What AI do you use

[Fraidycat]

Since they took Fable offline I have switched over to the new 'Le Chaton Fat' model,
the worlds most powerful model at 24 Trillion parameters.

[courtg9000]

Just fired up LM Studio on my "stay at home laptop". I tried the default "Google Gemma" and used some stuff with it that I previously put through Chat GPT and it was rubbish at it. I am going to try some other LLMs as I would quite like one on my webserver and hopefully can say more soon.
I will recommend LM Studio even though I am just getting to grips with it. I think it is a cool local AI tool so far, but you will need lots of disk space for the LLM's. Ie you probably want 100GB minimum spare depending on your own config's/parameters etc and it runs alright in 16GB ram (Microsoft Surface/Win11 - fully packed and patched)

[budgetastronaut]

Mostly ChatGPT and Copilot for me. I've tried Claude as well, but I wasn't quite as impressed as the rest of the community. That said, It really depends on what you're actually using these tools for.

[courtg9000]

Just fired up LM Studio on my "stay at home laptop". I tried the default "Google Gemma" and used some stuff with it that I previously put through Chat GPT and it was rubbish at it. I am going to try some other LLMs as I would quite like one on my webserver and hopefully can say more soon.
I will recommend LM Studio even though I am just getting to grips with it. I think it is a cool local AI tool so far, but you will need lots of disk space for the LLM's. Ie you probably want 100GB minimum spare depending on your own config's/parameters etc and it runs alright in 16GB ram (Microsoft Surface/Win11 - fully packed and patched)

Ok So I tried some of QWEN LLMS and they are much better.
Still trying to heavy duty LLM that is smaller than 8GB!
LM studio is still a cool tool. https://lmstudio.ai
With 16GB of RAM it can be a little sluggish. I might have to treat myself to a laptop with 32 GB RAM for an xmas pressie!
For those of you that need to churn code out and dont want to hit the limits of Claude, this might be worth a try https://www.localcode.codes/
It has produced a fairly heavy duty app for me.
This github page is also worth a look if you don't want to pay subscription fees. https://github.com/zebbern/no-cost-ai

Hope this helps!

[NotAllThere]

Since I started this thread, I've started using Claude Code with ABAP development. Two modes - MCP for read only access to the SAP server, and via git. With git Claude can change the local clone - while syntax checking via MCP. It doesn't do git operations at all.

Still in proof of concept, but so far, when well guided, it's pretty good at finding bugs. I also used it when i needed to parse 20 or so text audit reports in JSON. I gave it the structure and explained how the JSON should generically look. I was surprised to find that first it wrote test algorithms in python, and then when they were working connected to ABAP. I find the process of instructing Claude to be pretty much the same as the thought processes I go through coding by hand.

My son uses Claude extensively. He's a C# full stack developer. He finds that his detailed prompts take half an hour to run and produce pretty good code. His colleagues' take a few minutes and produce buggy code. He says Sonnet is like a junior dev, who is very fast and pretty much does what he's told, with some bouts of over-exubherance. Opus, though, is like a junior dev who thinks he's a senior dev.

[WooHooNomad]

I use Codex on my Mac mini. On my Windows laptop it’s just not as good. Usually GPT-6 Astra on High, or Ultra for more complex tasks.

I’ve tried different approaches, but what works for me is spending time on a detailed plan. Pointing it at another project really helps, whether that’s for the repository pattern or just a design I like. I reckon it saves me 60–80% of development time, sometimes more.

At the start of the year I quoted 25 man-days for a system. Spent six hours solid planning and had it finished in four or five days. Then added features I knew they would want and padded it out to 14 days.

Give it a crap requirements doc, though, and things can go wrong. My nephew recently gave me requirements for his business. I didn’t read them properly or understand his business well enough. The AI built an incredibly overcomplicated system trying to reconcile the inconsistencies. My fault, and a useful lesson. You still need to understand what you want and the business context.

I also built a multi-tenant system with requirements so restrictive and security-conscious that performance was terrible. I’m now changing the architecture with the AI. It can make those changes in 30 minutes, but still needs pushing on performance tuning. Really surprised by that, as I thought it would be perfect for AI.

Testing is great. I get it to test iOS apps in the simulator and watch it doing it. It also generates screenshots, formats them through a third-party website and uploads them to Apple, along with filling out the masses of questions Apple asks.

I wouldn’t go back to developing by hand. Saying that, I’m semi-retired and working with smaller businesses, so I’m a bit more relaxed about things.

[sadkingbilly]

I still can't find any use outside offhost dev environments.
we don't use 'apps', git or any 'frameworks' languages, and i cant find an LLM that can read local filesystems and train.

[DoctorStrangelove]

Skynet.

[Paddy]

Using OpenCase AI, very impressive and useful for uploading and checking contracts using commands such as "find unusual or negative clauses"
Legal issues and precedents; mainly correct but the "statement of truth" is incorrect however, even some real-world lawyers still get it wrong.