Fun with ChatGPT: Putting vast amounts of computing resources to work on important model rocketry tasks

The Rocketry Forum

Help Support The Rocketry Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.
For all of the acclaim that ChatGPT is getting about passing business school exams, I'm pretty sure it would flunk high school Geometry and it would definately flunk my Organic Chemsitry class.

And that says everything we need to know about business majors. Who run the banks. Errrr....
 
From the first set of single word not-real-word names, Neobolt, Fluxion, Fluxoray, and Eonexus are all pretty decent names for a hilariously begreebled sci-fi design. I may need to steal one of them.
Those are great names for models!

Begreebled is a really great word, too, thanks for that.
 
"In a simulated bar exam, GPT-4 passed with a score around the top 10 percent of test takers, compared to the bottom 10 percent where GPT-3.5 passed. On the SAT reading and writing section, GPT-4 scored a 710 out of 800, 40 points higher than GPT-3.5"

https://thehill.com/policy/technology/3901479-5-things-to-know-about-gpt-4/
As of a decade or so ago, to pass the bar exam (in most states), an individual had to pass 3 parts. One deals with ethics and is multiple choice. The second part is also multiple choice and deals with general concepts that apply to most states. The third part is mostly or exclusively essay and is basically a law school exam (ie issue spotting and word vomitting).

Curious as to which parts GPT-4 took. If GPT-4 had access to the Internet, I'm not surprised it passed, although being able to issue spot is a neat trick. The thing about doing well with law school tests and bar exams is that rote memorization with speed is one of the biggest skills to have, which is something machines have almost always performed better than humans at. So GPT-4 doing it is nothing to be proud about.
 
That just gives access to ChatGPT, need to pay for Plus account to use GPT4, at least that’s my interpretation (it’s a little confusing)
Maybe I have it screwed up too. I thought OpenAI upgraded ChatGPT to GPT-4 in the background. If that’s not the case then ChatGPT-3.5 got much better between the times that I gave it those prompts.
 
If anyone out there has GPT4 access (I do not) then by all means try some of the same queries and let's see what we get.
I think it's only available to select people (waiting list) who specify what they want to do with it, not anyone.

My input:

"Is this input processed with chatgpt or gpt-4"

Output:

"As of my knowledge cutoff date of 2021, there was no official release or announcement of the GPT-4 model by OpenAI. Therefore, I am a GPT-3.5 language model, and any input you provide is being processed by me, ChatGPT, using the GPT-3.5 architecture."
 
Putting it to good use:

Input:
"What is the best glue for building a model rocket?"

Output:
"The best type of glue to use for building a model rocket depends on the materials you are using to construct it. Here are some common types of glue used in model rocketry:
  1. White glue (PVA): This is a water-based glue that dries clear and is easy to clean up. It works well for attaching paper or cardboard fins to a rocket body.
  2. Epoxy: This two-part adhesive is strong and durable. It can be used for bonding plastic or metal parts, as well as for filleting fins.
  3. CA glue (cyanoacrylate): Also known as "super glue," this fast-drying adhesive is good for bonding small parts quickly. However, it can be brittle and is not as strong as epoxy.
  4. Wood glue: This type of glue is specifically designed for use with wood and is good for bonding balsa or plywood parts.
When selecting a glue, make sure it is suitable for the materials you are using and that it has enough strength to withstand the forces of flight. It's also a good idea to follow the manufacturer's instructions and to test the glue on a small area first to ensure it doesn't damage the materials."
 
Input:

“What nose cone shape is most aerodynamic for a rocket?”

Output:

“The most aerodynamic nose cone shape for a rocket depends on several factors, such as the rocket's velocity, altitude, and flight conditions. Generally, a nose cone with a streamlined, pointed shape is the most aerodynamic, as it helps to reduce drag and increase the rocket's speed and stability.

There are several common nose cone shapes used for rockets, each with its own advantages and disadvantages. Some of the most common nose cone shapes include:

  1. Conical: This is the simplest and most common nose cone shape, with a conical taper that comes to a point. This shape is best suited for low-speed, low-altitude flights, as it offers low drag and good stability.
  2. Ogive: This shape has a more gradual taper than the conical nose cone, with a smooth, rounded curve at the tip. Ogive nose cones are good for high-speed flights, as they offer low drag and reduced shock waves.
  3. Blunt: This nose cone shape is flat at the tip and has a larger diameter than the rocket body. Blunt nose cones are best suited for high-speed, high-altitude flights, as they create a shock wave that helps to reduce drag.
  4. Parabolic: This nose cone shape has a smooth, curved surface that creates a continuous curve from the tip to the base. Parabolic nose cones are best suited for supersonic flights, as they help to reduce drag and increase stability.
In general, the most appropriate nose cone shape will depend on the specific rocket design and its intended flight conditions. Aerospace engineers use computational fluid dynamics (CFD) simulations and wind tunnel testing to determine the best nose cone shape for a given rocket design.”

