So I'm currently working on the Tiny Cube and one of the features that I need is automatic Wi-Fi onboarding. Because this is going to be an applience in people's homes. So I had ChatGPT 5.6 Sol (high) come up with the infrastructure that we needed to build this and he built it. And for better or for worse: the actual thing worked perfectly.
However, I had the BRILLIANT [sarcasm] idea to show a QR code to connect to the phone in the projected geometry of the cube because I thought I would be cute.
I do know that QR codes have a built-in correction feature, so they have a fault margin. And so I hoped that the QR code being shown on the projected geometry would be readable for most or pretty much all QR readers. I was wrong.
As it turns out, the actual seams of the cube were actually interjecting somewhere in the QR code where it would actually destroy too much of the critical information, which made it fairly unreadable for pretty much any QR tool that I threw against it.
Not my phone, not QR reconstruction tools that I found online. As a least resort, I had ChatGPT 5.6 Sol (high) try. I gave it multiple photos of multiple angles. I told it the name of the network which we knew. And given that we know what/who generated the QR code and what the QR code should be and which settings were applied, we had a lot of information to be found inside of the QR code. That should have given it a reasonable chance of guessing what some of the stuff could be and be better able to reconstruct what the contents of the QR code would be. Kind of how they managed to break the Enigma code back int he 2nd World War. But all of this was to no avail. Not even OpenAI ChatGPT 5.6 could solve it.
So I was left with no other solution than to connect my ethernet cable to the Raspberry Pi 5 that is running within the cube and then connect it to my Mac and then I would be able to connect over the network to the device and read the correct files and solve the password conundrum that I was having.
However, I had one last idea...
I decided to ask Astra, OpenAI's newest most powerful model. Now one of the things I find interesting about the latest developments is that while Anthropic is releasing Fable 5.1 (which is semantically speaking, a minor upgrade), openAI is releasing Astra as a major upgrade from 5.6 to 6.0. So obviously they're claiming this to be the most powerful AI model ever released blah blah blah The question is, is it really that good?
So here we go, in the same thread, after changing the model from Sol 5.6 (high) to Astra Extra High, I'm asking Astra if it's really not capable of doing it.
Astra did it first try within the first two thinking responses it clearly identified a couple of problems and within 8 minutes, it actually produced the fully working QR code. I was able to access my device I didn't have to get the ethernet cables. I was a happy camper.
I am very much aware that this is not any kind of metric that defines how well or how bad or how performant an AI model is, but nevertheless I find it interesting that a problem that seems to be impossible to fix for one frontier model, just two clicks away sits another model that is perfectly capable of handling this. If this had happened just a few days ago, I would have been beaten.
Oh and by the way, the Tiny Cube is making real progress. Almost ready for those of you how care.
If you're interested, subscribe to the list so you're the first to know when they become available.
Tiny Cube Updates
Subscribe if you're interested in knowing once the Tiny Cube becomes available for order. I don't have any exact details or pricing information yet. So this is not an actual order that you're placing. It just allows me to inform you once this becomes available. I'll send you an email with the information and then you're free to decide whether or not you want to order one.