ In terms of test losses and downstream performance ）.

We assume that some tasks have the opposite trend ： With the increase of language model testing loss , Task performance becomes monotonous 、 The effect becomes worse , We call this phenomenon anti scale effect , Contrary to the scale effect .

This competition aims to find more anti scale tasks , Analyze which types of tasks are prone to show anti scale effects , Especially those tasks that require high security .

meanwhile , The anti scale effect task will also help to study the potential problems in the current language model pre training and scale paradigm .

As language models are increasingly applied to real-world applications , The practical significance of this study is also increasing .

Collection of anti scale effect tasks , It will help reduce the risk of adverse consequences of large language models , And prevent harm to real users .

 

Netizen disputes

But for this competition , Some netizens put forward different views ：

I think this is misleading . Because it assumes that the model is static , And stop after pre training .

This is more a problem of pre training on standard corpora with more parameters , Not the size of the model .

Software engineer James Agree with this view ：

Yes , This whole thing is a hoax . Anything a small model can learn , Large models can also .

The deviation of the small model is larger , therefore “ Hot dogs are not hot dogs ” It may be recognized as 100% Right , When the big model realized that it could make cakes similar to hot dogs , The accuracy will drop to 98%.

James Even further proposed “ Conspiracy theories ” View of the ：

Maybe the whole thing is a hoax —— Let people work hard , And show the training data when encountering difficult tasks , This experience will be absorbed by large models , Large models will eventually be better .

So they don't need to give bonuses , You will also get a better large-scale model .

Regarding this , Originator Ethan Perez Write in the comment ：

Clarify it. , The focus of this award is to find language model pre training that will lead to anti scale effect , Never or rarely seen category .

This is just a way to use large models . There are many other settings that can lead to anti scale effects , Not included in our awards .