Ngl from my experience its faster, more efficient, easier to maintain, etcetera . . . for coding one selves own code. Its just better on every ground, I dont even need a logical reason, (despite there being several)
(this is not meant to offend anyone by any means)
Youâre not telling me anything new.
I did make a game engine on OpenGL 1.2 back in 2002 and worked as a C++ programmer for several years. Despite everything Iâm not one of those without programming experience wanting to get the agent to do everything because Iâd be lost when having to so some actual coding myself.
Iâve looked at some of the code AI produced from time to time. While there is mainly spagetti code in the GDScript area, the C++ code looks quite well organized to me and also commented more then Iâd usually do.
But what baffles me a little is the bit where you say anyone can clone your work and fix it if you open source itâŚ
âŚwhy would anyone want to look at yet another hobbyist 's flight simulator? On top of which knowing how the code was produced?
Or will you be paying Microsoft for copilot code reviews?
To me one of the big issues with LLMs, beyond the ethics, morals, legal problems, environmental problems, impact on general cognition, how they reinforce prejudice and bigotry, etc., etc., is how sycophantic they are
Itâs one reason I wonât use them, I canât stand the way they âreplyâ is so, tuned to make you like it, and they donât give proper criticism, that sets you up for more painful rejection when you show a genuine human your work that an LLM has praised constantly and said âthatâs a great ideaâ to most coming out of your mouth
A proper coworker or employer doing work together with you will give feedback, will point out areas where things could improve, suggest better solutions, or outright reject ideas that are bad, will an LLM? Not only am I concerned that it lacks the ability to judge the viability of complex solutions, especially to novel problems, but will it actually challenge you? I doubt it
We humans need to be told weâre wrong occasionally, we need to get constructive feedback, to adjust, to learn, if you ask an LLM for feedback instead of a coworker, collaborator, or even on forums online, will it give you feedback that actually challenges you, that might make you feel bad even if phrased well because you thought you had a good idea?
Itâs no different from paying people that are only going to agree with you, or who donât dare to do so because you might fire them
A year ago I would have agreed with you on that. But the AI companies have had a learning curve that not all users want smiling and bowing LLMs. I donât remeber where, but I saw a documentary that claimed that the overdriven positivity of LLMs come from post training enforment made by human trainers. They disencurage the LLMs to give answers like âI donât knowâ. Thatâs one of the reasons LLMs tend to rather halucinate then saying ⌠nope canât find any infomation on that in my DB.
Iâd like to give an example from a question I asked Opus 4.7 just recently. This is unmodified including my typos:
Hasjime:
I have a flight simulator project which uses godot 4.6 and jsbsim as a basis.
What if I drop the dependency on Godot and try to make my own game engine by means of agentic coding instead. Would that be a good ideas?
Opus 4.7:
Honestly? For ProvPilot specifically, dropping Godot to build your own engine is probably a bad trade â but itâs a more interesting question than a flat âno,â so let me lay out the real tradeoffs.
What youâd actually be signing up to rebuild
Godot isnât giving you a game loop and a renderer. For a flight sim itâs giving you: a Vulkan/Metal rendering pipeline with frustum culling and LOD, a scene graph, an asset import pipeline (glTF, textures, your Blender workflow), input abstraction, an editor for placing things and tuning values, physics integration points, audio, and the packaging/export story for macOS that youâve already fought through. Terrain3D alone is a substantial chunk of engineering youâre currently getting for free. JSBSim is already external â that part of your stack doesnât even depend on Godot, so âmy own engineâ doesnât simplify the flight model at all.
Agentic coding is genuinely good at producing a lot of plausible code fast. It is much weaker at the things that actually make an engine: sustained architectural coherence across a large codebase, GPU driver edge cases, numerical precision over large world coordinates (a real flight-sim problem â floating origin / world shifting), and the long tail of âit renders but itâs wrongâ graphics bugs that take a human days to bisect. Youâd be the one debugging Metal validation-layer errors at 1am, not the agent.
