# Asked to draw electronic schematics without sucess

I am happy to see how chatgpt draws electrical networks when I ask “Please draw a voltage divider”:

`````` V_in
|
R1
|
+-------- V_out
|
R2
|
GND
``````

then “can you make it lossless?”

`````` V_in
|
\ /   R1
/ \   |
|    +-------- V_out
/ \   |
\ /   R2
|
GND
``````

with the text explanation:

“”
In this schematic, the resistors R1 and R2 have been replaced with ideal transformers with a turns ratio equal to the square root of the resistance ratio, so that:

N1 / N2 = sqrt(R1 / R2)

The input voltage is applied to the primary winding of transformer T1, and the output voltage is taken from the secondary winding. Since the transformers have zero resistance, there is no power loss in the circuit and the voltage division is lossless.
“”

this was a very clever answer, however, the electrical diagrams starts to get difficult to understand. I wonder how it draws an opamp:

I asked: “draw a basic opamp circuit”

``````   V_in
|
+------ R1 -------+
|                  |
|+|                  |
| |                 +-+
+------- R2 -------| |
| |                 |-|
|-|                 | |
|                  |
GND                GND
|                  |
V_out            V_out
``````

according to the explanation this must be a non-inverting amplifier … however, I can not understand the diagram. Then two questions appear:

1. Is that an standard way of drawing schematics in ASCII code that I am not aware?
2. Maybe a more complicated one: How can it know the component association (topology) of the electrical network? I just a lot of electronic recipes or chatgpt really knows what is answering?

M.

GPT, although at times can draw some beautiful ASCII art, it won’t be related to your query, or it will eventually deviate into a pit of madness. There are so many steps and logical gates required to draw electronic schematics, and GPT is simply not capable of it. These are two huge road blocks that cannot be passed.

Keep in mind that it’s a dialog agent. You are trying to jam a square piece into a triangle hole. Sure with enough time and training you may have some sort of coherent schematic AI, but why? Why not just use something that’s purposed for it?

4 Likes

@RonaldGRuckus is correct.

You could try to ask ChatGPT to write SPICE netlists for a circuit, and then render the code later on another rendering software.

2 Likes

Yes, thanks for the help. I asked for an LTspice netlist and for my surprise it could manage to output a netlist. I totally agree that chatGPT was not trained for synthesizing electric networks, but it tries to do it, and for some simple examples it does really well. So maybe there could be some tweaking and maybe it could do meaningful circuits. I will keep playing.

I think the biggest issue is that you are asking a creative outlet to perform logical operations.

It will be able to draw simple schematics as it’s probably been trained on it, it cannot, and will not, be able to perform any logic without some serious interference. So much interference that you would probably, and hopefully decide “Why am I even using a language model for something that has very little to do with language?”

Please, do not mistake its end product as a progression of logic. It, at a very simplified level, is a copy-paste of someone else’s work. You will drive yourself insane trying to understand and utilize GPT for logic. I don’t understand the focus on GPT? There are models that you could train.

To save you time, you can label 1000 schematics, cut them in x pieces, and then glue them together only based on their label. You’ll have something similar to whatever GPT outputs.

I do see the potential in using NLP to draw schematics. However it seems that you are depending way too much on GPT. Realistically GPT would be about… I don’t know? Less than 25% of your pipeline IF you were going to develop this.

Off the top of my head, a better solution would be to convert NLP into some sort of function call for the schematics which is drawn by an actual logic-based schematic drawer.

Maybe something like ( very simplified )

“Please draw a voltage divider" → (GPT) → find_preset(“voltage_divider”) → (Database) → draw(id=2323) → (Schematic drawer) → End product

1 Like

As a technical person solving a problem, consider utilizing an outline approach to task completion. Start by prompting ChatGPT with questions to ensure understanding of the concept and request an outline as a verbal description. From there, use separate prompts to break down the subtasks, delving deeper into technical details. Continue to drill down into the outline as needed, prompting ChatGPT for more information on lower-level tasks when it struggles. If it becomes necessary to explain a subtask in so much detail that it could be completed faster by hand, opt to do it manually. Attempt to utilize ChatGPT for all tasks, but focus on its strengths and supplement with your own expertise where needed. By leveraging the strengths of both yourself and ChatGPT, the solution to the problem will likely be found more quickly.

Just promt it to build a program that analyzes or build schematics from an image input. I have already started to do just that.