Reply to: What AI do you use

[ladymuck]

Have you tried draw.io?

That rings a bell, I possibly have, I'll take another look.

[WTFH]

Interestingly, with copilot, it keeps your message history somewhere. So I was using it to find out why I couldn't find a Visio installer that didn't come with all the M365 bloatware. It sent me dud links and told me to create batch files that didn't work, and such like. I then asked what alternatives there were to Visio. It came up with some options, and said "of course you're already considering LibreOffice so this may work". I hadn't mentioned that at all in this "conversation" and asked where it got that from. Which is when it said that although I've deleted previous chats, my entire chat history is still available to it and it was referencing that.

However, as you rightly say, when you're mid chat and fully expecting the whole context of the task at hand to be remembered, it will conveniently forget.

Have you tried draw.io?

[courtg9000]

my best use of AI this week!

[BlueSharp]

candy.ai - Miss NLUK has been augmented

[ladymuck]

Tried a couple of them recently. Grok was ace for turning a photo in to a video, had some great fun with that until they hid it behind a paywall. Never got to grips with Claude and not really needed a speciality ones. The three Im using heavily are Git Co-pilot, co-pilot and chat GPT. Git with Co-pilot to do some coding for home projects, very impressive. Got what I wanted but just been continuing on as it's so easy to ask for something and it does it so having fun messing around with what was a small app in to a massive program with web interface and all that. Got some really monotonous tasks to do at my current client so using co-pilot to add a ton of VB code to word and excel to automate which has been great fun and saved me hours. ChatGPT just because I added the app many moons ago and am just used to using it.

One thing I have noticed when using them heavily for the coding and work stuff is how inaccurate it can be and if it's a chain of work how often it loses the chain. Gets very frustrating at times having to remind it and question its results to which it says 'You are right to point that out and are correct because.....' so still have to do a lot of double checking and have half a clue what it's doing for me rather than relying on it. Light or home users think it's the best thing ever and it's perfect but start hammering it and the cracks start appearing.

Interestingly, with copilot, it keeps your message history somewhere. So I was using it to find out why I couldn't find a Visio installer that didn't come with all the M365 bloatware. It sent me dud links and told me to create batch files that didn't work, and such like. I then asked what alternatives there were to Visio. It came up with some options, and said "of course you're already considering LibreOffice so this may work". I hadn't mentioned that at all in this "conversation" and asked where it got that from. Which is when it said that although I've deleted previous chats, my entire chat history is still available to it and it was referencing that.

However, as you rightly say, when you're mid chat and fully expecting the whole context of the task at hand to be remembered, it will conveniently forget.

[northernladuk]

Reminds me of outsourcing...

and offshoring... but the work to reverse both that was forecast never materialised.

[northernladuk]

Tried a couple of them recently. Grok was ace for turning a photo in to a video, had some great fun with that until they hid it behind a paywall. Never got to grips with Claude and not really needed a speciality ones. The three Im using heavily are Git Co-pilot, co-pilot and chat GPT. Git with Co-pilot to do some coding for home projects, very impressive. Got what I wanted but just been continuing on as it's so easy to ask for something and it does it so having fun messing around with what was a small app in to a massive program with web interface and all that. Got some really monotonous tasks to do at my current client so using co-pilot to add a ton of VB code to word and excel to automate which has been great fun and saved me hours. ChatGPT just because I added the app many moons ago and am just used to using it.

One thing I have noticed when using them heavily for the coding and work stuff is how inaccurate it can be and if it's a chain of work how often it loses the chain. Gets very frustrating at times having to remind it and question its results to which it says 'You are right to point that out and are correct because.....' so still have to do a lot of double checking and have half a clue what it's doing for me rather than relying on it. Light or home users think it's the best thing ever and it's perfect but start hammering it and the cracks start appearing.

[NotAllThere]

The good thing about AI is there will be loads of work in the future fixing buggy and downright failed stuff with fewer people able to do the fixing due to dumbed down AI skills.

Reminds me of outsourcing...

[willendure]

The problem is it only knows what's out there and most of that is garbage

Thats why you give it what it needs to know. Want to know how to program a MikroTik router ? Give it the docs, then ask it.

