←
MG26,383 posts
MG avatar

MG

@_MG_

Nightmare Reifier. I sell some of them too: o.mg.lol & hak5.org/omg

◷ Joined April 2008▧ 3,459 Photos and videos
MG avatar
MG@_MG_·17 May 2021

OMG Cable - The New Batch Now in USB C, the implant is much smaller, but it’s even more powerful than before. Smartphone/tablet attacks, extreme long range triggers, geofencing, etc. o.mg.lol

1:23
MG avatar
MG@_MG_·18h

And it needs repeating: I was intentionally providing as little assistance as possible to see how much the model could fill in the gaps. It’s a big part of what the real job entails.

MG@_MG_23h

Just to be clear, this was a torture test worse than anything Anthropic bans you for exposing their models to: putting it in my shoes 6 years ago. But seriously, I’m playing the role of a sadistic ghost of Christmas future. A good designer isn’t just slapping together a pretty schematic and routing the copper. Its layer and layers of mind reading and crystal ball use. 🧠🔮 -Intent reconstruction based on sparse requirements -customer experience -product support implications -field recovery -Operating environment & real world context -User behavior and misuse anticipation -the difference between “it works” and “it always works” -Component selection based on margins -Interoperability -EMI -RF certification -other regulatory requirements -fault tolerance -safe failure -mechanical integrity -thermals -hardware & firmware co-design -design for manufacturing & assembly -supply chain resilience -product testing, calibration, & tractability -component aging And I am forgetting a lot. But many of these are competing constraints which complicates it all even more. I just wanted to see how far the latest models could go with close to zero help or setup beyond a fresh install of KiCad. Right now it’s still struggling with the copper design, but starting to poke deeper into the other layers.

MG avatar
MG@_MG_·18h

And on Astra 6.2 because it was just released today:

MG@_MG_18h

“do Astra 6.2” … fine. It’s mostly worse than prior tests with Astra 6 and Fable 5.2. (see replies below for prior tests) The one thing it did better than all others: anticipating OS behavior quirks that require extra components to work around. Everything else: Board is about as ugly as the rest, but at least the LED labels are next to the LEDs, and they are equally spaced. Better than Fable. Routing is the worst of all the tests. Not only did it use unnecessary vias and layer switching, it decided that this simple board required 4 layers. WTF!? The schematic is as painful to look at as Astra 6.

Post attachmentPost attachmentPost attachment
MG avatar
MG@_MG_·18h

I ran the same test on Fable 5.2 and saw some improvements:

MG@_MG_Nov 17

Hand-Made PCB vs AI-Made PCB Round 2! This time I used Fable 5.2 instead of Astra. Black is the hand-made PCB. Green is what Fable made. I’ll do a few comparisons below in the thread.

Post attachmentPost attachment
MG avatar
MG@_MG_·18h

Test with Fable 5.2:

MG@_MG_Nov 17

Hand-Made PCB vs AI-Made PCB Round 2! This time I used Fable 5.2 instead of Astra. Black is the hand-made PCB. Green is what Fable made. I’ll do a few comparisons below in the thread.

Post attachmentPost attachment
MG avatar
MG@_MG_·18h

First test with Astra 6:

MG@_MG_Nov 16

Hand-Made (black) vs AI reproduction (green) Experiment: use Astra to reproduce a simple design that I previously made by hand. I’m surprised what it produced on the first attempt after very basic instructions. It’s ugly, too big, signal quality is bad, lots of routing is really nasty, multiple unnecessary components, and there are multiple hardware bugs. But… it can actually function! I didn’t expect it would function without more guidance and revisions. It even figured out some “off label” uses of components that I had implemented… wtf man… I felt clever when I first thought of that and now AI is just one-shotting it in minutes.

Post attachmentPost attachment
MG avatar
MG@_MG_·18h

“do Astra 6.2” … fine. It’s mostly worse than prior tests with Astra 6 and Fable 5.2. (see replies below for prior tests) The one thing it did better than all others: anticipating OS behavior quirks that require extra components to work around. Everything else: Board is about as ugly as the rest, but at least the LED labels are next to the LEDs, and they are equally spaced. Better than Fable. Routing is the worst of all the tests. Not only did it use unnecessary vias and layer switching, it decided that this simple board required 4 layers. WTF!? The schematic is as painful to look at as Astra 6.

