So this morning I asked this of AI:
How should I make a 50 ohm dummy load capable of handling 5 watts from 330 ohm 1/4 watt resistors?'
Here was the answer:
Oops. Nope. That makes a load of about 20 ohms, not 50.
Wow, this seemed like a fairly simple request. AI got it wrong. If AI (LLM) can't get a simple question like this right, you have to wonder how many other things it is getting wrong.
I'm not sure if the AI has got through the licencing exams yet. Some of the circuit diagrams it comes up with are 'experimental'. HI HI HI
ReplyDelete:)
DeleteAlso, there's an AI practice exam tutor at
https://projecttoucans.com/
KO6BTY and I wrote the app. It uses ChatGPT under the covers.
I think it would be nice to know which AI model you used. Asking ChatGPT and Claude the exact same question gave me the correct answer.
ReplyDeleteYou can also give it directions to “think harder” or to be sure of its answer and not guess.
I just plugged the question into Google and selected their "AI overview."
DeleteOne thing of learned is different ai LLMs are better at some things than others. The scary part is the will sometimes be completely wrong but sound very convincing. One of my current favorites is Grok expert mode. I submitted your same question and the response can be found here:
ReplyDeletehttps://grok.com/share/c2hhcmQtMi1jb3B5_a858a562-f23c-4854-8259-76d919bda2ff
Hi Bill. LLMs are bad at math. If you're on an AI that can do it, you can try this
ReplyDelete"Please tell me how to build a 50 ohm dummy load out of 330 ohm, 1/4 watt resistors. Using those instructions, write python code to calculate how many resistors I'd need. Run the python on your code interpreter and print the result here in our chat."
The first time I asked, I forgot to put the power rating, and I got this answer back
"To build a 50 Ω dummy load from 330 Ω, 1/4 W resistors, the clean “exact math” approach is a series-of-parallel-banks network:
Build instructions (exact 50.0 Ω)
Make 5 identical banks.
Each bank is 33 resistors in parallel.
One bank’s resistance:
𝑅
bank
=
330
/
33
=
10
Ω
R
bank
=330/33=10Ω
Wire those 5 banks in series.
Total resistance:
10
+
10
+
10
+
10
+
10
=
50
Ω
10+10+10+10+10=50Ω
Total resistors:
5
×
33
=
165
5×33=165
Power rating (ideal sharing):
165
×
0.25
W
=
41.25
W
165×0.25W=41.25W
(Real-world: derate for heat, airflow, resistor tolerance, and uneven sharing—mount with good thermal management.)
Practical build tip: on perfboard, each “bank” can be two bus rails (top/bottom) with 33 resistors bridging across; then connect the five banks end-to-end in series. Use a metal enclosure and keep leads short to reduce inductance at HF."
Then, I ammended my quetion to
"Please try again keeping in mind that the dummy load only needs to handle five watts. The 50 ohm value can be plus or minus 12%."
and got this answer back
"Cool — with 5 W and ±12% allowed, you can use way fewer parts.
Key constraints
Each resistor is 1/4 W, so for 5 W you need at least:
5
/
0.25
=
20
5/0.25=20 resistors (ideal sharing)
Target resistance window:
50
Ω
±
12
%
⇒
44
Ω
to
56
Ω
50Ω±12%⇒44Ω to 56Ω
Smallest solution that meets BOTH power and resistance window
The smallest count I found is 24 resistors:
Option A (24 resistors, 55 Ω, 6 W total)
Build 2 banks in series
Each bank is 12 resistors in parallel
Math:
One bank:
330
/
12
=
27.5
Ω
330/12=27.5 Ω
Two in series:
27.5
+
27.5
=
55
Ω
27.5+27.5=55 Ω ✅ (within 44–56)
Power (ideal):
24
×
0.25
=
6
W
24×0.25=6 W ✅ (≥ 5 W)
Wiring:
Make Bank 1: tie all 12 resistors’ one ends together (node A), and all the other ends together (node B).
Make Bank 2 the same (node C to node D).
Connect B to C (series link).
Dummy load terminals are A and D.
A slightly “nicer” near-50 Ω option (still not many parts)
Option B (26 resistors, ~50.77 Ω, 6.5 W total)
2 banks in series
13 resistors in parallel per bank
Math:
Bank:
330
/
13
≈
25.385
Ω
330/13≈25.385 Ω
Total:
≈
50.77
Ω
≈50.77 Ω ✅
Power:
26
×
0.25
=
6.5
W
26×0.25=6.5 W ✅
Python code I used + result (run)
This searches “k banks in series, m resistors in parallel per bank” and finds the smallest build meeting 44–56 Ω and ≥20 resistors: ..."
But yeah, without running code, they're pretty bad at math.
Here's the link to the chat
https://chatgpt.com/share/6969179d-6c08-8001-b93b-c20d2990ff34
I've used AI for quite a few things now and been disappointed by pretty much all of the results, on of the most recent things was a 'vibe coded' digital VFO which worked right up until it needed to change the output frequency, then it went haywire.
ReplyDeleteIt was using bits of code ripped from all sorts of places and the documentation for the hardware side of it was, at best, not going to work, at worst would have left you with a fried Si5351, TFT display and Pi Pico.
I've tried getting it to correct itself but even though the AI acknowledges it's got things wrong, it's not capable of fixing them and often digs itself into a deeper hole before it errors out and gives up.
There's no current substitute for doing the work for yourself or at least being able to so you know how to fix the garbage AI spews out.