More happy use of AI in amp designing...

[nuke]

I know the AI thing makes some people groan. I'm planning a new scratch build and doing the math to figure out things, as that's what engineers are supposed to do.

So I'm working out the mains side power circuit. I wanted to calculate in a defensible way, things like fuse ratings, whether I wanted an NTC inrush and what value and part number, the switch ratings, the hardware, terminals for the wires and circuit board to make it pluggable, and select parts that are well stocked and available, add up the prices, estimate floor planning for the board size.

Part of the in-rush calculations involve the rectifier and filter capacitors and loads on the secondary side, and how they reflect from the secondary side to the primary side. Also goes into selecting a fuse that provides safe interruption, without giving rise to nuisance fuse trips.

It's all stuff I know how to do. But it is also a good bit of tedious work. Then there's the running down of parts, searching online catalogs, getting CAD drawings and footprints, comparing cost and availability.

It's this kind of stuff that agentic AI can help with. I put in some data from a transformer I selected already, and some proposed component values for the rectifiers and filters and the count and types of tubes on the filament string. Then I tasked it with calculating the current loads as seen from the primary, along with the math to support the numbers. It correctly said it was lacking some information about the transformer, so I put in that it was a commercial standard EI core and gave the approximate dimensions of the care stack and to infer the values by estimating.

In just about a minute, it produced all the calculations supporting the numbers. I checked the work, as I would check a junior engineer's work. It was all correct. And fast.

Then we went down the path of selecting and vetting parts, sources, costs, wire gauges, crimp terminals (I have PIDG tooling) assembly methods, estimated board space for components and necessary trace clearances, and even what hardware and standoffs were needed.

In the end, I produced a BOM table, with actual costs and availability, which it downloaded directly from distributor online price lists.

In about an hour.

That's a lot of leverage and time savings.

[jbrrrrr]

It's great, in the hands of people that can do what they're asking the model to do for them.

It's a horror show in the hands of people that are using it to do things they aren't able to do themselves.

[R.G.]

Aye, there's the rub.

If you already know how to do it, an AI can run through the calcs and list making very, very quickly. And your expertise can spot any oddities that it just made up. But for those of us who watched Disney's Fantasia, the Sorcerer's Apprentice lurks in the shadows.

The process of a person doing all the grunt work themselves is what develops the understanding needed to check the AI's work. Starting as a beginner with an AI helper directly impedes and slows developing the skills to distinguish solid work from hallucinations; and worse, little hallucinations hidden inside a big slug of solid work.

[Snicksound]

Yeah it can be a double edged sword... I like it for brainstorming ideas and doing quick math I can't be bothered with when just ideating stuff.

But then I start running the numbers myself and realize a crucial element was missed.

But it sure "feels" good to have a willing assistant you can talk with about this stuff all day long.

It's pretty good at math, it's also pretty good at being very confident while hallucinating.

[Snicksound]

Snippet from a funny recent discussion (note this is the "Pro" version of Gemini). I asked a question to the Gemini iteration within Google search, then asked the same thing to the standalone Gemini Pro on my phone and the two answers were contradictory (had to do with voltage handling of a specific IC)

At least it was honest this time... Don't trust it. Double check.

Screenshot_20260817-102545.png

[xtian]

I use Zoho Billing to track my time/customers/invoices. But creating an invoice requires a lot of clicks, repetitive steps that are the same each time, and Zoho won't save my settings. I asked ChatGPT to help me automate the drudge work. It looked up the API, and we spent a couple hours working through the steps and crafting a solution. Now I have a Python script that launches a local web portal that shows me my clients with outstanding hours, and with one click, it creates a draft invoice. A second click lets me review the invoice, and a third click sends it. So much faster!

If I had hired a programmer, it would have cost a couple thousand dollars.

[martin manning]

It's great, in the hands of people that can do what they're asking the model to do for them.
It's a horror show in the hands of people that are using it to do things they aren't able to do themselves.

If you already know how to do it, an AI can run through the calcs and list making very, very quickly. And your expertise can spot any oddities that it just made up. But for those of us who watched Disney's Fantasia, the Sorcerer's Apprentice lurks in the shadows.

The process of a person doing all the grunt work themselves is what develops the understanding needed to check the AI's work. Starting as a beginner with an AI helper directly impedes and slows developing the skills to distinguish solid work from hallucinations; and worse, little hallucinations hidden inside a big slug of solid work.

Spot on!

[nuke]

I'd say a year ago, we weren't there yet. I was using GPT5.6 in normal (fast) response mode for this particular task, because it wasn't a hard task and it didn't require a lot of deep reasoning.

At this point, it is a good torque multiplier for getting white collar grunt-work done.

In the Garage section of the forum, I posted how I used it to create drawings for an adaptor plate to fit a smaller power transformer into a chassis that was cut for a larger transformer. All I told it was the aftermarket Hammond replacement part numbers and to create an adapter plate. It went off and downloaded the PDF drawings from Hammon, read the dimensions of the mounting screws and in less than two minutes, I had a first draft in autocad and PDF formats. I knew the center cutout size wasn't depicted on the Hammond data, so I supplied the dimensions I wanted for the center cutout, to reduce the outer dimensions to the same as the larger transformer, told it to enlarge the screw holes to 1/4" and to select appropriate rounded corners. In about 45 seconds, I had updated drawings. I checked them in a .dxf viewer and dropped them to Send-Cut-Send for fabrication.

