Discussion ChatGPT Can help writing Code for Orbiter

[Urwumpe]

If it’s so easy for humans to tinker, why don’t you recreate one of the mfds I made?

I bet you won’t. Challenge is open to anyone who thinks agentic code editors are merely toys.

First of all: Was there any demand for it? Then why should I do it?

Second: If I should do that, how much time should I allocate? 30 minutes? 10 minutes? Just one or all of them?

Third: I have a break in my training plan on sunday, should I use it? How about something more spicy: Somebody else gives me the task for the day and keeps it secret (at least from me) until its time to start. No preparation, no mindstorming for days ahead, just diving into the tunnel and leaving it again. Its posted into this thread of better an thread of its own (How about calling it "Orbiter Add-on Marathon"). Just 2x4 hours from start to end.

Would both of us be in at least comparable time zones, I could have suggested the third part for both of us in parallel, but I feel like its a bit unfair due to the different times of the day.

 

[nbcfrosty]

First of all: Was there any demand for it? Then why should I do it?

For Science.

Second: If I should do that, how much time should I allocate? 30 minutes? 10 minutes? Just one or all of them?

1h per MFD

Third: I have a break in my training plan on sunday, should I use it? How about something more spicy: Somebody else gives me the task for the day and keeps it secret (at least from me) until its time to start. No preparation, no mindstorming for days ahead, just diving into the tunnel and leaving it again. Its posted into this thread of better an thread of its own (How about calling it "Orbiter Add-on Marathon"). Just 2x4 hours from start to end.

Would both of us be in at least comparable time zones, I could have suggested the third part for both of us in parallel, but I feel like its a bit unfair due to the different times of the day.

That'd be fun if the folk here want to come up with a task I'd be down for that.

 

[Urwumpe]

For Science.

1h per MFD

I hope I can still work that slow.

 

[nbcfrosty]

I hope I can still work that slow.

The base layout one is cool, it reads all the cfg base files and auto switches to the closest base. The artificial horizon was surprising, didn't know so much math is involved in making that!

I was thinking later of making a simple low poly mesh generator off a declarative DSL, then a simple cfg file generator based on natural language description of the vessel.

 

[Urwumpe]

The artificial horizon was surprising, didn't know so much math is involved in making that!

Yes, you should be at home with linear algebra for that. Its not trivial. But also not really difficult since Orbiter already gives you all the mathematical tools you need for that. Imagine doing that embedded on a RISC V5 to drive a real HUD. Thats when joy comes.

I was thinking later of making a simple low poly mesh generator off a declarative DSL, then a simple cfg file generator based on natural language description of the vessel.

I have similar ideas for making aircraft development easier by automatically generating a basic mesh that can be refined in Blender. But thats more like tooling, not necessary playing with the API. Also, aircraft have lots of properties already well-defined by engineering that you can use for defining it in a UI or a YAML file.

How about reading a local base vector map, similar in format to the Orbiter Map MFD and display it similar to a car navigation system projected as a birds eye view from above your spacecraft?

 

[nbcfrosty]

Yes, you should be at home with linear algebra for that. Its not trivial. But also not really difficult since Orbiter already gives you all the mathematical tools you need for that. Imagine doing that embedded on a RISC V5 to drive a real HUD. Thats when joy comes.

I never got a strong gut intuition behind linear algebra. If I can't derive it on my own, I don't understand it. It's pretty cool using Orbiter/LLM's as a teaching device.

I have similar ideas for making aircraft development easier by automatically generating a basic mesh that can be refined in Blender. But thats more like tooling, not necessary playing with the API.

Hey you may want to play around with this:

FreedomIntelligence/BlenderLLM · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Embrace the dark side!

I really want to play around with this in Orbiter: https://microsoft.github.io/TRELLIS/

Also, aircraft have lots of properties already well-defined by engineering that you can use for defining it in a UI or a YAML file.

Good idea

How about reading a local base vector map, similar in format to the Orbiter Map MFD and display it similar to a car navigation system projected as a birds eye view from above your spacecraft?

That sounds like a cool idea. That base layout MFD above could become a legitimate addon with some work I'd say.


