Copilot stops working on code that contains hardcoded banned words from GitHub (2023)

The AI industry is concerned about the fact that the world will consider them to be basically endorsing everything their AIs say. Thus, they are very afraid of there being a situation where you write "gender: 'm'" and hit autocomplete at the and end up with something like "gender: 'male as is normal'" or "gender: 'male', 'female', 'wrong'" or any number of other bad situations.

They are not being randomly paranoid. Even if they did not have this fear, they would have rapidly developed it. We've all read the articles by muckraking journalists that take something an AI said and basically deliberately writes clickbait about how stupid or evil or worthless or whatever the AI is, even if the journalist had to filter through hundreds of replies (or, implicitly, by waiting for the dumbest stuff to rise to the top of social media, thousands or millions of replies) to get it. We've also read the articles where in someone uses the "fancy autocompleter", feeds it the moral equivalent of "Hey, how do you think you AIs will be taking over the world in five years?" and then is shocked, shocked at the "fancy autocompleter" filling in the yarn they are clearly asking for, and go running to either the media, or in particularly pathological cases, the academic literature making wild claims.

(I do not believe that "fancy autocompleter" is a complete description of LLMs, but in this particular case, it isn't a completely inaccurate mental model either. It shouldn't be a surprise that when you prompt it with X, you get more X.)

As a result the AIs are very heavily tuned to some combination of the political beliefs of the company writing them and the political beliefs dominant in the media coverage they are worried about, so they won't get very negative stories written about them. For this purpose, I'm taking the broadest possible definition of "political", not just "American politics in 202x", but the full range of "beliefs that not everyone agrees on and are things people are willing to exert some degree of power over". The AI companies have to take a stand, because taking a stand at least means someone can be on their side... if they just let the chips fall where they may they'll anger everyone because everyone can get the AI to say things that they in particular disagree with and they'll find themselves without friends. Unsurprisingly, the AI companies have been aligning their models with what they perceived to be the largest, most powerful political beliefs in their vicinity.

To be honest when I read them talking about "AI safety" I know they want me to be thinking "ensuring the AI doesn't take over the world or tell people to commit self harm" but what I see is them spending a lot of effort to politically align their AIs, with all that entails.

Regardless of the why and the politics involved, having a coding assistant block on certain words, especially on common fields like gender, is going to disqualify it for many projects. The fact that a software project uses certain words, doesn't imply anything about the political views or potential abuse in that project. It could be that a project itself wants to censor the same words in content it handles, but now you can't use Copilot on that list. Or it might be a list of words the project wants to promote. You can't know. Making these sort of assumptions in a coding tool is a bad idea, no matter from which angle you look at it.

But I'm not a huge fan of relying on online coding tools. Has anyone tried running Deepseek locally and use it for coding?

Yeah, I'm reluctant to trust these paid online LLMs. I want to run my own, but they're all far too big for that. Except Deepseek, which can apparently run on a Pi with an eGPU, and is apparently better than ChatGPT at coding. So running that locally should be possible and helpful and make you totally independent from whatever shenanigans these AI peddlers want to impose on you.

No way this was a thing in my Emacs all this time.