AI Development general thread.

[L-Dog]

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.

Yeah reformatted my old gaming pc and in a few discord channels where people are messing with AI. First tip is to create its own email and not give it access to any of your personal files. Seen many posts about it deleteing peoples emails or messing up their calendar. Some might be fud but just going to give it viewing access where I can on personal stuff

 

[Kosmik]

Yeah reformatted my old gaming pc and in a few discord channels where people are messing with AI. First tip is to create its own email and not give it access to any of your personal files. Seen many posts about it deleteing peoples emails or messing up their calendar. Some might be fud but just going to give it viewing access where I can on personal stuff

Sandboxing is best, especially if you still learning to ring fence it.

 

[L-Dog]

Yeah guide is slightly outdated or just doesn't work on windows 10. Might try again tomorrow or just use gemini / chatgpt to set up a different model

 

[Kosmik]

Gemini/chatgpt will require api access. Lots of folks use Olama and local llms rather.

 

[L-Dog]

Yeah I see i just had to adjust setting in my bios now I can upgrade to the windows 11. Will try model again tomorrow. Still not sure if I should go openclaw or olama route just don't want to burn money which might be issue with openclaw

 

[Kosmik]

Yeah I see i just had to adjust setting in my bios now I can upgrade to the windows 11. Will try model again tomorrow. Still not sure if I should go openclaw or olama route just don't want to burn money which might be issue with openclaw

If I undetstand it right , and thats just based on research vs actual use, OpenClaw is just a connector broker. So it creates local connections on the machine via whatever you allow it but its powered via api billing based connections to your respective llm option. The closed models and openclaw from what I read don't allow integrated connection, you must use api billing keys which are more expensive in comparison. For use as a "pa", I would go with a lower local hosted model. Cheapest.

Based on a chatgpt breakdown I was looking into, it said this:

Available

General

Model Family	Typical Sizes	Notes
Llama2	7B, 13B, 70B	Standard general-purpose open models. (Ollama)
Llama3	8B, 70B	Modern “instruct + reasoning” models. (Ollama)
Mistral	7B	Efficient general model. (Ollama)
Mixtral	8×7B, 8×22B	Mixture-of-Experts models (larger, MoE configs). (Ollama Operator)
Phi-2 / Phi	2.7B, 14B	Microsoft open model family. (Ollama)
Orca-Mini	3B	Lightweight “mini” model. (Ollama)
Gemma	2B, 7B	Google’s open models. (Ollama)

Coding

Model	Sizes	Notes
CodeLlama	7B, 13B, 34B, 70B	Code generation / reasoning models. (Ollama)
Qwen-Coder / Qwen3-Coder	various (e.g., 30B+)	Larger coding-focused series from Qwen. (Ollama)

Multi/Vision

Model	Sizes	Notes
LLaVA / Llama3.2 Vision	7B, 11B, 90B	Vision + language models (image understanding). (Ollama)
MiniCPM-V	~8B	Vision-capable model. (Ollama)

Compact

Model	Sizes	Notes
smollm2	135M, 360M, 1.7B	Very small models for low-resource hardware. (Ollama)
TinyLlama	~1.1B	Ultra-light general LLM. (Ollama)

This is the explanation on the size requirements and hardware:

What the Sizes Mean (Roughly)​

• ~1B–3B — Can run on modest hardware (like 8–16 GB RAM, sometimes without a GPU).
• ~7B — Standard baseline for general LLM tasks. Good quality without huge GPU requirements.
• ~13B — Better quality & reasoning; often needs more RAM/VRAM.
• ~70B and larger — High capability but hefty hardware requirements (lots of VRAM/RAM).

On my old box with decent CPU/RAM but low GPU it recommended this set:

Default

Llama 3.2 3B​

• llama3.2:3b is explicitly positioned as strong at instruction following, summarization, prompt rewriting, and tool use

or
phi3 includes a 3.8B “Mini” variant designed to be lightweight.

More Agent workflows

Qwen2.5 3B / 7B​

• Qwen2.5 has many sizes (including smaller ones via tags).
  On your CPU, I’d start at 3B; try 7B only if you’re happy with slower responses

 

[L-Dog]

YEah heard teh Qwen model is the best or deepseek depending on what you want. Openclaw is just teh interface and can do a lot more than some of the other "brokers". Upgraded my windows 10 to 11 so will try the setup from begining again tonight otherwise I will probably just use chatgot / grok / gemini and tell it to assist me with the most basic set up

 

[L-Dog]

https://twitter.com/x/status/2026662838022832472

 

[L-Dog]

Set up my openclaw this weekend but ran into numerous errors. Chat gpt often gave me linux code and had some rate limit issues due to local model running, kept looping teh same 3 responses. Think I will just start a fresh set up using claude code (once back up)

 

[L-Dog]

Hey guys, so after tesing AI a but think I am goign to sell my old pc and just use my main gaming pc as I don't have much time for gaming anyway. Was told to use WSL2 and linux to set up my openclaw. Has anyone tried this and is it the best way on a windows pc ?

 

[Kosmik]

Hey guys, so after tesing AI a but think I am goign to sell my old pc and just use my main gaming pc as I don't have much time for gaming anyway. Was told to use WSL2 and linux to set up my openclaw. Has anyone tried this and is it the best way on a windows pc ?

I did similar to a old laptop and ran ubuntu directly. Openclaw was a pain with talking to my local ollama, but I think its a resource issue ie: Ram. Ollama and models ran fine but Openclaw needs a certain level of model to run properly, round the 4B and higher.

Was playing with the qwens.

