heh good luck with that.
This whole circus is out of control.
1 Like
LLMs are training data dependent (bad input leads to worse outputs), and too this day LLMs fail at basic logic questions, and it is highly suspected that many of the models have the answers to the ātestsā hard coded into them (it is amazing how ever time there is a new test the LLMs do horrendously at them, and then amazingly they do a lot better)
they donāt understand labels after 6 years, they donāt understand spelling after 6 years. Why have they not bean able to fix these things after such a long time.
LLMs are deemed most proficient by users on the same metrics as Gelman Amnesia.
the LLMs are based on pattern matching, and probability determinations; that āintelligenceā puts them on a highly literate Middle school student. when I can put in a coding question of current models, and get back outdated/deprecated information, or information that was NEVER accurate. I have seen the tokens being burned on LLM agents going down paths that had no merit to begin with, and a human with any level of proficiency would have never started on, and so many people are happy because the LLM returned in 30 minutes spending the a weeks salary in tokens, and still need to be outright corrected.
I donāt care what āthe futureā models will do because they donāt exist. The same way a publisher wonāt give me a contract unless I have something tangible to show or a proven portfolio (if you find a publisher that will give a contract on a pitch deck without a portfolio you have found fools); which is the same exact BS āAI [psychosis] brosā are doing.
Verse might stand a chance because there is less ābad codeā for the LLMs to reference, so until there is a proliferation of ābad codeā the LLMs should do okay, but after that they will be just as error and hallucination prone as they are with C++, or C#.
The only people who should be afraid of being replaced by LLM AI are the people who never understood their job, or had a job that never needed subject matter competence. heck it is the Middle managers of established companies that should be the most concerned about these LLMs because they tend to lack subject matter expertise.
I kinda want to agree with this, and itās quite optimistic, but I also see a difference between previous languages and verse.
Verse development will be entirely after the discovery of LLM so even though older languages examples in various communities can sometimes suck but it was still human made, those ecosystems was developed by humans first. Thus I am not sure with Verse and LLM training data would have that case, it can be a self eating snake for LLM. 
Generated Verse creates more public examples, then those examples influence more generated Verse.
1 Like
Switch your modelā¦ if youāre getting those results youāre missing out. I use ChatGPT Paid in visual studio and itās been pretty flawless with unreal c++ since ~5.2. Iāve been able to nearly one-shot complex systems that use iris networking, niagara systems as a service, mass entity and other highly specific unreal systems. It will hallucinate a function name every once in while, but thatās been the biggest issue for quite some time.
I am experienced with unreal c++, so Iām making very specific well designed prompts, and the LLM generally responds with very well designed systems that follow design patterns and unreal best practices. Often the LLM corrects for poor assumptions I made or suggests better designs. The generated code is clean and very readable. I can build out a framework in an afternoon that would have taken me a week or two to get the same results pre-AI. Itās made me fall in love with programming in a whole different way.
Iāve found Gemini/Google AI to be really sloppy with unreal. Like really really bad at times. Maybe people just try that because of ease of use and form their opinion there. Either way, I kinda feel bad for people missing out on this.
2 Likes
I know, I built a small model myself and trained it with Unsloth.
i.e. Programming your game. One personās creative profession is anotherās tedium that can be replaced with AI
I agree, its good in strong hands. In defined system operations or with well designed prompts with extensive knowledge which is again actually a system with pseudo code for machine.
Itās āutterly disgustingā with unsolved problem spaces, approaches, offering or constructing a problem space or solution space even in strong hands. However its still handy in high level tech design or gameplay design so that we can organise ideas, rather then making mad internal talk : we can actually talk with something back & forth.
That is actually basis of my concern about the data set and self eating snake of a new language ecosystem mentioned above (model collapse). I am not sure itās just me but I canāt even find a hit in google searches anymore accurately. Maybe I am just paranoid but I feel that is something of a fast growing dataset. It feels like watching internet becoming garbage in realtime. Even with a huge data-set can become like this, what about a new ecosystems quality?
Either search ecosystem is vulnerable to this already or there is intentional no action against it for various reasons (money). So the AI response in the every search is I feel like more likely a product mindset rather than an innovation, or maybe thay are just trying to adapt cause they already know we canāt fight it? New ecosystems like Verse can be more vulnerable to this.
1 Like
The world of coding (not engineering) has changed.
Even the maker of Linux admit it, and now uncle bob is vibe-coding tooā¦
Bizare!
1 Like
Thatās Unc B, not Joe Dirt.
This guy gits it.
2 Likes
Can I point people coming to this topic fresh to this excellent blog post by Adam Hafez which is what you need to read to know the what and why and to give you some hope. The .uasset Problem, Revisited: Epic Bet the Engine on Text
Read the article. Everything in the article is covered and discussed here alreadyā¦ Including the psuedo node editor.
One small nuance: although MCP requires the Unreal Editor process, you do not need to launch it yourself. The agent can start the editor silently in the background, either from Rider or from a console opened in the folder containing the .uproject file, and then send MCP commands to that hidden editor instance.