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