Thread

Interesting article from Tristan Greene that paints quite a dark picture of the state of AI research. I dont' know whether there really is rampant fraud in this area, this area has definitely underwent radical change in the last decade or so. twitter.com/Kingwulf/statu…

Deep learning has brought immense improvements, but it is also very opaque. Breakthrough papers present new architectures that performed better on benchmark data sets, but there is little scientific understanding why that is so.

If you look at a random paper (e.g. transformers), you'll see that there many design decisions that went into this, but no analysis of which of these choices really brought the biggest differences. At most there is a discussion of structural differences to existing work.

Add to that that the models become bigger and bigger and it takes real money to train these models. How would a reviewer test a paper? You can look at formal correctness and whether you agree with the experimental setup, but aside from that there isn't much you can do.

Now add to that the immense financial investment and interest that companies like Google have. It's not just using AI in their products, they also produce tensorflow, and provide a cloud service that people can use.

I'm just saying that all of this creates a huge incentive which need not be aligned with the scientific process.