- - - - -

Well here's a thought: if you were worried about robots replacing humans for jobs, how about chatgpt almost looking like it could replace TRF. 🤔
 
Input:

“What is the easiest way to locate the center of pressure of a model rocket?”

Output:

“The center of pressure (CP) of a model rocket is the point where the aerodynamic forces acting on the rocket can be considered to act upon. The location of the CP is an important design parameter for a model rocket, as it helps to determine the rocket's stability and flight characteristics. There are several ways to locate the CP of a model rocket, but one of the easiest methods is to use a simple tool called a CP calculator or a CP calculator app.

A CP calculator works by taking into account the shape and size of the rocket's fins, body, and nose cone, as well as the position of the rocket's center of gravity (CG). To use a CP calculator, you will need to measure the dimensions of your rocket's components and input them into the calculator. You will also need to determine the position of the rocket's CG, which can be done by balancing the rocket on a narrow object such as a dowel or ruler.

Once you have inputted the necessary data into the calculator, it will generate an estimate of the CP location. The CP location is typically given as a distance from the rocket's nose cone tip, and it is often expressed as a percentage of the rocket's length.

It's important to note that while a CP calculator can provide a good estimate of the CP location, it may not be 100% accurate. Factors such as air density, wind speed, and rocket spin can affect the actual location of the CP during flight. Therefore, it's always a good idea to perform a few test flights with your rocket to verify its stability and make any necessary adjustments.”

- - - - -
 
Input:

“What nose cone shape is most aerodynamic for a rocket?”

Output:

“The most aerodynamic nose cone shape for a rocket depends on several factors, such as the rocket's velocity, altitude, and flight conditions. Generally, a nose cone with a streamlined, pointed shape is the most aerodynamic, as it helps to reduce drag and increase the rocket's speed and stability.

There are several common nose cone shapes used for rockets, each with its own advantages and disadvantages. Some of the most common nose cone shapes include:

  1. Conical: This is the simplest and most common nose cone shape, with a conical taper that comes to a point. This shape is best suited for low-speed, low-altitude flights, as it offers low drag and good stability.
  2. Ogive: This shape has a more gradual taper than the conical nose cone, with a smooth, rounded curve at the tip. Ogive nose cones are good for high-speed flights, as they offer low drag and reduced shock waves.
  3. Blunt: This nose cone shape is flat at the tip and has a larger diameter than the rocket body. Blunt nose cones are best suited for high-speed, high-altitude flights, as they create a shock wave that helps to reduce drag.
  4. Parabolic: This nose cone shape has a smooth, curved surface that creates a continuous curve from the tip to the base. Parabolic nose cones are best suited for supersonic flights, as they help to reduce drag and increase stability.
In general, the most appropriate nose cone shape will depend on the specific rocket design and its intended flight conditions. Aerospace engineers use computational fluid dynamics (CFD) simulations and wind tunnel testing to determine the best nose cone shape for a given rocket design.”

- - - - -

Well here's a thought: if you were worried about robots replacing humans for jobs, how about chatgpt almost looking like it could replace TRF. 🤔
If someone selects a nose cone based on that description, well, good luck. conical nose cones are the least drag for slow flights and parabolic the least drag for supersonic? 🤣🤣
 
When talking about AI, I get a bit confused as to what makes up the 'language model' - is it select information provided by human beings or is it basically a bunch of google searches that are databased, linked and then goes through a magic box to create the output? If so, seems like it would very quickly become a positive feedback loop, as each input results in some output and then that output would be included next time there was a similar input request. I'm 100% sure that there is way more to it than that and that people have already thought about this more than I will ever be able to understand, but it seems kind of scary - not from the robots taking over perspective, but from the mass information unification standpoint. Unified information could be wonderful, unless it is misinformation.

Sandy.
 
Input:

“What nose cone shape is most aerodynamic for a rocket?”

