More AI-enabled amp design work...

[nuke]

I've been chugging along on a from-scratch build. I'm repurposing stuff in my garage to build a 4x10 of my own design. Using a left-over Fender Bassman re-issue cabinet (someone wanted a hand-built cabinet 25 years ago) and a Weber 5F6A chassis that never got used.

I'm using the "pro" ChatGPT in medium reasoning mode. Lately, I'm working on designing the front panel control placement and layout. I want to repurpose the existing chassis as much as possible, and fit more controls and pots than the original bassman-style front panel has. Problem is the holes are not all the right sizes. The existing holes for input jacks and pots are 3/8", the hole for the pilot lamp is 11/16" and the other 4 holes are 1/2" for power switches and fuseholder.

To make it even more fun, the 4 jacks do not share the same centerline as the rest of the controls. The existing holes for pots are spaced on 1.5" centers, as is the pilot lamp hole center, but the 1/2" holes are irregularly spaced.

My thinking is to make a new brass grounding plate, that runs the entire length of the jacks, controls and switches, with holes cut in it that reduce the size of the 11/16" hole and one or two of the 1/2" holes, down to 3/8" for the pots, and down to 1/4" for the signal switches. Also, I'll be including the alignment notches for the signal and power switches.

I'll need a few special washers the thickness of the chassis, to reduce the chassis holes and level them out flush with the top of the existing chassis, so whatever faceplate I come up with, won't get crushed by the nuts holding pots and switches down.

Where AI comes in:

I grabbed the drawing from WeberVST website of the chassis, and checked mine very carefully to make sure it matched. The drawing is pretty good, but missing some dimensions, such as the hole diameters and the spacing between pairs of the input jacks. They also rounded the fractional inches down to only two decimals.

Took a screenshot of the section of the drawing of the Face of the chassis. I uploaded that into my session. Gave ChatGPT a directive to examine the drawing, convert the rounded decimal offsets of the holes back to fractional inches, and build a numbered table of the holes, with their x-y coordinates and diameters.

It examined the drawing in about a minute and half, produced the table and to my amazement, accurately estimated the exact size of each hole from the scale drawing, even though the dimensions were not called out. It didn't quite get the position of the upper two jack holes relative to the lower. I just prompted it with the measured offsets and the table was spot on accurate.

Then I iterated several times with possible control, switch and jack placements on those coordinates, used drawings from a number of knob styles, and produced 1:1 scale drawings in PDF format of the entire 20 inch front panel.

This let me print out the various iterations on my laser printer, on 8.5x11 paper with alignment marks, so I could cut them out and tape them together accurately, then tape them on the chassis to see how it looked. The paper lined up perfectly with the holes in the chassis, verified by shining a flashlight from the inside of the chassis.

That way, I went through several iterations of knob styles and positions, in just about 2-hours of total time.

Now I'm just waiting for a few sample knobs and a sample of a new kind of power switch I'm considering using to finalize which layout I'm going to use. I can even mount the actual chassis with the paper depictions in the actual cabinet to get a really good look and feel before committing.

Since all the drawing info is now fully captured digitally, it will be a simple matter to order a laser cut brass plate and laser-cut special washers from Send-Cut-Send or other online service, once I decide the final candidate.

The savings in time are enormous, not to mention wasted parts orders and all the trial and error of fitting and changing my mind once having things in-hand to look at.

The same captured digial info will feed my front overlay design as well. Once I get some sample knobs in, I'll tackle that next. Still pondering the faceplate implementation, overlay, metal, plastic, or something else. The something else I'm pondering may be pretty cool.

In the meanwhile, I'm going to design a circuit board for the preamp tube sockets to mount on the chassis. I know that goes counter to the "hand-wired" tradition, but dang, it saves me a lot of assembly labor, and I can do it in a way that will allow easy disassembly for tweeking components later. I'll probably do the main board as a printed circuit, just a nicer one than is typical in retail guitar amps and that is designed to be easily accessible. If I can get the board in and out in minutes, I don't really need the top-access of eyelets or turrets, and I can use modern parts that are a lot smaller.

I've yet to find an AI board layout tool that actually works. But honestly, I spend most of my time getting the physical edges, holes, and parts placements to line up with the chassis. AI really, really helps with that.

