Polite AI

[WooHooNomad]

I use Claude. Today I used Claude Code to finish off organising my photos. I'd start a few years back, putting them into folders of YYYY-MM, but never got back to it. Claude Code did the entire job within about half an hour.

I use Codex with the GPT-6 Astra model set to high reasoning for coding tasks. I bump it up to ULTRA when I’m planning development work. Just like before AI, the more thought you put into the planning, the better the end result. So, I guess that is providing a better prompt in the end.

I haven’t properly tried Claude yet, but I might give it a go now that OpenAI has reduced my token usage on the £200 plan.

[sadkingbilly]

well its cool for code monkeys and document faffing, but no use to me.

[jamesbrown]

Claude is definitely usable and very useful, not least because it has a VM/sandbox where it can run code and hence perform unit/integration test, so it doesn't delivers code that is broken - might not deliver what you intended, which may be because the prompts were poor/ambiguous or because it made flawed choices - but it won't deliver code that is broken in elementary ways, unlike all the free AIs, which don't test anything before delivering it. They are all still over-confident, but the free AIs are miles behind the latest subscription models - 6 months is a long time in AI dev.

[NotAllThere]

I use Claude. Today I used Claude Code to finish off organising my photos. I'd start a few years back, putting them into folders of YYYY-MM, but never got back to it. Claude Code did the entire job within about half an hour.

I used the Claude desktop to build a 3D model of the theatre were using for the Panto this year. I fed it details of the lighting set up, 2D plans of various parts of the auditorium and stage. It created the necessary data in various formats to import into ETC/EOS.

It's done a pretty good job given the information at hand. Very much usable.

For work I use Claude Code for programming, like when I had 20 reports of similar but sufficiently different formats and needed to parse them into JSON. It went away and created the parsing classes. It spotted some bugs that had been lurking for a few years. It's also build a few prototypes using Power Automate/Teams/Sharepoint.

Using an MCP for ABAP (ARC-1), it's pretty good at analysing issues on our code base as they come in. Sometimes needs a few pointers.

If you create good prompts, then you get good output. The problem is to create good prompts you need to be an experienced senior dev. If AI is doing the junior dev roles where do we get senior devs from?

Copilot (free consumer version) assured me five times yesterday it could creat a docx file of the information I'd asked it to find. Before finally admitting that it can't. "I definitely can create a DOCX, but I just need to know one last thing", it said about three times...

[WooHooNomad]

I've only used the stuff clients have allowed, the usual Microsoft junk. Took me ages to craft a prompt that actually produced a non-hallucinated response. It was quicker to print out two documents and read them side by side with a highlighter to mark the differences.

The fluffy language every time I pointed out an error or a baseless assumption drove me potty.

Yeah i'm having issues with tasks that required the AI to input my data and register details. Now it wants me to fill out half the data and it blames itself for not informing me. Annoying as hell. I've found spreadsheet comparisons to be ok in the latest models, never really compared docs so can't help. Perhaps we need AI's to interact with AI's

[ladymuck]

Which models are you using, just out of interest? I also feel personally attacked by your last statement

I've only used the stuff clients have allowed, the usual Microsoft junk. Took me ages to craft a prompt that actually produced a non-hallucinated response. It was quicker to print out two documents and read them side by side with a highlighter to mark the differences.

The fluffy language every time I pointed out an error or a baseless assumption drove me potty.

[cojak]

Because AI doesn't understand what it's saying. It constructs sentences based on probability and context. If you actually want it to compare two documents and highlight the exact changes then your prompt needs to explicitly tell it to only use the source information you specify, not to add detail that doesn't exist, and not make assumptions that are not supported by the contents of the V1 and V2 documents. If you're going to give it a lazy prompt, you'll get a lazy response.

The language it uses (all sycophantic and ego fluffing) is purposely designed to make the inept feel like they're gods.

That just sounds like a natural language request of PL/1.

Or am I missing something?

[DoctorStrangelove]

I also feel personally attacked by your last statement

[sadkingbilly]

Well i haven't yet found any that are any use to ME.

[WooHooNomad]

Because AI doesn't understand what it's saying. It constructs sentences based on probability and context. If you actually want it to compare two documents and highlight the exact changes then your prompt needs to explicitly tell it to only use the source information you specify, not to add detail that doesn't exist, and not make assumptions that are not supported by the contents of the V1 and V2 documents. If you're going to give it a lazy prompt, you'll get a lazy response.

The language it uses (all sycophantic and ego fluffing) is purposely designed to make the inept feel like they're gods.

Which models are you using, just out of interest? I also feel personally attacked by your last statement

[WooHooNomad]

I would also say which model you use matters. Some of the dumber models are frustrating but have their uses.

[sadkingbilly]

Algorithmic Inference. - bah!

[ladymuck]

Because AI doesn't understand what it's saying. It constructs sentences based on probability and context. If you actually want it to compare two documents and highlight the exact changes then your prompt needs to explicitly tell it to only use the source information you specify, not to add detail that doesn't exist, and not make assumptions that are not supported by the contents of the V1 and V2 documents. If you're going to give it a lazy prompt, you'll get a lazy response.

The language it uses (all sycophantic and ego fluffing) is purposely designed to make the inept feel like they're gods.

[tazdevil]

Had a conversation with Google AI asking about an API and it provided me with wrong information. I was trying to find out where the location information had gone which was present in API v1 but not v2. So I told it the info was wrong and it came up with a bit more b*llocks until finally admitting I was right and that it had rescanned the documentation and apologised for guessing the answer up till then! Why have the developers built AI to guess rather than do its homework?