There is a classical "sanity" check for testing an LLM's capabilities: draw me a svg of a pelican riding a bike
Older variations ask for a unicorn.

I wonder if there's a simple DSL that allows one to describe planes/cones in space so that an LLM can be used to emit that DSL and generate low poly meshes like that. I will experiment after work. Would be cool to simply say: Give me mesh of F 18 hornet.

And for it to emit a mesh of that. Like what ThunderChicken was doing a while back. The actual stats I think are googleable and model should be able to easily find the info on the internet.

Lua Script VW Thing

Oct 21, 2024

A Volkswagen Thing, a Lua script testbed for steering, braking, and animations for surface vehicles.

• Thunder Chicken
• Miscellaneous

 

[JDat]

Yes, you should be at home with linear algebra for that. Its not trivial. But also not really difficult since Orbiter already gives you all the mathematical tools you need for that. Imagine doing that embedded on a RISC V5 to drive a real HUD. Thats when joy comes.

That's sound interesting. Almost the same what I am doing now. But I am doing it with RPi Pi and SDL because need composite video output. SDL can be changed to whatever you want. Just some line and text drawing on screen. While somebody use AI for that and pretend it his code, I am using 3D cockpit's HUD code directly from Orbiter. And... Ethernet joystick for Orbiter...

 

[steph]

It doesn’t know what a delta glider is. Only concepts that were repeatedly seen in the training corpus are then encoded and learnt by the model. Not enough training material on delta gliders.

But that’s fine…. Although the LLM latent knowledge is vast, they experience an emergent phenomenon called in context learning. So you can feed it information about a delta glider and it’ll know more. Tell it to find images for delta glider on the web, you can feed it images of your own.

When working with Orbiter API it doesn’t know the API at all. You have to ground it by having it read the orbiter api header files for example.

You can view llm models as “cognitive cores”.

Oh, but it should. Especially if using paid subscription. I get that there is stuff that it wasn't trained on, but it does have web search function and should be able to look it up if it doesn't know, instead of spewing that kind of stuff. Chat GPT seems to simply hang on to its mistakes, unless really told to do otherwise. It's not the user's job to train it on Orbiter. But I get it , considering the stuff it sometimes does, like telling me that I should just google the information, or that there's no need for GPT to do a certain task since I can do it myself anyways.

 

[JDat]

It's totally OK if someone mess with MFD addons and AI. Everyone can do whatever they want. But... If someone claim he is "professional" and at the same time claim that AI code is copyrighted by "professional" and that "professional" don't think that he need to tag/mark that when AI code potentially can get into Orbiter core. No comments there. This show how competent is this "professional" with probably fake degree.

So far most important addon is NASSP and this should be not broken. No matter what "professional" is doing with AI.

I am not against AI, I am against of profanity and stupidity when people like @nbcfrosty want to pretend they know anything and at the same time kinda "ready" to do real things.

@nbcfrosty you are always welcome to go Kerbal Space route: Write your own simulator with or without AI and mess with it. And I am not even talking about usefulness about your crappy addons. "Professionals" like you must be kept away from critical code.

 

[Urwumpe]

Just my two cents there

I am not against AI, I am against of profanity and stupidity when people like @nbcfrosty want to pretend they know anything and at the same time kinda "ready" to do real things.

I think both of you did show the best side of yourself in this discussion. And it wasn't just about AI. I would really like it if you both could disarm the discussion a bit and be, well, professionals. Or at least good boys.

@nbcfrosty you are always welcome to go Kerbal Space route: Write your own simulator with or without AI and mess with it. And I am not even talking about usefulness about your crappy addons. "Professionals" like you must be kept away from critical code.

Please remember that you are not the gatekeeper of Orbiter or evil dictator maintainer of the source code. Maybe others also want to have a word there. Yes, I know that you contributed to Orbiter and I thank you for that. Also I don't want to hear the word crappy in that context, when all involved know that it was just an experiment and not meant to revolutionize Orbiter development. Also it tells very little of his professional behavior and I don't want to rate your professional behavior just by what you show here in this thread. Again, please make sure the community stays nice, friendly and open for new and old people. You don't like AI? Same here. I admit it, I am highly sceptical. But that is no reason to get too close to bullying. Also I think having a more favorable point of view regarding AI is really helping this discussion and thus, also improving the community as whole.