[Mark]

Tried to use Grok to assist my son with a school project and Grok made impressive drawings and videos, but they were rubbish. I persisted with it for about two weeks and I had to abandon it. I tried Gemini and it doesn’t do drawing which was helpful as I didn’t waste time. I didn’t have Claude at the time. I have ChatGBT but I generally favour Grok.

[Snicksound]

Tried to use Grok to assist my son with a school project and Grok made impressive drawings and videos, but they were rubbish. I persisted with it for about two weeks and I had to abandon it. I tried Gemini and it doesn’t do drawing which was helpful as I didn’t waste time. I didn’t have Claude at the time. I have ChatGBT but I generally favour Grok.

What I did is I had Gemini help me learn to do Onshape and Inkscape to design my own chassis and labeling. I don't want AI to do the work for me, I want it to help me learn new skills. That's how we're directed to use it at work too. And it's great cause when I'd get stuck I'd just take a screenshot and show it to Gemini and it could figure out what was wrong

It also helped me figure out clever way to implement my channel switching using logic ICs and such. It couldn't produce a schematic worth crap, but it was very good at explaining what I needed to do, then review my work.

Still have to keep AI on its toes though, hallucinations show up in ways you don't expect so you still need to double check the math and the datasheet. I'm fine with that. I enjoy refining my schematics myself.

The thing is: if AI makes the schematic itself and you never learn the skill, how can you check its work? That's what we're taught at the office: never use AI to do a task you wouldn't be able to do yourself.

[xtian]

I recently used ChatGPT in live chat mode, using the desktop app, which can see your screen when you ask it to (automatically shoots a screenshot--you just say, "look at my screen"), in order to get up to speed quickly with Blender 3D. It was an amazing experience, like having an expert human on the phone with me. Conversation was fluid and natural. After an hour or so, I bumped up against my daily token limit--either have to pay more, or wait a few hours. Anyway, we may not have flying cars, but we have Dick Tracy wrist watches and expert AI systems. Hello, future!

[Mark]

I recently used ChatGPT in live chat mode, using the desktop app, which can see your screen when you ask it to (automatically shoots a screenshot--you just say, "look at my screen"), in order to get up to speed quickly with Blender 3D. It was an amazing experience, like having an expert human on the phone with me. Conversation was fluid and natural. After an hour or so, I bumped up against my daily token limit--either have to pay more, or wait a few hours. Anyway, we may not have flying cars, but we have Dick Tracy wrist watches and expert AI systems. Hello, future!

I can see why AI is so popular I find it to be a great tool, even if I have a difficult fault I can ask it questions and it searches the internet faster than I ever could for an answer and summarises them. I do look at the source material because it can be wrong, but it’s a great tool to use.

[Mark]

Tried to use Grok to assist my son with a school project and Grok made impressive drawings and videos, but they were rubbish. I persisted with it for about two weeks and I had to abandon it. I tried Gemini and it doesn’t do drawing which was helpful as I didn’t waste time. I didn’t have Claude at the time. I have ChatGBT but I generally favour Grok.

What I did is I had Gemini help me learn to do Onshape and Inkscape to design my own chassis and labeling. I don't want AI to do the work for me, I want it to help me learn new skills. That's how we're directed to use it at work too. And it's great cause when I'd get stuck I'd just take a screenshot and show it to Gemini and it could figure out what was wrong

It also helped me figure out clever way to implement my channel switching using logic ICs and such. It couldn't produce a schematic worth crap, but it was very good at explaining what I needed to do, then review my work.

Still have to keep AI on its toes though, hallucinations show up in ways you don't expect so you still need to double check the math and the datasheet. I'm fine with that. I enjoy refining my schematics myself.

The thing is: if AI makes the schematic itself and you never learn the skill, how can you check its work? That's what we're taught at the office: never use AI to do a task you wouldn't be able to do yourself.

In my case I used AI to come up with possibilities that I hadn’t considered and it came up with hallucinations. I dare say it’s getting better all the time and in a year or will be capable of the task.

[Snicksound]

Tried to use Grok to assist my son with a school project and Grok made impressive drawings and videos, but they were rubbish. I persisted with it for about two weeks and I had to abandon it. I tried Gemini and it doesn’t do drawing which was helpful as I didn’t waste time. I didn’t have Claude at the time. I have ChatGBT but I generally favour Grok.

