AI Development general thread.

Need to do some more research but a mate told me about claude terminal and how it uses it daily to automate some basic research task and to stream line his calendar. Think you need the $17 dollar subscription and then you download the terminal but apparently its pretty good, here is also an intresting but older video.

 
Lately, I’ve been using both ChatGPT and Gemini, and honestly, Gemini outshines GPT when it comes to development, coding, and debugging. Yesterday, I spent hours trying to fix issues with an app using ChatGPT. This morning, I switched to Gemini, and everything went smoothly, just got things done in a fraction of the time compared to GPT, which is ironic because theoretically OpenAI's new 5.2 model is "smarter" and "better" at coding than Gemini 3.
 
konfab said:
Yup. OpenAI et al have poured far more money into training these models to be good at coding than solving arbitrary word puzzles themselves.
Click to expand...


So far I could use AI for some useful things. But mostly it is like a glorified search engine. It's also an analysis and decision making tool.
 
MnrWerkswinkel said:
So far I could use AI for some useful things. But mostly it is like a glorified search engine. It's also an analysis and decision making tool.
Click to expand...
Things I have done this week with Claude Sonnet 4.5 and Opus 4.5
1) Migrated our application away from Webpack to Bun
2) Upgraded our component library
3) Swapped out an unmaintained and broken tabling library with one that is being actively maintained.
4) Writing more comprehensive integration tests that test all the above.

But yeah, it is just a glorified search engine.
 
konfab said:
Things I have done this week with Claude Sonnet 4.5 and Opus 4.5
1) Migrated our application away from Webpack to Bun
2) Upgraded our component library
3) Swapped out an unmaintained and broken tabling library with one that is being actively maintained.
4) Writing more comprehensive integration tests that test all the above.

But yeah, it is just a glorified search engine.
Click to expand...

Been working on a distilled model with opencode with Opus and Codex as a review agent to help me understand the processes in place.

Did a few other things as well. Not very entrepreneurial since I still haven’t built a SaaS to sell to the masses
 
konfab said:
Yup. OpenAI et al have poured far more money into training these models to be good at coding than solving arbitrary word puzzles themselves.
Click to expand...

If you have the $$ try Pro model of GPT. It seems more human like. Similar to Gemini Thinking bit better. Just painfully pricey
 
Pho3nix said:
If you have the $$ try Pro model of GPT. It seems more human like. Similar to Gemini Thinking bit better. Just painfully pricey
Click to expand...

I honestly couldn't really be bothered. I will try it if it is GH Copilot's model selection and that is it.
 
Vorastra said:
What vision model that's tiny (12GB VRAM Blackwell, NVFP4 would be bae) do you okes recommend?

I'm looking for something I can run images through, and then spit out tag-like descriptors.

I'm seeing Qwen2.5-VL-7B is apparently quite good and comes in NVFP4!
Click to expand...
Not AI (what is really anyway), but OpenCV with models work well. Haar cascade models are fast though an older method. Also runs on next to nothing hardware (a 10 year old Android TV box would be powerful enough).
 
Vorastra said:
What vision model that's tiny (12GB VRAM Blackwell, NVFP4 would be bae) do you okes recommend?

I'm looking for something I can run images through, and then spit out tag-like descriptors.

I'm seeing Qwen2.5-VL-7B is apparently quite good and comes in NVFP4!
Click to expand...
Yolo models work quite well

https://github.com/ultralytics/ultralytics

If you have an iPhone, you can run YoloV8 on it.
https://t.co/HYqmsxtE6D

For what you are wanting to do, an LLM is overkill.
 
Vorastra said:
What vision model that's tiny (12GB VRAM Blackwell, NVFP4 would be bae) do you okes recommend?

I'm looking for something I can run images through, and then spit out tag-like descriptors.

I'm seeing Qwen2.5-VL-7B is apparently quite good and comes in NVFP4!
Click to expand...

Check huggingface. You should find a model that does exactly this
 
Pho3nix said:
Check huggingface. You should find a model that does exactly this
Click to expand...
I used EfficientNetV2B2 as a base pre trained image model and employed transfer learning for my use cause case. Specifically around plant pathogen detection/classification. My additional images used in the learning was in the regions of around 200K labeled plants, down to pathogen level.
 
semaphore said:
I used EfficientNetV2B2 as a base pre trained image model and employed transfer learning for my use cause case. Specifically around plant pathogen detection/classification. My additional images used in the learning was in the regions of around 200K labeled plants, down to pathogen level.
Click to expand...

Trying to do something similar but with Llama at the moment. No training data so have to do it all myself which is a pain. Let me give EfficientNet a look
 
Which code based agents are folks using? Running some experiments with local cli powered agents for Claude and Codex. Codex allows us to also plug directly into our own models in Foundry which is a plus I reckon.

Not talking about ide helpers but actual standalones where you can set environment and give it full repo access to work with and on what you stipulate.
 
austenallred.gumroad.com

Automate Everything: The OpenClaw Handbook

Automate Everything: The OpenClaw HandbookShort Description:Master AI-powered automation with OpenClaw. From setup to advanced workflows, this 400-page handbook teaches you to automate anything.---## 🤖 AUTOMATE EVERYTHING: THE OPENCLAW HANDBOOKStop doing repetitive work manually. Learn to build...
austenallred.gumroad.com austenallred.gumroad.com

Openclaw handbook how to set up your own model +- 200 pages, going to start with this tonight and see how it goes
 
L-Dog said:
austenallred.gumroad.com

Automate Everything: The OpenClaw Handbook

Automate Everything: The OpenClaw HandbookShort Description:Master AI-powered automation with OpenClaw. From setup to advanced workflows, this 400-page handbook teaches you to automate anything.---## 🤖 AUTOMATE EVERYTHING: THE OPENCLAW HANDBOOKStop doing repetitive work manually. Learn to build...
austenallred.gumroad.com austenallred.gumroad.com

Openclaw handbook how to set up your own model +- 200 pages, going to start with this tonight and see how it goes
Click to expand...
Yeah, just be super careful with it.

I was reading about the whole meta ai head and openclaw, but what she didn't say was WHICH models she was using that wiped out her email. Bet it was meta's.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X