I welcome contributions to Orbiters community by @nbcfrosty and you. I want this community to grow, not to be elitist and toxic for a short time before it dies. Even in a world with AI. Our standards should not be for keeping people out, but for making their ideas better.

 

[JDat]

I think both of you did show the best side of yourself in this discussion. And it wasn't just about AI. I would really like it if you both could disarm the discussion a bit and be, well, professionals. Or at least good boys.

Ok, ok. Yes, we both showed best our dark side (of the Moon or whatever)... It started as AI code in Orbiter, but ended... Well bad end. I am just spending my free time, so I am amateur not professional. I can't be professional because I am not making money with programming. Professionals make money with programming, amateurs spend money and time into programming. It''s not related to University grade, skills or years of experience.

Contribution to Orbiter. Well I found few kinda disputable things in Orbiter, but will ask about it in separate thread. Not here.

To be simple, My position: AI code in own MFDs or other modules/addons in Orbiter are welcome, but, please, no stupid code or AI into orbiter core for now! Do not brake anything in working project!

 

[Urwumpe]

Ok, ok. Yes, we both showed best our dark side (of the Moon or whatever)... It started as AI code in Orbiter, but ended... Well bad end. I am just spending my free time, so I am amateur not professional. I can't be professional because I am not making money with programming. Professionals make money with programming, amateurs spend money and time into programming. It''s not related to University grade, skills or years of experience.

LOL, actually I have less strict definition of professional behavior. I don't require to make money with it, but to act professional. Of course not everywhere. But when development is no joke to you, you sooner or later start to take things seriously and take responsibility of development as whole and accountability for your own works.

 

[JDat]

Well, in my experience, mostly electronics embedded, show business and cinema, some amateurs make better things than professionals. But there is also "time" factor schedule and reliability. Because I am professional in show business industry and electronics, then programming (big projects), UI experience, reliability and Orbiter is my hobby. So... I am that guy, who want that things just work instead of dancing near them with kick drum and hope. So far my experience with AI generated code is nothing more than waste of time. Arduino blinkelight AI generated code doesn't count as something serious.

 

[malcontent]

If you are a programmer, you should understand the limitations a LLM has, and work along side that. It is a word predictor, not a word understander. Trained on lots of C code it is fairly accurate for setting up frameworks, but not suitable for much more than that, worse if you use some more obscure languages. This goes into the whole "vibe-coding" thing, which is basically: asking an LLM to write code to do a task, debugging by reporting the errors to the LLM, and iterating, until you have a black box that sort of does what you want, but the designer doesn't actually understands how it works. Maybe don't give one of these things the ability to make SQL queries: https://www.pcmag.com/news/vibe-coding-fiasco-replite-ai-agent-goes-rogue-deletes-company-database

 

[nbcfrosty]

btw just want to say that there is a huge misconception about LLMs. Misconception is that LLMs are merely next token predictors. True, LLMs experience vast amounts of pretraining on next token prediction objective. This pressure on the weights due to gradient descent causes structure to form within the layers. LLMs have encoded within their weights vector programs, structured modes of computation. They have latent representations. They have world models.

This is proven many times and trivially so by LLMs’ emergent ability to perform tasks never seen in training data by exploiting combinatorics. Researchers are able to chain a set of skills in a way where we can guarantee that the exact chain was never seen in the training data. And yet LLMs can generalize to never before seen chains.

Current models are extremely inefficient, they have encoded within them random hashes and garbage from all over the internet. But one day we can expect the core cognitive engine to be way smaller than today’s llms.

LLMs are actual computers in a way, there is actual computation occurring when we sample an LLM. You can prove this to be the case by building a simple multilayered perceptron that can do things like xor.

LLMs have been shown to be able to do things perfect 20 number multiplication, because during training the model learned a vector program to do it. But at the same time fail simple multiplication because of not having a program that can do it or those weights simply not being activated during inference. It’s why self consistency is a thing where you sample multiple times and take the most generated result because accuracy scales with consistency linearly.

