Using AI to speed up development

[Jon]

CC does it extremely good. As an example, I refactored tonight like 10000 lines of code in 320 subs ... just made 2 or 3 bugs (2 of them, he forgot a CLNG, and the other he forgot to make a sub altought he documented it) because when context window gets full (about 200k tokens), it begins to forget some stupid things (like all LLMs). Just mantain it under 200k, and Claude Code + opus 4.6 with mcp is really a game changer (nothing similar I've ever tested... and believe me that I have slept little this year testing things hahhahah)

I'm finding Opus 4.6 excellent. However, in a years time, local LLM's will be about the strength of Opus 4.6 today. They are improving fast. Consequently, I'm getting a mac mini pro with 64gb of RAM and have AI agents create the apps. That is the idea. Will see how well it works in practice. There is some great stuff out there for agents now. For example, gstack that can be found on github. It is basically a full team of agents, each with its own speciality. Amazing!

 

[unmateria]

I'm finding Opus 4.6 excellent. However, in a years time, local LLM's will be about the strength of Opus 4.6 today. They are improving fast. Consequently, I'm getting a mac mini pro with 64gb of RAM and have AI agents create the apps. That is the idea. Will see how well it works in practice. There is some great stuff out there for agents now. For example, gstack that can be found on github. It is basically a full team of agents, each with its own speciality. Amazing!

Mmmm... is going to be difficult... the layered system that we use today is extremely big even for big computers. We really cant imagine "the machine" that moves Opus and all the system about it. I dont think that that movement will come soon. We have nice models on local to do nice specific tasks... we can finetune others and we can even train from scratch... but what makes a model really capable is the amount of layers, abilities with tools etc. That is not going to be easy to fit on 512gb of ram and some tons of VRAM. And surely not this year with the ram prices!! hahhahah

About gstack... give a try to claudish with LM Studio or ollama is lovely, but is not the same.

And about Opus 4.6... is a beast. I have tested claudish with every openrouter model, in little and big projects (yeah, i have wasted money like if i was rich just for testing hahahha). No one gets near (and cheaper) than Opus.

 

[RaptorRaptur]

I have been using Copilot for coding support. The latest was a SQL Server query to close stale connections. While it did produce functional T-SQL, I find it to be a sloppy programmer. The query required editing for simple things like defining variables at the top of the code. Does Claude do a better job with the basics?

Claude certainly does the basics really well. Compile errors almost don't exist. I'm not skillful enough to judge Claude as to the efficiency (cleverness) of its code but can attest to its effectiveness. I haven't worked enough with Gemini to categorically say much other than the fact that I've been impressed thus far. I can't say the same for CoPilot and certainly not for ChatGPT

 

[RaptorRaptur]

Hm. It's almost like you have assumed the role of "Project Manager" and assigned your AI the role of "Able Assistant".

I think that that is not a bad way to not only think about it, but do actually go forward. My only problem is when one uses that approach with human programmers I can manage 2 or 3 projects at a time. With AI I can only manage one at a time (albeit a much shorter time.
I think, within the next 12 months, we will see staggering improvements to our "AI able assistants"