Post attachmentPost attachmentPost attachment
MG avatar
MG@_MG_·23h

Just to be clear, this was a torture test worse than anything Anthropic bans you for exposing their models to: putting it in my shoes 6 years ago. But seriously, I’m playing the role of a sadistic ghost of Christmas future. A good designer isn’t just slapping together a pretty schematic and routing the copper. Its layer and layers of mind reading and crystal ball use. 🧠🔮 -Intent reconstruction based on sparse requirements -customer experience -product support implications -field recovery -Operating environment & real world context -User behavior and misuse anticipation -the difference between “it works” and “it always works” -Component selection based on margins -Interoperability -EMI -RF certification -other regulatory requirements -fault tolerance -safe failure -mechanical integrity -thermals -hardware & firmware co-design -design for manufacturing & assembly -supply chain resilience -product testing, calibration, & tractability -component aging And I am forgetting a lot. But many of these are competing constraints which complicates it all even more. I just wanted to see how far the latest models could go with close to zero help or setup beyond a fresh install of KiCad. Right now it’s still struggling with the copper design, but starting to poke deeper into the other layers.

MG@_MG_Nov 17

Hand-Made PCB vs AI-Made PCB Round 2! This time I used Fable 5.2 instead of Astra. Black is the hand-made PCB. Green is what Fable made. I’ll do a few comparisons below in the thread.

Post attachmentPost attachment
MG avatar
MG@_MG_·Nov 17

What stood out most was how Fable was thinking about downstream consequences of a design when it comes time to actually use the hardware. No other model caught this. I distinctly remember getting bitten by this 6 years ago when I was first designing it. It only cost me an extra day of work because I could just re-spin a new prototype on my PCB mill, but this extra set of eyes would have been very nice.

Post attachment
MG avatar
MG@_MG_·Nov 17

Fable’s schematics were much nicer for human eyes. Symbols for ground and power. Things actually connected with visual lines. Etc

Post attachmentPost attachment
MG avatar
MG@_MG_·Nov 17

Fable’s routing was much better. It didn’t do all the unnecessary vias, layer changes, & redundant traces. Routing was much more optimized. The only recurring issue was the 90deg bends everywhere.

Post attachmentPost attachmentPost attachment
MG avatar
MG@_MG_·Nov 17

Fable took 2x longer than Astra. Looking at the 3D render, Fable's looks a bit worse than Astra's. The LEDs aren't evenly spaced out and the labels are too far away. Fable also chose to add a 5th LED that I didn't initially ask for.

Post attachmentPost attachmentPost attachment
MG@_MG_Nov 16

Hand-Made (black) vs AI reproduction (green) Experiment: use Astra to reproduce a simple design that I previously made by hand. I’m surprised what it produced on the first attempt after very basic instructions. It’s ugly, too big, signal quality is bad, lots of routing is really nasty, multiple unnecessary components, and there are multiple hardware bugs. But… it can actually function! I didn’t expect it would function without more guidance and revisions. It even figured out some “off label” uses of components that I had implemented… wtf man… I felt clever when I first thought of that and now AI is just one-shotting it in minutes.

Post attachmentPost attachment
MG avatar
MG@_MG_·Nov 17

Hand-Made PCB vs AI-Made PCB Round 2! This time I used Fable 5.2 instead of Astra. Black is the hand-made PCB. Green is what Fable made. I’ll do a few comparisons below in the thread.

Post attachmentPost attachment
MG avatar
MG@_MG_·Nov 16

All that setup before you can start laying down copper is tedious and AI largely handled it fine with no real world implications on the PCB. At least for this simple design. Except if a human has to review the schematics… No visual symbols that speed up visual understanding, like visual symbols for GND. But I will give allowance to the AI on this one. All those visual improvements are going to be extra clutter for how AI processes this data.

Post attachment
MG avatar
MG@_MG_·Nov 16

The copper routing is probably worse than the already terrible but way more computationally efficient auto-router. It doesn’t just look bad, it has real world implications. Ex: any high speed signal would likely fail over routes like the above post. But there is a ton of setup that needs to happen before you can start drawing traces. Schematics, symbols, footprints, pinouts, etc.

MG avatar
MG@_MG_·Nov 16

An example of “technically functional, but WTF is this routing”. AI vs how I would do it with 10sec of effort:

Post attachmentPost attachment