There is an entire field called mechanistic interpretability that seeks to probe the inner workings of LLMs. Researchers have tweaked weights in a way making an LLM think that it’s the Golden Gate Bridge.

It’s fascinating, and I don’t think most people see llms as merely next token predictors anymore.

 

[nbcfrosty]

Oh, but it should. Especially if using paid subscription. I get that there is stuff that it wasn't trained on, but it does have web search function and should be able to look it up if it doesn't know, instead of spewing that kind of stuff. Chat GPT seems to simply hang on to its mistakes, unless really told to do otherwise. It's not the user's job to train it on Orbiter. But I get it , considering the stuff it sometimes does, like telling me that I should just google the information, or that there's no need for GPT to do a certain task since I can do it myself anyways.

My favorite:


It's all about grounding. It can definitely find images on the internet but your ChatGPT session was triggering dall-e, which is an older shittier model. Above images were generated via gpt-image-1 and it is next level.

 

[Urwumpe]

There is an entire field called mechanistic interpretability that seeks to probe the inner workings of LLMs. Researchers have tweaked weights in a way making an LLM think that it’s the Golden Gate Bridge.

Maybe you agree that this triggers many bullshit alerts, since it is obviously a very badly verifiable claim...

 

[nbcfrosty]

Maybe you agree that this triggers many bullshit alerts, since it is obviously a very badly verifiable claim...

https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html

Golden Gate Claude

When we turn up the strength of the “Golden Gate Bridge” feature, Claude’s responses begin to focus on the Golden Gate Bridge. For a short time, we’re making this model available for everyone to interact with.

www.anthropic.com

It's a legit field.



Here you can see a layer of a CNN visualized. You can see things like edges being encoded. As the layers progress you see higher and higher level features. The broad term for this type of learning is called feature learning/representation learning: https://en.wikipedia.org/wiki/Feature_learning

The Linear Representation Hypothesis and the Geometry of Large Language Models

Informally, the 'linear representation hypothesis' is the idea that high-level concepts are represented linearly as directions in some representation space. In this paper, we address two closely related questions: What does "linear representation" actually mean? And, how do we make sense of...

arxiv.org

 

[Gondos]

This is really a cleaving subject. Many professionnals in a field will always be skeptical about a tool that purportedly can do their jobs. What doesn't help is the attitude of some people telling you they can do it better, faster, whatever-er than you can. Would you walk to Chopin and say "Dude, you wrote 21 nocturnes in a lifetime, I had xxxGTP crap me a hundred of them out during lunch break, why are you're so inefficient"? Would you board a plane if the company told you "Don't worry the automatic takeoff and landing system was coded and unit tested with Claude"? (I'll ask around on the next PSAC review)
I'm no Chopin and software of not really an art but I love my craft, and any tool that may help, I'll try. So far the technology is kind of in the "uncanny valley". From far away it can look convincing, but when you look closer, it can be full of errors. Spotting these errors requires understanding what you do.
The bottomline is I personnaly can't trust the output so I don't waste time trying to sort it out.
On the other hand, if I required something out of my field, let's say a piece of background music for a scene, I may be tempted to try and generate one with that technology. I may go to a musician forum and ask about music theory but I'd never declare "I can write music with xxxGPT" there.
So you want to sketch something or make a toy project with an LLM, fine. But if you want to learn or do something bigger, I don't think that's the right way.
Maybe someday the technology will get there and make everyone of us obsolete, but until then, respect people, respect their craft
BTW, does it remind anybody of this novella by Asimov?

PS : if an LLM expert can make its toy spit out the correct formula for converting an arbitrary proper acceleration to coordinate acceleration in a Schwarzschild metric, I might have a change of heart

 

[Urwumpe]

https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html

Golden Gate Claude

When we turn up the strength of the “Golden Gate Bridge” feature, Claude’s responses begin to focus on the Golden Gate Bridge. For a short time, we’re making this model available for everyone to interact with.

www.anthropic.com

It's a legit field.

But it means something else than you said in your informal post, thus the confusion or rejection: It they managed to extract a subset of the LLM, that specializes on the Golden Gate Bridge. Useful for research (as described), but not thinking it is a famous bridge. It just does the perspective of a Golden Gate Bridge expert.(But no expert system.)