BBS: TELESC.NET.BR Assunto: Using AI for coding De: DaiTengu Data: Thu, 26 Feb 2026 12:44:00 -0600 ----------------------------------------------------------- hbRenb: hcUsing AI for coding bBynb: hcMRO bto cDaiTengu bon cThu Feb 26 2026 04:45 amn >> That's not how LLM models work. they don't "learn" in real-time. They have >> a limited memory based on your current conversation with an agent (some >> agents can share the data with other agents, but you're not going to find >> that through a web interface chat) MR> how do you know that though? maybe it talked to someone about synchronet. MR> maybe they made scripts with that person and the person uploaded code. the MR> AI banked all that info for use later. It's part of my job to know these things. unfortunately. MR> we don't really know what these things are doing behind the scenes, do we? this statement is more accurate than people realize. You're right to a degree, but probably not in the way you think. We do know that AI Models don't "learn" and "share" input among sessions until the model is rebuilt/updated. building/training a model is VERY expensive and incredibly time consuming, require multiple datacenters worth of compute power for things like Claude, OpenAI and Gemini. What we don't really know is how these models work internally. Even the people who create and train them don't really understand how they work. That's why they can be jailbroken, and why models like Grok can suddenly start calling itself "MechaHitler", and go on to generate nude pictures of real people. MR> regardless, they can make code. i would call claude.ai the most superior. MR> it will be done with the ai editor in a bit. it just needs some visual MR> cleanup. Claude Opus 4.5 has been my go-to for code generation. I'm told 4.6 is even better, but it's expensive as hell to use. That said, it is nice that my job pays for me to use it. I'm on the low-end of our teams budget, so if I use it for personal projects, no one is really going to notice. A thing I did that's 95% generated by Claude Opus 4.5: https://github.com/DaiTengu/slaane.sh It's basically what i use in my bash environment. It did a decent job of generating "thematic" documentation in the style I wanted (Warhammer 40K), and only took me swearing at it and threatening it a handful of times to get things right. The whole project is really only useful for me, but I had it create a "modular" system that allows for extensive modification, people can easily create new modules that contain tools they want in their bash environment. I even have a "TOOD.md" file with stuff I want to add in the future. I spend a day or two a month working on it. I'm not super-fond of the bash code it's written, as it's a bit of a mess, but it works. I initially created it with the idea that it would run on Debain, Redhat, SuSe, Arch, Gentoo and other distros out of the box, but it really struggled with the testing of that, so right now it's only verified to work on RedHat-based systems. And to head off the question of "Why would you create something like this? how many servers do you need to log into?" I only need to log into a server if there's a problem, but when I manage thousands of servers, it's nice to run a curl command that will set up my bash environment just the way I want it quickly. ...A pnp transistor will be an npn. n --- gSynchronetn War Ensemble BBS - The sport is war, total war - warensemble.com ----------------------------------------------------------- [Voltar]