Output:

“The most aerodynamic nose cone shape for a rocket depends on several factors, such as the rocket's velocity, altitude, and flight conditions. Generally, a nose cone with a streamlined, pointed shape is the most aerodynamic, as it helps to reduce drag and increase the rocket's speed and stability.

There are several common nose cone shapes used for rockets, each with its own advantages and disadvantages. Some of the most common nose cone shapes include:

  1. Conical: This is the simplest and most common nose cone shape, with a conical taper that comes to a point. This shape is best suited for low-speed, low-altitude flights, as it offers low drag and good stability.
  2. Ogive: This shape has a more gradual taper than the conical nose cone, with a smooth, rounded curve at the tip. Ogive nose cones are good for high-speed flights, as they offer low drag and reduced shock waves.
  3. Blunt: This nose cone shape is flat at the tip and has a larger diameter than the rocket body. Blunt nose cones are best suited for high-speed, high-altitude flights, as they create a shock wave that helps to reduce drag.
  4. Parabolic: This nose cone shape has a smooth, curved surface that creates a continuous curve from the tip to the base. Parabolic nose cones are best suited for supersonic flights, as they help to reduce drag and increase stability.
In general, the most appropriate nose cone shape will depend on the specific rocket design and its intended flight conditions. Aerospace engineers use computational fluid dynamics (CFD) simulations and wind tunnel testing to determine the best nose cone shape for a given rocket design.”

- - - - -

Well here's a thought: if you were worried about robots replacing humans for jobs, how about chatgpt almost looking like it could replace TRF. 🤔

Well it probably parsed the responses on TRF as part of its training...
 
When talking about AI, I get a bit confused as to what makes up the 'language model' - is it select information provided by human beings or is it basically a bunch of google searches that are databased, linked and then goes through a magic box to create the output? If so, seems like it would very quickly become a positive feedback loop, as each input results in some output and then that output would be included next time there was a similar input request. I'm 100% sure that there is way more to it than that and that people have already thought about this more than I will ever be able to understand, but it seems kind of scary - not from the robots taking over perspective, but from the mass information unification standpoint. Unified information could be wonderful, unless it is misinformation.

Sandy.
According to the piece I linked above, the model has statistical associations with words. “Julius Caesar was murdered” is strongly correlated with “Brutus”, so if you ask who murdered Caesar, it will give you an answer. However it doesn’t appear to try to use search engines to find the answer and then use the language model to write some sentences about it.
 
Good names and very cool task for Chat GPT.

I asked Chat GPT for the ratio of a BT-60 to BT-80 body tube and it gave me a wildly incorrect answer (it assumed that BT-60 was 60mm and BT-80 was 80mm OD and sounded very confident in its reply). I corrected each measurement and it updated its result (wonder if it updated its data). I wonder how Chat GPT would do with stability calculations.
I asked it how high a rocket would go. It asked some pretty specific questions about the weight, propellant type, diameter.. etc

I did not have any of the exact info in front of my at the time but I will input the data requested.

I plan on comparing the AI answer to OpenRocket…
 
If someone selects a nose cone based on that description, well, good luck. conical nose cones are the least drag for slow flights and parabolic the least drag for supersonic? 🤣🤣
Definitely a good idea to double check!

Well it probably parsed the responses on TRF as part of its training...
Yeah, if it gets any of its training data from the internet, it probably knows about this place.

- - - - - - - - - -

Another one:

Input:
“Which product works best to fill the spirals of a model rocket body tube?”

Output:
“To fill the spirals of a model rocket body tube, a few options can be considered:
  1. Sanding Sealer: This is a solution that is applied to the body tube and fills in the spirals. It dries to a hard finish and can be sanded smooth.
  2. Spackling Compound: This is a putty-like substance that can be applied to the spirals and smoothed out with a putty knife or your fingers. It dries hard and can be sanded smooth.
  3. Wood Filler: This is a putty-like substance that is specifically designed for filling in gaps and holes in wood. It can also be used to fill the spirals of a model rocket body tube. It dries hard and can be sanded smooth.
  4. Epoxy Resin: This is a two-part solution that can be mixed together and applied to the spirals. It fills in the spirals and creates a hard, smooth finish when it dries.
The best option for filling the spirals of a model rocket body tube will depend on the specific materials and preferences of the user. Sanding sealer and spackling compound are popular choices and are readily available at most hobby and hardware stores.”

