Chat GPT & Generative AI tools

[firlandsfarm]

IMO Claude Code (anthropic) is the go for coding atm. It's good because it has access to your filesystem. (Some people might care about privacy...I don't...)

You obviously wanna know what you are doing or it'll make a mess...but if it fks up, ya just revert and try again... It's exceptionally good at writing things from scratch, but gets less good as your app gets more complex or ambiguous. One ends up spending their time documenting coding standards, rather than actually coding....then it gets a bit boring repeating, 'follow this directive you...suck...get it right... ok thx'

But still 50% faster than doing it yourself....

ChatGPT is cheaper because one quickly ends up on the max plan with anthropic.

Edit: The real advantage of these tools is learning... learning to code is never just the syntax, it's the patterns and practices and architectures that REALLY take ages to learn.......IMO one would do well to short places like Pluralsight for learning those things.... that used to be the goto... LLM's do it better.

The problem I've found with using it for coding is stopping it running away with itself. l guess it thinks it's being helpful by designing something from scratch through to finished product without asking you if it's on the right track and more worrying without checking not just that it's good but making sure you understand what it doing and how. I find I have to keep asking it to explain what its doing and dragging it back to what I want. Yeah, it's ideas may be right or they may not and I want to be able to check which is which. I think getting it to code without explanation is very dangerous.

Thanks for the Claude tip, maybe I should check it out.

[Big Bad Barney]

The problem I've found with using it for coding is stopping it running away with itself. l guess it thinks it's being helpful by designing something from scratch through to finished product without asking you if it's on the right track and more worrying without checking not just that it's good but making sure you understand what it doing and how. I find I have to keep asking it to explain what its doing and dragging it back to what I want. Yeah, it's ideas may be right or they may not and I want to be able to check which is which. I think getting it to code without explanation is very dangerous.

Thanks for the Claude tip, maybe I should check it out.

In my corporate day job I was using ChatGPT and cut pasting a lot...(AI policies and privacy concerns, etc)......for my own stuff Claude is the go...100x better....

Claude WILL run away from you also if you let it....it's solution to that is plan mode...you switch in and out of it it...in plan mode it won't do anything but plan and tell you what it's going to do.... then at the end you tell it yay or ney.......you could if you desired, just stay in plan mode and learn all day long.......but yeah...my experience is the same....the challenge is keeping it consistent, keeping it on track........

The way I get around that is document coding styles and architecture examples and EVERY PROMPT, I tell it about the relevant important ones.....because it does forget... even when you setup pre/post command hooks, etc......it's too much info for it to cope....

[firlandsfarm]

In my corporate day job I was using ChatGPT and cut pasting a lot...(AI policies and privacy concerns, etc)......for my own stuff Claude is the go...100x better....

Claude WILL run away from you also if you let it....it's solution to that is plan mode...you switch in and out of it it...in plan mode it won't do anything but plan and tell you what it's going to do.... then at the end you tell it yay or ney.......you could if you desired, just stay in plan mode and learn all day long.......but yeah...my experience is the same....the challenge is keeping it consistent, keeping it on track........

The way I get around that is document coding styles and architecture examples and EVERY PROMPT, I tell it about the relevant important ones.....because it does forget... even when you setup pre/post command hooks, etc......it's too much info for it to cope....

I'm in the middle of something just now, Phase 2 is about to start ... maybe I'll give Claude a go and see how it handles it. Thanks for the suggestion.

[firlandsfarm]

I'm in the middle of something just now, Phase 2 is about to start ... maybe I'll give Claude a go and see how it handles it. Thanks for the suggestion.

OK, took Claude around the block as a 'Pro' and initially liked it.

It was fast and easier to follow than ChatGPT but it was just as verbose and continually repeated itself if I queried a point in a response by answering my point but then repeating everything from the earlier response as if I hadn't noticed them (I prefer to pick up one point at a time, with ChatGPT I find if my response to it gets too long it seems to forget to answer some of my issues}.

But then came the shock of being told I had run out of allowed conversation! On querying what the problem was the 'token' system was explained ... it seemed to me Claude had used most of the tokens with it's verbosity! The irony (stupidity?) was it ended the conversation when producing a summary of three long responses and there wasn't enough tokens left to complete the summary.

