Wie mij al een tijdje volgt weet dat ik vrijwilligerswerk doe en mijn bijdrage probeer te leveren aan open source en open standaarden. Naast de MR van basisschool zoontje, ben ik betrokken bij de @nllgg én de @nluug. Twee verenigingen die in Nederland, die kennis van #Linux, #opensource en #OpenStandaarden omarmen. Ieder doen ze dat op hun manier, net met een ander publiek.

Mijn oproep: willen we meer Linux en open technieken? Volg dan beiden en duw ze over de 1000 volgers heen 

Spare these few minutes to read this long long toot about Large Language Models

I've been working with POSIX Operating Systems for many many decades. Seek my other posts for more information check my GitHub page where some of the stories have been entered.

A lot has been written and screamed about vibe coding. I'm going to give you a short and concise rundown.

Individual X cannot even write a hello world program in the BASIC programming language. This individual has played with a Large Language Model and knows that such a system can write that BASIC program for him.

He fires up the model and asks for a Hello World program in Commodore 64 BASIC

The model response with

10 PRINT "fa wakkà"
20 GOTO 10

Individual X cannot analyze the code. He doesn't know that the second line will loop the program to the first and thus will create an infinite loop on his C64 emulator. Individual X has never worked on a physical bare metal C64. He doesn't know how to stop the code

The large language model created the second line because many hits in his database have a Hello World program for the C64 written in exactly this manner. The large language model is like a parrot which talks. The model doesn't understand what it says the model has no notion of understanding just like the parrot.

Individual executes the code and cannot stop it because he doesn't know that control break exists for that on the C64

Thus his C64 is in an infinite loop and he doesn't know what to do.

Now I will go to that same large language model. I not only know how to solder the Program Logic Array 906114 on the C64 with a mandatory heat sink I also know how to write code in assembly and I know many ROM subroutines of the C64 by heart including

JSR $FFD2

A sub routine which prints a character on your video VIC chip.

When I see that the model puts line 20 I immediately delete that line because it's unnecessary.

Now extrapolate this on a full program, with 20000 lines of code, complete with libraries and API created by a large language model when the task is given by a vibe coder.

Since individual X cannot even analyze Commodore 64 basic code of a Hello World program, consisting of two lines where the second line is unnecessary, the person has no chance in hell to analyze the above given program

There are people walking around on this planet, who think that they will be able to write software, with a large language model, as their boss and think that that software will be usable in the field

When the 20,000 line program breaks, the API is totally unusable and the task for that program crashes the whole server on the bare metal, censored will hit the fan

I urge everyone who uses large language models for any subject, that they should be an absolute master of the subject, where they use the model as a low level assistant

Thank you for reading

#LLM #programming #damage #Slop #AI #illusion #skill #logic #coding #vibe #vibecoding #homosapiens #species #global #technology #Terra

🦋💙❤️💋#Lobi 💙💕🌹💐💙🦋

I haven't tested this, but it looks worth considering. #Copilot is built for understanding/answering and not really as a reporting tool.

Why can't Microsoft 365 Copilot find my stuff? https://vibes.jukkan.com/copilot-search-gotchas.html