InternetException

Nonsensical LLMs writing

· 586 words · ~2 min read · in AI

#coding agents #llm

I am experimenting lately with code written by models in some of my pet projects. It is a bit easier to do side projects this way: I can for example instruct it to make some specific change and then do something else entirely - some chores or just get a nap.

It mostly works, but the most problematic part is comments and documentation written by the model. I am using Qwen 3.8 Flash Next lately and it is a great model for more complex tasks, and it can be run fairly fast on Strix Halo with halogen server. But its writing is so Claude-like.

I am fighting this and most of the time I resort to just removing them entirely, because they are so hard to understand. The value they add, in comparison to the cognitive load of trying to make sense of them, is so low that pressing delete is much less work. Especially because this is just a fun experiment I do on the side, so documentation in code practically does not matter.

One of the examples of such comments is:

The culture of the request is one this runtime knows no value to read in.

I know the intention here, because it was written as a summary of an enum value that is an error code of an HLinq query. If you do not know what HLinq is: it can be used to write Linq-like queries for your data in an HTTP query string. Pretty neat. I use culture in it to specify translation of query keywords and value formats. Specifying “pl” allows you to use a sane date format for dates - yyyy-MM-dd - instead of the ridiculous US format.

So I knew that specifying “124” as a culture is invalid. Obviously we need an error code for that. Obviously I understand that the comment should describe that case.

Nevertheless, this feels like it was written backwards, in a way that it does not describe what the error code is used for, but what an invalid culture is. It does not matter what an invalid culture is, besides the fact that it is invalid.

How would I write it?

Culture was invalid and not recognized by the runtime.

And that is it.

Even that feels somewhat redundant when you do not keep in mind that the set of valid cultures depends on .NET runtime settings and on the OS libraries that are available. But for most cases it is most probably the standard set that is available by default. So if a culture is unavailable, it means that the user sent a typo or some silly value like ‘foobar’ instead of a valid culture.

I think OpenAI models do not do that, at least not that much. I think Anthropic trained their models to write text that way so they appear ‘smart’ and ‘sophisticated’. Why pay a million dollars for tokens that will appear as written by an ordinary human being and not by a PhD scientist?

For better or worse, Qwen models are distilled Anthropic models. So they kind of inherit their style of writing.

Even instructing a model to rewrite some text in a terse way, using only simple words, is usually not enough, because it will write a sentence like the one above.

For now this is a losing battle for me, and it is probably solvable only by changing the model. But I am happy with its technical side for now, so I can write the documentation myself.