Unbelievably I was told to start a new convo. and summarise the previous convo by manually copying sections from the old into the new! I started the new convo by copying the whole previous convo and pasting it as the opening prompt for the new convo requesting a summary which it performed satisfactorily. (I must say I prefer it's layout of a summary and it's use of colour and in this context I appreciate it's verbosity by producing longer, more detailed summaries than ChatGPT which I find a little too short sometimes.)

I discussed the inefficiency of the curtailing of a conversation only for it to be summarised and reposted into a new conversation for continuity of content. It agreed with me! Maybe I had flogged it into a mindset of "let's agree with this guy and get rid of him"! Anyway that's how I will be using Claude in the future ... copy a related previous convo and add it into the new convo. maybe as a file rather than pasting thinking pasting will use more of the 'tokens'.

[firlandsfarm]

I'm in the middle of something just now, Phase 2 is about to start ... maybe I'll give Claude a go and see how it handles it. Thanks for the suggestion.

One thing I found after my previous post is that on reverting to a related ongoing ChatGPT conversation ChatGPT was fully aware of my Claude conversation in fact it opened the posting window already filled with the Claude chat as if I had copied and pasted it myself! When using Claude I had noticed in some of the things it said that it seemed to be aware of my ChatGPT conversations. I'm not sure if I should be impressed by their apparent sharing of conversations (is it all in the cookies?) or concerned that no conversation with them is truly private!

[wearthefoxhat]

I'm in the middle of something just now, Phase 2 is about to start ... maybe I'll give Claude a go and see how it handles it. Thanks for the suggestion.

When using Claude I had noticed in some of the things it said that it seemed to be aware of my ChatGPT conversations. I'm not sure if I should be impressed by their apparent sharing of conversations (is it all in the cookies?) or concerned that no conversation with them is truly private!

I'd lean into, "that no conversation is truly private."

Extend that a little to someone developing a prediction strategy. The LLM's will use elements of, if not all, to assist you in developing your own, the trade off is, that it will then share your final work with others. Your own strategy could be untested, but it's gonna get intergrated with someone else's, the trick is, to sift through the chaff for the wheat, then you know your winnowing.....

[firlandsfarm]

When using Claude I had noticed in some of the things it said that it seemed to be aware of my ChatGPT conversations. I'm not sure if I should be impressed by their apparent sharing of conversations (is it all in the cookies?) or concerned that no conversation with them is truly private!

Extend that a little to someone developing a prediction strategy. The LLM's will use elements of, if not all, to assist you in developing your own, the trade off is, that it will then share your final work with others. Your own strategy could be untested, but it's gonna get intergrated with someone else's, the trick is, to sift through the chaff for the wheat, then you know your winnowing.....

I understand the trade-off but I guess I wasn't aware of the sharing ... makes me wonder if they are all sharing from the great big melting pot. I was surprised the sharing seemed to be so direct.

[Big Bad Barney]

I understand the trade-off but I guess I wasn't aware of the sharing ... makes me wonder if they are all sharing from the great big melting pot. I was surprised the sharing seemed to be so direct.

You can turn sharing/learning off in both ChatGPT and Claude. At least I am fairly certain you can. It's just not labeled as...'Do not share any of my stuff with anyone' since it's a bit nuanced...

I'm a bit surprised they shared that quick too!

[Big Bad Barney]

OK, took Claude around the block as a 'Pro' and initially liked it.

It was fast and easier to follow than ChatGPT but it was just as verbose and continually repeated itself

Yep it does that...it can be painful... I wouldn't say it's any 'better' than ChatGPT in terms of what it does or returns....what I fundamentally like about it is access to ones filesystem...(as far as I am aware chatgpt doesn't and isn't like to offer that any time soon)

It can connect to all kinds of other servers/services... I haven't given much crap about that though....

But yeah...I'm on the max plan, and it chews that baby up...especially if you use Opus instead of Sonnet...

[firlandsfarm]

I don't think Claude is for me in it's current form ... shame really because I liked it (compared to ChatGPT). But I can't be having the "end of conversation" lark! Some of my discussions are pretty long, currently I'm on my 3rd chat discussing the legal aspects of a website I am building and it will probably extend into more chats. The trouble is Claude doesn't know what has been discussed in the previous chats. I've called for a summary to load into the new chat but that didn't help. I think Claude would be great for quick answers to individual questions but not a full length project.