 

[L-Dog]

https://twitter.com/x/status/2034655113961455651

Crazy how fast you can built an app with zero effort

 

[L-Dog]

https://twitter.com/x/status/2043770365369876979

Anyone tried Hermes yet, will be intresting to see how it competes with openclaw

 

[quovadis]

https://twitter.com/x/status/2044370455260176713

 

[NarrowBandFtw]

what's your toolset / environment looking like for AI dev these days?

went deep down the rabbit hole recently:
- Claude Code ... nice, but too expensive for casual exploration
- Antigravity ... decent, but why the custom IDE?
- VS Code, Roo, Hermes (the agent, not the model) to hook up any model, except:
-- all the half decent free models accessible via OpenRouter are permanently out of free tokens
-- Google AI studio is a decent alternative, but Hermes has a bug and can't connect to it
-- Ollama is a decent alternative, but skullfcks your hardware in no time, better prepare to wait for any responses
-- oh and VS Code's April update seemingly makes Roo obsolete

so currently:
- VS Code, Roo, Google AI Studio API key ... free models still run out quite fast
- Hermes, Ollama, Gemini4 ... edge models to limit VRAM requirement, but hardware still keeps asking me why I'm asking the impossible

also playing around with Claude Projects, very first try the thing has zero project context and keeps guiding me how to properly start a chat within the project ... which is what I was already doing, proven with a screenshot, it asked me to log a bug with Anthropic

EDIT: and as if to prove just how fast these things move, that Hermes bug was fixed in a weekend update, now talking to Google AI Studio models without issue ... also realised Roo is kinda not needed so I nuked it

 

[dualmeister]

YEah heard teh Qwen model is the best

I'm getting some really good results with Qwen 3.6
Very detailed guides. Impressive.

 

[L-Dog]

I'm getting some really good results with Qwen 3.6
Very detailed guides. Impressive.

Yeah currently using Qwen3.5-9B but think 3.6 might just be above my Pc's threshold. Fine tuning my trading bot over the next week and might dive into the bets available model again

 

[dualmeister]

Yeah currently using Qwen3.5-9B but think 3.6 might just be above my Pc's threshold. Fine tuning my trading bot over the next week and might dive into the bets available model again

Seems Qwen2.5-VL-7B will do all that I need for now.
Gemma 4 almost choked my PC to death

 

[konfab]

So I have been using Github copilot pro and its extremely generous subsidisation for a while now and basically view AI assistance as a critical part of development. That is coming to a close at the end of the month as they are moving to API pricing. My main usage till this point was sonnet 4.5 and then 4.6 with bits of Opus.
My main way of using these models has been with OpenCode. If you are using these agentic harnesses, you can see what your usage will roughly be with a little cli utility:

npx copilot-arewecooked

As you can see, most of that is due to Claude Sonnet and Opus. The GPT-5.5 and others we will come to later:

Which gives you how much API fees you will be paying. For me, it was about $200 USD average, which is quite a lot going up from $40 per month.

Obviously this is a problem. To help others, I will document the methods I used. The trick I found out is that you cannot really compare token per token costs between models. A model that outputs the right answer in 5x more tokens, but being 2/3 the cost per token will not be cheaper to run. To quantify this, there is a site that measures test scores vs cost to run.



Source
This was very interesting to me, as you can see that according to this metric, GPT 5.5 low reasoning is basically equivalent at producing the same sort of results as sonnet and some of opus reasoning, but for 1/4 the price.

To me, this is a slam dunk against Anthropic's models, which is why I have moved over all my work to mostly using GPT 5.5 low reasoning.

Using a slightly shorter period, you can already see the difference.


I will be moving over to a ChatGPT pro plan for now, and see how that goes. If it isn't enough, then we will need to switch to the $100 USD a month plan.

But this is worth the price as opposed to the $hyteshow with Anthropic who will rugpull whatever tools you like using.

 

[semaphore]

So I have been using Github copilot pro and its extremely generous subsidisation for a while now and basically view AI assistance as a critical part of development. That is coming to a close at the end of the month as they are moving to API pricing. My main usage till this point was sonnet 4.5 and then 4.6 with bits of Opus.
My main way of using these models has been with OpenCode. If you are using these agentic harnesses, you can see what your usage will roughly be with a little cli utility:

npx copilot-arewecooked

As you can see, most of that is due to Claude Sonnet and Opus. The GPT-5.5 and others we will come to later:

View attachment 1908309Which gives you how much API fees you will be paying. For me, it was about $200 USD average, which is quite a lot going up from $40 per month.

Obviously this is a problem. To help others, I will document the methods I used. The trick I found out is that you cannot really compare token per token costs between models. A model that outputs the right answer in 5x more tokens, but being 2/3 the cost per token will not be cheaper to run. To quantify this, there is a site that measures test scores vs cost to run.

View attachment 1908312

Source
This was very interesting to me, as you can see that according to this metric, GPT 5.5 low reasoning is basically equivalent at producing the same sort of results as sonnet and some of opus reasoning, but for 1/4 the price.

To me, this is a slam dunk against Anthropic's models, which is why I have moved over all my work to mostly using GPT 5.5 low reasoning.

Using a slightly shorter period, you can already see the difference.
View attachment 1908324

I will be moving over to a ChatGPT pro plan for now, and see how that goes. If it isn't enough, then we will need to switch to the $100 USD a month plan.

But this is worth the price as opposed to the $hyteshow with Anthropic who will rugpull whatever tools you like using.

Yeah , just go pro. The GPT plus accounts are effectively free accounts with a bit of limit. I can use up a plus account 5 hour session in about 5 minutes. Go proo 100$ or 200$