Microsoft rejects documentation PR because AI chatbots can't display tables | Hacker News Reader

Microsoft rejects documentation PR because AI chatbots can't display tables

github.com

108 pointspanic30 comments

Maybe documentation shouldn’t be written for humans anymore. Maybe nothing should. Maybe we should all stop talking to one another and just communicate through AI chat bots. My chatbot will ask your chatbot if you’d like to go out tonight.

This sounds really close to what the CEO of Zoom promised back in June:

Let’s assume, fast-forward five or six years, that AI is ready. AI probably can help for maybe 90 percent of the work... I can send my digital version — you can send your digital version. Again, not like an in-person meeting.

https://www.theverge.com/2024/6/3/24168733/zoom-ceo-ai-clone...

I don't think that's what Alan Kay had in mind when he proposed replacing "data" with "ambassadors" in https://news.ycombinator.com/item?id=11945722 (Ctrl-F "ambassador" in this thread)

This, surely, must be peak AI mania. Can't make the docs readable to humans; it might upset the robots.

It does look like they've backed off, now:

> I've learned today that you are sensitive to ensuring human readability over any concerns in regard to AI consumption

I mean, you'd think that would go without saying.

the wording reads like this response was outputted by an ai.

He didn't back off; he locked the thread because people got too "uppity" and left the issue open.

Chances of it getting merged: 0%.

Bots 1: humans, 0.

Good news: he did some minor touch ups and then approved the PR. Seems like humans did win this round. Let's see how long that lasts.

Thank goodness.

Ironic that their build bot uses a table to present its validation check.

Ah yes. The two times I had to use Microsoft documentation for something, it boiled down to:

a) we've detected that your browser is set to Kazakh. Please enjoy our automatic translation of technical documentation to Kazakh. You want English? Haha that's a hidden option

b) the field "WSFaddr" contains the WSF address. Its type is "string"

Because of the above, I'll never ever use Azure again

I've always believes that MS has bad docs for azure (they can be quite good for other subjects) because they charge for support

Did you solve (a)? I found the “globe” button but have to press it for every single page which gets tedious.

Find the locale string somewhere in the URL and change it to something like en_US.

Alternatively, stop using Microsoft products. Tell your manager that their poor documentation incurs direct cost because instead of being productive, you spend your day clicking globe buttons.

I also had much trouble with this. I solved it with an extension called "FFS MSDN in English"

the way forward should be making llms understand tables better. those documentations should be written for humans, as an alternative you could have an llm friendly version

LLM could understand tables pretty well. Maybe the small local llms like Microsoft Copilot have some issues.

:rolleyes:

AI should be a tool to aid humans, not humans trying to aid AI. When you start valuing the tool over humans things start going awry rather quickly.

I'm not against AI, but I don't think we should be bending the world to make AI fit better.

In before there is a request to change the docs into pure tokens, as human language is far too inefficient for the AI to process.

He's so incredibly personally offended and passive-aggressive, it's wild. Especially when he can only accept SOME of what he's read after apparently pretending half of it was his idea!

Just reading the actual content of his (minimal) replies, he constantly avoids addressing the content of what anyone said over acting personally hurt and calling everyone "sensitive" and acting like it's arbitrary and unfair anger. I feel like that says a lot about microsoft, and a lot of Corporate types in general...

"Edited"

MS admitting their artificial "intelligence" is too stupid to understand tables, not even 20 hours after HN discussed how are developers preparing for AI to takeover our jobs. https://news.ycombinator.com/item?id=42431103

This is our future. I hate it, but it feels inevitable.

My job is going to be unrecognisable in 10 years time, to the point where I'm not sure I'll be particularly skilled to do it. And the tools I need to do my job are being made worse, just so that I can be overtaken by the next generation who are agile enough to switch to leveraging LLMs better.

The life of the programmer. Nothing new!

Just move to embedded. The tools were never good to begin with! LLMs might be useful for docs or tests, but I have found them lacking when it comes to writing firmware, especially for more niche hardware with proprietary compilers. I certainly would not use AI to write functionally safe code. I recently had an intern submit code for a simple SPI based driver that randomly called calloc (thanks CoPilot!).

I plan to leave tech by then and become a full time fursuit maker.

Is the documentation for AI chat bots or for humans?

The PR is now closed but seems like we should be striving to do better.

I'm not sure how I feel.

On the one hand I absolutely understand what all people are mad about. Docs are for humans, so LLMs understanding them shouldn't be a reason to reject something as simple as a table.

However, it also makes sense to make sure we not intentionally put things in a format that's bad for LLMs when other formats are available and just as useful for humans.

Personally, one of my favourite uses of ChatGPT is to just throw in any manpage followed by a question. If all manpages had specific sections that an LLM couldn't process I'd be a little annoyed at least.

I mean at that point why not take the (future) LLMs into acount?

to answer your question — YAGNI

i’ve seen plenty of questions like this over the years and they invariably lead to premature optimisation nightmare land.

build (or write the docs, in this case) for what’s needed today. what’s needed tomorrow is tomorrow dijksterhuis’ problem.

half the time i’ve been the one asking that question. so it’s not like i’ve been a saint in avoiding this trap.

>build (or write the docs, in this case) for what’s needed today. what’s needed tomorrow is tomorrow dijksterhuis’ problem.

But LLMs exist today...

but not everyone is relying solely on an LLM for documentation.

humans are still the primary audience for documentation.

if it ever gets to a point where no human is reading the docs, worry about it then (or close to no humans).

just because a new frontend framework gets released doesn’t mean we should all start switching to it today. it’s the same principle.

Putting aside the insanity of rejecting a docs PR on the basis that "it's better for the AI," if their docs AI can't easily express tables yet, there should be a way of telling the user "Hey, this answer contains a table; go [here](https://path.to.docs) to look at it," just like audiobooks do.

Reply on news.ycombinator.com