🤖 🤖 🤖 🤖
 
When talking about AI, I get a bit confused as to what makes up the 'language model' - is it select information provided by human beings or is it basically a bunch of google searches that are databased, linked and then goes through a magic box to create the output? If so, seems like it would very quickly become a positive feedback loop, as each input results in some output and then that output would be included next time there was a similar input request. I'm 100% sure that there is way more to it than that and that people have already thought about this more than I will ever be able to understand, but it seems kind of scary - not from the robots taking over perspective, but from the mass information unification standpoint. Unified information could be wonderful, unless it is misinformation.

Sandy.

I understand it like this:

Human beings have provided it with enormous amounts of selected data ("training data"). Our input makes it decide what information to select (statistically, as boatgeek says) and how to organize it as an output. If new training data is fed to it (by the programmers), two similar inputs, provided at different times, could generate two different outputs, the later one considering the newer data. I don't think the outputs are fed back to it.

According to the piece I linked above, the model has statistical associations with words. “Julius Caesar was murdered” is strongly correlated with “Brutus”, so if you ask who murdered Caesar, it will give you an answer. However it doesn’t appear to try to use search engines to find the answer and then use the language model to write some sentences about it.
I don't think it uses search engines. People (programmers) have already fed it with training data. The higher and better quality is the training data, the more accurate the response. I think. I'm not sure how often new training data is provided to improve the model.

M'eh ... why don't I just ask it:

Input:
"Do you require an internet search to provide an output?

Output:
As an AI language model, I do not always require an internet search to provide an output. When you ask me a question or input, I analyze the text and use my knowledge and understanding of the language to generate a response.

However, there may be times when I need to access external sources of information to provide a more accurate or complete response. In those cases, I may use the internet to search for relevant information to include in my answer.

For example, if you ask me a factual question like "What is the capital of France?", I can provide the answer based on my pre-existing knowledge. However, if you ask me a more complex question like "What are the best restaurants in Paris?", I may need to search the internet to provide a more informed answer.

In summary, while I can provide an output without an internet search, there may be cases where I need to access external information to provide the most accurate and complete response possible."
 
Last edited:
ChatGPT is getting all this attention for three reasons.

First, it’s a generative AI, which means it creates brand new content instead of just reacting to existing content.

Second, it’s based on a language model. This is nothing new, as math has been used by computers to predict human communication. I imagine Grammarly and Hemingway ring a bell for some of you and those have been around for a while.

A language model works by using math to predict how to assemble words to create a coherent sentence, phrase or paragraph. This is not that hard for a computer to do. For example, fill in the blank:

“The Apollo moon program was run by _________________.”

A “regular” language model AI would fill in the blank with almost any word(s) that sounds plausible, such as USSR, United States, Lockheed, McDonnell Douglas, Elon Musk, etc. Most of these potential blank-fillers are factually wrong. This is what makes ChatGPT special, in that it has the uncanny ability to fill in the blank with a factually correct word or set of words, without requiring an unmanageable amount of training (in terms of manpower and technology).

Third, ChatGPT accomplishes its training by using neural networks. It has the ability to be given large amount of raw information (from the Internet, presumably), and process it in a useful way, WITHOUT that information first being annotated by humans. After this automatic training is complete, do humans step in to fine tune and tweak (aka: “teach”) the AI. Some of this training is to place safeguards in ChatGPT to avoid it giving harmful or offensive information/responses.
 
I do wonder if those people who are in charge of giving the AI information do so neutrally, or introduce some bias by what they give to the AI?
 
I do wonder if those people who are in charge of giving the AI information do so neutrally, or introduce some bias by what they give to the AI?
Humans always have bias. The real question is whether that bias is intentional or not.

Then there's the question of where do you draw the line between bias and trying to correct factually incorrect information?

Many people are trying to report current political events as objectively as possible. Yet, there's a reason that if you take a U.S. History class in college or high school, your syllabus will probably stop covering topics up to the year 2000 or so.
 
I do wonder if those people who are in charge of giving the AI information do so neutrally, or introduce some bias by what they give to the AI?
let me guess, you are on the alert about "bias" because you watch fox news? yeah, that's a healthy reaction.

have you read any of the chatgpt answers? it's obsequious trying to present several "sides".

but yeah, GIGO. And you'll never know, because it can not explain how it arrived at any answer. that is, what were the inputs and weights. it is inscrutable.
 
Back
Top