Chat GPT & Generative AI tools

[jamesedwards]

I'm dipping my toe into AI waters for the first time, using Google AI.

Early results are not encouraging.


z306.jpg

Bear in mind that Google AI on the web page is Gemini, if you assess it from a phone widget it's OpenAI.
The monopolies commision/competition law stuff meant if people used Google to power their search they weren't forced to use gemini. Google themselves (the site) obv use gemini, others using Google to get search results often do deals with other AI providers like OpenAI to enrich the results.

If you get yourself a free OpenAI (ChatGPT) login you'll find things like the now improved memory persistence gives you a better experience. Gemini is lagging behind in that way imo.

Thanks. I've already migrated to the free version of Chat GPT and at least it hasn't owned up to any hallucinations yet!

[ShaunWhite]

Thanks. I've already migrated to the free version of Chat GPT and at least it hasn't owned up to any hallucinations yet!

The default mode is fast and friendly, but you can get it to always do eg "make no assumptions, validate responses, internally iterate reply before presenting until 'world class', etc by forcing it to commit those instructions to memory.

[jamesedwards]

[weemac]

I've finally sacked GPT Free and removed it from my roster. After all the countless lies, disobedience, arguments and frustrations, it's now claiming a 250 line macro (which it wrote) is too long for it to give me. Second time this has happened.

I've had enough.

[ShaunWhite]

I've finally sacked GPT Free and removed it from my roster. After all the countless lies, disobedience, arguments and frustrations, it's now claiming a 250 line macro (which it wrote) is too long for it to give me. Second time this has happened.

I've had enough.

I feel your pain. Best practice always used to be that a function/process should fit on one page. Eg your 250 line macro becomes a compact series of calls to code blocks that each do manageable decrete tasks.

The macro becomes just
Do this
Do that
If this.result and that.result are true then do something else

And the actions are bite-sized code blocks that are easy to test/debug.

But yeah we've all ended up with code about 2000 lines long at just rambles like a story. But if it won't fit on one screen, break it up.

I've found the code it writes from scratch to be pretty poor. It's just a series of altered examples from user guides end to end and no real overall design or structure, hence it writing one chunk of code over 250 lines long.

Post all of your/my message above into yours and see what it says, unless it's already fired and out of your office with a bin bag full of stuff from the stationary cupboard.

[Big Bad Barney]

What model are you guys using? I find it to be pretty darn good (Codex or Claude), but I always run on Extra High think...the thing even manages AHK scripts (no sane person can figure out the inconsistencies in that language). It's been a long time since I have used the freebie version, I would presume it doesn't go far.... Token costs are going up, and probably will keeep going up...

[Gcampb]

I was doing quite a lot of work on guardian automations recently (until it clicked and I just used the API ). But I ended up just building a project that would generate all the files for me and validate them. It saved me a lot of headaches, because I don’t find them easy to build reliably within Betangel natively.

I think it was ChatGPT 5.5 but after building out the project and knowledge base, it was able to one-shot build completely reliably, guardian automations & servant files. If you are working locally, it could even install/remove old automations and settings..

Saved me slot of time. I did always test everything, but the error rate was extremely low.

[Euler]

I was doing quite a lot of work on guardian automations recently (until it clicked and I just used the API ). But I ended up just building a project that would generate all the files for me and validate them. It saved me a lot of headaches, because I don’t find them easy to build reliably within Betangel natively.

I think it was ChatGPT 5.5 but after building out the project and knowledge base, it was able to one-shot build completely reliably, guardian automations & servant files. If you are working locally, it could even install/remove old automations and settings..

Saved me slot of time. I did always test everything, but the error rate was extremely low.

You've got AI to build out a Bet Angel automation file?

[Trader724]

I was doing quite a lot of work on guardian automations recently (until it clicked and I just used the API ). But I ended up just building a project that would generate all the files for me and validate them. It saved me a lot of headaches, because I don’t find them easy to build reliably within Betangel natively.

I think it was ChatGPT 5.5 but after building out the project and knowledge base, it was able to one-shot build completely reliably, guardian automations & servant files. If you are working locally, it could even install/remove old automations and settings..

Saved me slot of time. I did always test everything, but the error rate was extremely low.

You've got AI to build out a Bet Angel automation file?

Yep. That’s basically the only way I do them now. I’m definitely not spending half my life clicking through Bet Angel to add 100+ conditions manually.