When it would actually make sense
Thereâs a version of this thatâs defensible. If your real goal is learning how engines work rather than shipping ProvPilot, a from-scratch engine is one of the great learning projects in software, and you have exactly the right background for it â youâve already poked at Godotâs rendering architecture, the Metal backend, and shader optimization, so youâre not starting cold. A focused renderer (something like a flight-sim-shaped scene: terrain, skybox, a cockpit, instruments) is a bounded target, much more so than a general-purpose engine. Agentic coding shines when you can hand it small, well-specified, testable units â a quaternion camera, a glTF loader, a tile streamer â rather than âbuild me an engine.â
The trap is the middle path: trying to ship ProvPilot and write the engine. Thatâs two hard projects where you currently have one, and the engine work doesnât make the sim better â it just moves the goalposts back a year.
A cheaper way to scratch the itch
If whatâs bugging you is feeling boxed in by Godot rather than a concrete blocker, consider a custom rendering layer or GDExtension inside Godot (C++/Rust), keeping the editor and asset pipeline while owning the parts you care about. You also keep JSBSim integration untouched. Or build a standalone renderer as a separate learning project with no obligation to ProvPilot â thatâs where agentic coding is most fun and least risky, because nothing ships-critical depends on it.
Whatâs actually driving the question â are you hitting a specific wall in Godot (perf, rendering control, the Terrain3D/JSBSim seam), or is it more that the build-my-own-engine itch is back?
In this case it missed that I had removed the dependency on Terrain3D some time ago. There is some flattering involved : â and you have exactly the right background for itâ. Otherwise Itâs about the answer I would expect to get from a human as well.
Iâd say that you asking âwould that be a good idea?â changes the probability of it replying like this in the first place, and a lot of people using these tools wonât do that, and I doubt itâd give you feedback like that if you didnât ask for it, or just asked how rather than if
How would you ask then ?
Ask what? I wouldnât first of all, but Iâm talking about people presenting an idea and asking an LLM how to do it, rather than asking if itâs a good idea, like âhow would I do X, Y, Z using A, B, Câ
Thatâs not how I understood your above statement:
I have no clue what you are talking about, what are you replying to/saying?
Results are all that matters. Letâs see some.
Do you not cringe at the thought of screaming GPUâs overheating the rack and boiling the cooling water and stoking up the coal plants somewhere in a hot Texas desert that couldâve used the water to feed the locals?
All in order to tell you stuff the people on this forum, or maybe reddit or whatever, wouldâve told you for free? And more reliably? People here require a lot less nourishment. Forums like this feed the machine you proclaim is worth listening to. Would you rather listen to it?
What response are you looking for? Just genuine connection with actual humans? Why then commit yourself to listing off arguments we all heard before? Because your LLM was trained on those arguments⌠At worst by its own outputs.
On another note..
Do you not find it odd that CEO-like humanoids would rather burn money on machines than pay us humans a decent wage?
Iâve pushed scenery3d to github.com
For ProVPilot there are still too many hardcoded paths that I want to fix before making it available.
Also the scenery data is much too large to push to Github. I have a webspace, Iâll see if I can make it available there.
Cheers ! 
Yet another terrain demo? Everyone has a couple of those in their drawers. Iâm not gonna compile this just to wasd through some heightmaps and trees.
By âresultsâ I meant a released game that can be played.
I agree fully. Texans need the water. Dont mess with Texas.
(apologies, had to lol)
Itâs almost uncanny how I can predict the future. I knew youâd say that.
I want to laugh but the US is rapidly descending into poverty right now⌠âŚ
Thankfully most promised data centers are not going to be built any time soon with GPUâs gathering dust all over the place.
Perhaps because itâs true 
Those terrain thingies are all over the place. Itâs pretty much boilerplate at this point.
A competent developer could knock something like this up in a week. âAiâ is supposed to 10x it so it should be a single day job for it. If this is all you have to show for a yearâs worth of work, hardly anyone will be impressed.
Itâs obvious from the bot exchange you posted above that youâre already beginning to get stuck with the integration of whatever youâre trying to integrate via gdextension. Youâre trying to get some âassuranceâ from the bot that they perhaps can do a better job by writing the whole âengineâ form scratch (because word is on the street that bots are not that good with Godot, so duh, letâs do it all in C++). Thatâs already quite desperate. And I can tell you right now that it ainât gonna happen. Itâll just prolong the context rot agony and suck more token money out of your pockets.
You donât need to believe some random anti-ai luddite on the internet though. Go and experience the whole thing for yourself.
You completely missed the point of that post. But what does it matter. This forum is nothing but a soap opera. Nothing to gain here.
No, I got the point you were going for. I just additionally read between the lines.