[tazdevil]

The good thing about AI is there will be loads of work in the future fixing buggy and downright failed stuff with fewer people able to do the fixing due to dumbed down AI skills. I asked Google recently for a summary of a large local authority contract that failed but it thought it was brill and still on the go as it was and totally missed that yours truly had picked up the pieces (but not the big money unfortunately as that had been squandered) It also coughed my own words back at me when asking about what I did, mixed in some other stuff and a few falsehoods for good measure. The problem is it only knows what's out there and most of that is garbage

[sadkingbilly]

TL: DR

aye, very good.
I'm sure you, and your Algorithmic Intimate will go far.
I've seen no particular use for it in my field, except maybe scanning docs, but even that's beyond some of the so called 'AI's.

[willendure]

and it'll recover your £1800 in a few weeks? really?
fine.

I have no intention of recovering the £1800 - actually more like £450 because I've only been on it for a few months, and will scale back to a lower tier soon anyway. Its money that I wanted to spend to develop this new compiler because its a project I believe in and am personally invested in.

But consider this, if I really have done 1.5 to 2.5 years of work, which is not an unreasonable claim, and I have only put in 0.5 years of my own time (which adds up to a lot more lost income as a contractor than the £1800), then to do this by employing someone to do the extra 1 to 2 years work would have cost me how much ? At least £100K, maybe £200K. So when you put it like that, I have indeed got a bargain.

Last contract I worked with a bunch of useless offshore types - there was 1 sprint where 4 of them wrote 3 lines of code between them all - I guess that was 1 line each and one guy just totally out to lunch! Of course these were scrape-the-barrel offshore bums on seats so towards the cheap end. I would rather work with, and could accomplish more with, a £20/month AI subscription! and it would be less hassle to boot!

Part of the reason for the maxxed out AI subscription though is that I want to get the first release of this compiler out soon, before they start charging the real cost of AI, because it almost certainly is being sold at "loss leader" prices today to get everyone dependant on it. Also I need to get it released so I can find some paid work.

Part of the reason for doing this was also for me to skill up using AI for software development, which I can now confidently claim to be an expert in. Its obvious to me that this is so new, no-one really has the answers yet for how to do everything and create a good result on large software engineering projects. I have had to invent my now techniques even, and experiment a lot with it to undertand how to use it. War stories like this are invaluable in an interview to show that you have real hands on experience of the types of problems that arise that are unique to AI coding.

Its been an epic challenge working on something like this, and finding out what AI can do, and proving by doing that it really can develop something as sophisticated as a compiler - the sort of work that PhD students might do. Part of the reason for doing this is also that I turned 50 this year, I rarely get to do genuinely interesting work as a contractor, so I get to prove to myself that I am still capable of doing this kind of work.

[ladymuck]

A friend of mine used a lot of AI to challenge something with Islington council who co-own / manage the flats where she lives. She owns her flat and so there's maintenance charges, as one would expect. However, the council were taking the proverbial a bit it would seem and so she used an AI tool to help dig into the detail and find all the right things to make their life a misery. No idea if the cost of the credits outweighs any financial savings...

[hobnob]

Using AI for writing a Skeleton Argument for a county court case, it has saved about two days work with a BIG however, six out of eight precedents were incorrect or non existent, wrong section number for CPR (regulations), statement of truth was the outdated version.

Like any AI, it needs checking otherwise, egg on face.

Here's a good example of what happens when you don't check:
Malcolm Cork & Anor v Smith [2026] EWHC 1199 (Ch) (22 May 2026)

Short version:
* Legal company applied to a court, asking them to do X.
* Judge wrote back, asking "Why do you think that I have the power to do X without involving the secretary of state?"
* Legal company: "You can do it because of this piece of legislation [..]"
* Judge: "That rule doesn't say anything like that."
* Legal company: "Ah, just to clarify, we weren't quoting a specific rule, we were summarising. Sorry for any confusion."
* Judge: "You're lying. Did you use AI?"

It turned out that a junior lawyer had a 59 page chat with an LLM and took the output at face value, even when the LLM said "This might be wrong, you really need to double-check it." Their boss approved the letter without knowing about the LLM and without checking the info. The big boss then did the same thing with a follow-up letter.

The judge has "publicly admonished" the law firm (name and shame), and I suspect this won't look good in the junior lawyer's performance review...

[Fraidycat]

I have been pair programming with Claude:

Claude is given the driver/typist role.

While I take the Navigator role and focus on looking out for logical flaws, edge case errors, high-level architecture and overall goals for the codebase.


The Clankers tend to do a good first pass on a brand new code file, but after a while the code can decay into a mess if you don't keep a close eye on them.