I'd been procrastinating on that little project for a long time, mostly because of all the tedious steps and I'm not great at mechanical CAD. Send-Cut-Send is only 1-day by regular mail and I had my plates in hand in two business days.

For workflow automation, have a look at n8n. It's a directed graph system where you literally draw out your steps. It isn't directly AI itself, but it can make calls directly into AI engines of your choice, which is great if you need to automate something that requires interpretation. Say you wanted to watch a busy contact email box and escalate issues and make summaries of the last 24 hours. I implemented such on an email system, with local AI, a smaller faster LLM running on local hardware. It gathers email and processes them all, creates 24-hour activity summaries, looks for known customer addresses, and sifts through spam and sorts out and summarizes possible important emails.

It's the kind of stuff a junior office assistant might do, but it all runs locally on less than $10k worth of hardware.

So for me, it is an effort and productivity amplifier.

[Bretly]

Good luck trying to get it to help source parts for a Music Man 2475-65. Or to help understand the hand drawn spec. Google image itself was surprisingly good at that part though.

My overall experience is thats it is great at well documented stuff like an AB763. Not so much at all for less famous less documented stuff.

I find it to be lazy and the gaslighting is next level. No Claude, everything isn’t a pro move and wasn’t boss at all.

Problem being they are incentivized to answer, right or wrong. They will never say. “I do not know” or “I’m not sure”

You gotta put hard rules and guardrails down to get them to act the part.

[R.G.]

I'm doing some coding work for a client. It's tinkery, fiddly work, so it's been slow. I got multiple suggestions to just use AI to help me code it. Got it to working, but needs polishing by hand.

OK, OK, I can be open minded. I described the app in pretty clear detail, parts, pinouts, connections, what controller insides got used for what, and tossed it into an AI coder. It "thought" for about 30 seconds and came back with "OK here's your code. I took the liberty of..."
Looked like code, all right. A couple of places looked funny, and I asked questions. The responses were variants of "Oh, I made a mistake there. The actual [stuff} should be...."
After a few go-rounds of this it produced listings that looked pretty good. I started reading through it and ran into places where things were just flat wrong, but in detail and assumption. Asked the AI to recheck all its register usage and initializations.
That got about seven pages of "Oh, there are some problems there. You were justified in asking for a re-check. Here are the things I found wrong on the re-check, and suggestions on [stuff] for several pages.

This went on for three hours before I got the mistakes corrected by pointing them out. Some places got fixed more than once. And it still won't compile cleanly.

That is, it was very much like dealing with a complete newbie.

[drewspriggs]

As an experiment, I asked ChatGPT about finding a transistor with a certain pinout and specs I could use.

I did this three times on three separate days.

I was given 7-13 options each time. First time had one that was correct, second time had zero, third time had one. None of the transistors with the correct pinout are available.

All of the "evidence" it linked showed the exact opposite to what it was claiming.

If you know what you're doing, it can make simple jobs very slightly simpler but hard jobs much harder. If you don't, it's never going to work.

Is that worth destroying the planet and worlds economy for?

[R.G.]

By accident, I ran into a pertinent quote:

"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." - Edsger Dijkstra.

[didit]

I'm doing some coding work for a client. It's tinkery, fiddly work, so it's been slow. I got multiple suggestions to just use AI to help me code it. Got it to working, but needs polishing by hand.
[...]
That is, it was very much like dealing with a complete newbie.

Curious. Would you be able to expand a bit more on the target uController system, and the AI coding toolkit and workflow models you experimented/experienced? Not to pry into the client's applications, simply the foundation technology spaces.

..

[R.G.]

Curious. Would you be able to expand a bit more on the target uController system, and the AI coding toolkit and workflow models you experimented/experienced? Not to pry into the client's applications, simply the foundation technology spaces.

You're thinking fancier than I used. The task was to use a controller to monitor certain real-time events, collect analog data both at the events and at intervals between them, condense the readings down into lights and beeps for a human to interpret, and to do protection tasks that the human might be too slow to interpret before damage happened. It was at the edge of being too big for a PIC controller, and too expense sensitive for an overkill ESP or Arduino to be embedded. PIC it was, then.

I used C language for the code. I had a hand-coded version operating, but some real time events that needed cleaned up. Given that, I didn't feel called to go research tookits and workflow in the fancier meaning that those have today. I tried dumping my existing code with additional design instructions into the AI. The AI didn't like my code.

Then I backed off to just the description of the design and let it go code it from zero. It liked its own code fine - except that the code left things out, didn't cope with the real time events well, and had both logic and coding errors. I wound up leading the AI through the code because I had a good understanding of it from my coding. I got the funny feeling that the AI was using ME as a coding assistant.

Probably if I had decomposed the whole task into a series of sub tasks - initialize the internal peripherals, describe the interrupt structures, subroutine the basic tasks, then integrate the sub modules into an overall schema - it would not have been as trying. It just seems that if I've done all that ahead of time, the AI isn't doing me much good except for looking up the syntax and generating the obvious code statements.

Dag nabbit.

[didit]

Thanks. I've an indefinitely deferred fairly complex PIC based project that a few former colleagues have been taunting me to delegate to a coding harness. Fortunately I'm still too busy with other stuff to delve recklessly, at least for the moment. Maybe in calming coming doldrums of an inevitable Ottawa January.

..