I haven't touched Go in over a decade (since before generics!) but I can see why this would be true. My theory is that LLMs absolutely love very tight, focused context. Go inherently restricts how many abstractions you can stuff into your code, and more abstractions tend to make the context a lot more complex and noisy. So LLMs love Go code because it keeps things simple.

The thing about Go, which some have complained bitterly about and others (and TFA) have touted as a strength, is the limited expressiveness of the language (hence my remark about generics!) This is what restricts the number of abstractions in Go code, leading to more verbose but much simpler code all around. Choosing between simplicity and expressiveness is a matter of taste, but also organizational dynamics; for larger organizations which require a large amount of context shared amongst a large pool of employees, it's better for the code to be simpler and locally understandable. As TFA indicates, this has been a guiding principle for Go.

I think what is happening with AI coding is similarly related to context. Consider that while more expressive languages enable more abstractions, they can make the code more concise, but critically, this also spread the logic around. E.g. in large Java codebases you will find deep inheritance hierarchies with class and method definitions spread around a dozen different source files and JavaDoc references.

This necessitates finding and stuffing a lot more information into the context for any given task, a lot of it irrelevant and all of it more complex, because it requires making multiple hops of reasoning to figure out the logic. On the other hand with fewer abstractions, all the necessary code and logic though verbose is right there. It's much easier for a human and an agent to follow that code.

The difference is a human gets tired reading a lot of code, which is what pushes us to devise more abstractions, whereas an AI does not get tired.

I get the sense that if a context is stuffed full of highly relevant information, the agent will perform well regardless of the size of the context window. But the moment you pollute it with noisy irrelevant information, performance will drop regardless of the size of the window. (There are some papers showing this effect IIRC.) Hence simpler code, as encouraged by simpler languages like Go, are more amenable to tighter and simpler contexts, which work better for AI.