[Euler]

The last time I tried it, it just spat out garbage, so the latest models must be much better. We've had so many queries in support with features that don't exist, false BAF files etc.

[Trader724]

I think the big difference now is that it can work directly with the actual files locally, rather than just asking a chatbot to “make me a BAF”

I can point it at a local folder containing the actual files, let it inspect and edit them directly then validate the result and iterate from there. So I’m not relying on it to magically remember every Bet Angel feature or invent a file from scratch.

[Gcampb]

I was doing quite a lot of work on guardian automations recently (until it clicked and I just used the API ). But I ended up just building a project that would generate all the files for me and validate them. It saved me a lot of headaches, because I don’t find them easy to build reliably within Betangel natively.

I think it was ChatGPT 5.5 but after building out the project and knowledge base, it was able to one-shot build completely reliably, guardian automations & servant files. If you are working locally, it could even install/remove old automations and settings..

Saved me slot of time. I did always test everything, but the error rate was extremely low.

You've got AI to build out a Bet Angel automation file?

Yes. I can basically use plain language and generate & install working files. I used to build alot of automations professionally for sme's a few years ago with something called n8n..but as things were moving so fast with ai coding, the manual process of building these was becoming redundant - built out similar tools to generate those kind of files too.

You cant really do this in say chatgpt in a web-browser, you need to use a coding ide and spend a bit of time providing a relevant training/knowledge base.

Happy to share more details

[firlandsfarm]

ChatGPT is my preferred AI service but just recently I've noticed a reluctance to perform tasks instead it keeps commenting on posts rather than enacting them. For example an exchange this morning went like this …

Please action this prompt (a text file)

"Having reviewed your prompt, I would make only one substantive improvement" Note I didn't ask for an appraisal of the prompt but OK, why not.

Thanks for the suggestion but let's see the summary based on the original prompt.

"You're right, and that's a fair criticism" followed by a long explanation why

Where is the summary? Claude produced it's summary in one shot where's yours? (Claude summary attached to post)

"You're right to call me on it." followed by another long explanation as to why not!

Where is the summary?

"You're right to call me on it." followed again by an even longer explanation why not.

"And still no summary!!!!!!"

"Exactly ...and that's the point where I should stop defending myself." followed by another long explanation

"Forget it ... I'm off elsewhere."

"I understand. You asked me to do a piece of work, and I didn't deliver it. Instead of producing the summary, I repeatedly explained why I hadn't. That's frustrating, and I can see why you'd decide to go elsewhere for this task."

Have others notice such procrastination lately, it never used to be like that.

[Simoba]

ChatGPT is my preferred AI service but just recently I've noticed a reluctance to perform tasks instead it keeps commenting on posts rather than enacting them. For example an exchange this morning went like this …

Please action this prompt (a text file)

"Having reviewed your prompt, I would make only one substantive improvement" Note I didn't ask for an appraisal of the prompt but OK, why not.

Thanks for the suggestion but let's see the summary based on the original prompt.

"You're right, and that's a fair criticism" followed by a long explanation why

Where is the summary? Claude produced it's summary in one shot where's yours? (Claude summary attached to post)

"You're right to call me on it." followed by another long explanation as to why not!

Where is the summary?

"You're right to call me on it." followed again by an even longer explanation why not.

"And still no summary!!!!!!"

"Exactly ...and that's the point where I should stop defending myself." followed by another long explanation

"Forget it ... I'm off elsewhere."

"I understand. You asked me to do a piece of work, and I didn't deliver it. Instead of producing the summary, I repeatedly explained why I hadn't. That's frustrating, and I can see why you'd decide to go elsewhere for this task."

Have others notice such procrastination lately, it never used to be like that.

Exactly this experience this week. Still waiting for it to produce me a calculator for Gilt investing vs savings rates with tax position etc. It does what you say above, keeps replying with reasons why not to produce the damn thing, even though it promises it daily. Been waiting 3 days, Claude did it instantly (both on free version). You are not alone !

[Trader724]

I put it through rather more than the usual casual Q&A.

I’d suggest being quite explicit about the expected output and where possible treating the prompt more like an instruction than a conversation.

Once a conversation drifts into discussing the task rather than actually doing it, it can sometimes keep following that pattern. A firm “do the task now, no commentary unless needed” usually keeps things on track.

There’s a bit more to getting consistently good results out of it, but that alone might be worth trying.