Is there any blueprint assistant available thats worth it?

thanks pixiebell

another point i just remember that people don’t talk about, is that

  1. llms knowledge is outdated the second it starts training. as it has no capacity to learn.

  2. llms have no capacity for logical inference. so they can’t extract wisdom from their knowledge.

  3. llms knowlegde is pretty much as good as the documentation and code. so if you read it yourself you’ll be wiser than an llm.

  4. use it or lose it. if you depend on an llm you dont develop your muscles. If you practice reading the engine source one day youll become comfortable and it will be a huge power. If you use llms youll only confirm to you that you CANT read the code without help, and this will invariably stunt you. Ive seen this far too many times in my career, personal life, and others. Human tend to depend on what makes the comfortale. Specially if it creates a habit. Dont forget that they are “cheap” now but their price will increase soon and they are already doing so.

  5. vendor lockin. The more you use it the more dependent you become.

  6. llms tend to hyperfixate on the text you give them. so when watching logs , they really fixate on what the error message says and can’t really find the actual source. i see that issue a ton here on the forums. If you work with cpp you will know that just forgetting a “;” can trigger the weirdest of errors. Llms take everything at face value, everything. They dont have the capacity to reason or question, and they wont. Is fundamental to how they work.

  7. And more importantly, your question is about BPs. Llms cant understand bps visually. They cant read the nodes and understand the connections and flow. Even though bps can be serialized to text (eg when copying to clipboard), they dont work like that. In fact llms dont understand.

  8. Most of the people i know they say that llms slow them. And only barely usable for writing boilerplate code, and even there they are technically, ethically, resource-wise and carreer-wise problematic. And you can already boilerplate code with templates.

  9. Also the cost of the llm is absorbed by the person, while you might think youre more productive, your company might benefit from that production but still pay you the same. So in the end you lose money by paying for these things. And while one might think “oh, i’ll tell my company to pay for it”, in my experience they always have a problem with licenses. Most of the companies i’ve worked for (90%) did NOT wanted to pay for my rider’s license. And that thing has so many tools and shortcuts that it truly speeds my productivity.

  10. Most of the times i find the answers i need with duckduckgo, and ive timed it, its even faster than waiting the llm to finish writing. And i dont have to write an overly complicated prompt. Just the keywords “unreal vsm performance” done.

  11. Llms are by definition dogmatic. and as anything that is static, prone to go stale. In the forums, you can ask a question and get 10 different options. and experienced people will remind you that the right solution depends on your priorities and specific problem. And performance must be correctly measured, planned, acted, and meausered again. Llms are quick to tell you the “Best” answer, which bites you later, but more importantly makes you a bad dev. Which is a huge huge problem, since juniors tend to want to find the “best, one and only, do once never evaluate again” solution, and those are the ones that are more seduced to use llms. you can take a look at this question. it’s extremely basic, but you get a few different options, and considerations to the project. How to limit how far a ++ or -- function can go?

  12. Cont: look at this other question Using Lumen/Nanite/VSM/PLA with low polygon meshes... where the person is trying to find a definitive answer to a grey problem. An llm will try to give you the definitive answer without much consideration for how grey it is. In the forum a person can get other point of views, and help exploring the situation to arrive at the solution that fits them more appropriately. An llm never shipped a game, it has no real experience on how to deal with these things. And you could coerce it to do a similar job, it’s not there by default.

  13. llms are very sensitive to certain invisible triggers. once an llm replied that the problem is X, it’s really hard to make it change its position. sometimes they can tell you the best answer is X, you use a different word and now it’s Y. most people only ask once.

  14. They fail to have a holistic view of the issues. and question what and why you’re approaching a problem in a certain way. in the last link you’ll see how the issue of “remember the cost of your workflow and maintenance” came up.

  15. an llm never says “i don’t know” they always reply something over confidently.

  16. an llm can’t even count the number of characters in a sentence… let alone understand a piece of code. they can’t really follow the flow of the code and detect runtime issues. (there are compile issues, logic issues, runtime issues).

  17. if you need to use an llm, that runs on a gpu farm, cooled with drinking water from 3rd world countries without water, raising the prices of electricity, to solve an issue that can be solved by your IDE, with a script, with a google search. you’re just being wasteful.

i got this from google trans the other day, i think they changed to an llm recently as i’ve been getting these types of errors consistently.

2 Likes