What I did is I had Gemini help me learn to do Onshape and Inkscape to design my own chassis and labeling. I don't want AI to do the work for me, I want it to help me learn new skills. That's how we're directed to use it at work too. And it's great cause when I'd get stuck I'd just take a screenshot and show it to Gemini and it could figure out what was wrong

It also helped me figure out clever way to implement my channel switching using logic ICs and such. It couldn't produce a schematic worth crap, but it was very good at explaining what I needed to do, then review my work.

Still have to keep AI on its toes though, hallucinations show up in ways you don't expect so you still need to double check the math and the datasheet. I'm fine with that. I enjoy refining my schematics myself.

The thing is: if AI makes the schematic itself and you never learn the skill, how can you check its work? That's what we're taught at the office: never use AI to do a task you wouldn't be able to do yourself.

In my case I used AI to come up with possibilities that I hadn’t considered and it came up with hallucinations. I dare say it’s getting better all the time and in a year or will be capable of the task.

Yes. Hallucinations are a huge issue and there may not be a solution in sight for it. It's a byproduct of how LLMs work and it's important to understand that. That's why you should never use it to do a task you don't understand yourself, because then you can't review its work.

Treat as an assistant that helps you speed up mundane tasks, but do not fully delegate to it.

I have yet to see a single schematic or layout produced by AI that isn't garbage.

But for years I just couldn't figure out CAD software and thanks to Gemini I was able to break that barrier and get a fully custom chassis built. It didn't do the work for me. It helped me learn the software and troubleshoot when I got stuck. And the fun part is that I could get mad and scream at it, it didn't care.

[nuke]

For me, it's getting the machine to do the grunt work, calculating stuff that I know how to do, but would take me more time, and doing stuff like taking a map of all the screw holes and tube socket holes, and control holes in an existing chassis, and translating that into a coordinate system for a brass subplate, and adpator plate for one different footprint device, or a printed circuit board that lines with exact holes in a chassis that I don't have the original drawings for.

On the more EE side of things, calculating surge currents, peak values, bandwidth limits and so on that I also know how to do. That sometimes lets me iterate down a few rabbit holes to see if the numbers work out, or they bottom out and point to something else. It's stuff I'd do on a whiteboard, while doing back of the napkin math. It's just a lot faster with a machine that can do the math a lot faster and revise the drawings.

I've also used it to help "train" me with software packages I'm less familiar with, or to help me update when I've hung unto something like KiCAD for a few versions and updated to the latest stable build.

Being able to play "paper dolls" with front panel layouts was especially helpful. I get spot on accurate, 1:1 print outs that I can tape together and get right on the money with where knobs and jacks and switches go, and decide whether it is sensible or feels too crowded or not logical. In this example that prompted this thread, I even taped the printouts onto a real chassis, then mounted in a real cabinet, and set various sample knobs exactly where they'd go on the real thing. Also mess around with the labels and scales for each control and jack.

That's a thing that's always cost me time, money and aggravation in the past. Thinking something will fit, or look right only to find out that it was wrong. Leads to ordering parts and stuff I didn't need and having to rethink something that I'd been designing everything else around for a while.

LLM hallucinations are a real issue, and they do make mistakes, big ones. So in the hands of someone who can't discern whether stuff is right or wrong, that's the recipe for trouble.

Like "product managers" who don't need software engineers anymore. Like a pilot that doesn't know how to fly and just hopes the autopilot gets him to Denver somehow. It might work, but hope ain't a plan.

[Guy77]

I recently used ChatGPT in live chat mode, using the desktop app, which can see your screen when you ask it to (automatically shoots a screenshot--you just say, "look at my screen"), in order to get up to speed quickly with Blender 3D. It was an amazing experience, like having an expert human on the phone with me. Conversation was fluid and natural. After an hour or so, I bumped up against my daily token limit--either have to pay more, or wait a few hours. Anyway, we may not have flying cars, but we have Dick Tracy wrist watches and expert AI systems. Hello, future!

This is an amazing feature Xtian! I could definitely use this. Is this the $20usd/mnth plan that gives you this feature?

I have been using the free version to help me calculate plate voltages for preamp tubes based on the dropping resistors used in an amps power supply and on the preamp tube anodes, it surprised me how accurate it was.

Cheers
G

[xtian]

Is this the $20usd/mnth plan that gives you this feature?

Yes.

[nuke]

Yeah, the paid-for versions are much better, and setting the reasoning level to a higher setting gets much better results. Takes more time and burns tokens faster, but the quality improves dramatically.