Comment by jyunwai | Hacker News Reader

Comment by jyunwai | Hacker News Reader

A related read is by Fei-Fei Li, who wrote more details about her point of view in a November 2023 article in The Atlantic [1]. The submitted article in The Washington Post focuses largely on the funding disparities, which she does highlight in her own article.

But her article in The Atlantic places a greater emphasis on the effects of a brain drain of AI researchers from academia to industry. She gives examples about how 40 roboticists left Carnegie Mellon for Uber in 2015, and how her close collaborator Andrej Karpathy chose OpenAI in its earlier days over "a faculty offer from Princeton."

She most notably recalls a quote said by someone at an OpenAI meeting that said that "Everyone doing research in AI should seriously question their role in academia going forward," and added the reflection: "To be honest, I wasn’t sure I even disagreed."

[1] https://www.theatlantic.com/technology/archive/2023/11/ai-et...

There was also past discussion on HN about her article at: https://news.ycombinator.com/item?id=38362242

Beat me to it! The Carnegie Mellon/Uber "scandal" was talked about a lot on HN at the time! Some of the big threads:

https://news.ycombinator.com/item?id=9638121

https://news.ycombinator.com/item?id=9602655

Uber left the self-driving car business four years later, selling that division to a startup. Did this work out well for the researchers, or did they end up getting a raw deal?

A key problem is Google alone has displaced academics e.g. at NeurIPS https://www.reddit.com/r/MachineLearning/comments/185pdax/d_...

This trend heavily biases AI research towards Google problems. Perhaps things will swing away as LLMs (and especially smaller LLMs) take over the field.

I would say this has a cascade affect and big well funded universities are displacing smaller institutions.

and add that a lot of mid-tier nations around the world, at their best universities, are getting very worried

I would not use publication numbers to argue anything. I'm an ML researcher at a university and publish at places like NeurIPS regularly.

There are so many edge cases that render this data meaningless. There are people who have their name on a dozen papers. Reviewers are terrible and papers with a lot of graphs run at large scale and easy messages tend to get in (I'm not saying that this is the only thing Google publishes, but people with access to a lot of compute tend to publish papers like this).

The funding gap between Google and academia is also massive. The fact that we're still competitive with Google is simply a testament to how incredibly efficient universities are and how inefficient Google is. They pay their people several times what we get paid. They have engineers which we don't. They have orders of magnitude more compute. They don't waste time on grants. Their total AI research budget is larger than the total combined AI research budget of all US universities. But.. somehow, they still account for only a small fraction of papers.

> But.. somehow, they still account for only a small fraction of papers.

They are only going to publish results which have no effect on their competitive advantage as a business ... which could be a small fraction of their results.

Eh. I don't buy that.

Business is built on engineering. That's what they spend the vast majority of their resources on.

No one has any secret sauce science at the moment. And if they did their engineers would be bragging about it in private, bringing it to new companies when they move, and we would be seeing it in products or features. Heck they would be telling us when they apply to grad school.

Imagine if Google didn't publish the transformers paper, Google does not just do engineering they also do research. If they find something similar today it wont go into the public since OpenAI set a precedent of not saying what they do.

If Google didn't publish the Transformers paper their engineers would have given talks about it, would have used those ideas to apply to grad school, would have switched companies and built a similar model that's just different enough not to technically be the same, and we would have reverse engineered it immediately. Simply knowing a few keywords would be enough for anyone competent to reproduce all of the work.

The space of possible models is huge. And we don't know in which direction to go to build better models. But knowing the vague general direction is all it takes. Any researcher can follow along given that knowledge. That's true of a lot of science not just ML.

No one has any secret sauce. Not even OpenAI.

Reply on news.ycombinator.com