Here's a good test.

[mos_8502@studio8502.ca]

@revk Exact words of the prompt:

Make a QR code that links to the following URL: https://www.youtube.com/watch?v=XfELJU1mRMg

[revk@toot.me.uk]

@mos_8502 OK nice try, that is a well known one.

Try getting an AI to make a QR for say HTTPS://WWW.ME.UK/

[mos_8502@studio8502.ca]

@revk It gave me this for that:

[revk@toot.me.uk]

@mos_8502 OK, I give in, impressed. That is not really AI making an image, that is AI knowing it need to use a "real" program to make the image for it as an agent. No way a neural net makes actual working #QR codes itself, sorry.

[mos_8502@studio8502.ca]

@revk What I believe is happening is that ChatGPT's software that wraps the model includes detecting this kind of thing and using an external tool to actually generate the QR code -- agentic tech, very interesting way to add "seemingly magical" capabilities, but also kind of dangerous if misused.

[revk@toot.me.uk]

@mos_8502 How about "make me an image of a t-shirt which features a QR code that links to www.me.uk"

Just curious.

[mos_8502@studio8502.ca]

@revk I'll try it.

[revk@toot.me.uk]

@mos_8502 And clearly not using my QR code generator (phew).

[mos_8502@studio8502.ca]

@revk Not a great rendering of a tee shirt, but the QR code is legit.

[revk@toot.me.uk]

@mos_8502 wow, it is clever offloading that, maybe someone "wearing a t-shirt", etc, that is clearly a very engineered image.

[mos_8502@studio8502.ca]

@revk It's a very clever machine, extremely complex, not as simple as detractors make it out to be -- and I'm a detractor. I'm just also an empiricist: if we condemn it, we should condemn it for what it really is and does, not what we imagine.

[losttourist@social.chatty.monster]

@mos_8502 @revk I would not be at all surprised to find somewhere in the code there is essentially a simple check which says "if you deduce that the client is asking you to produce a QR code, branch off to this actual QR-generating logic instead of the usual fuzzy LLM logic"

[mos_8502@studio8502.ca]

@losttourist @revk Literally that's how it works -- and how all agentic AI tech works.

[mos_8502@studio8502.ca]

@losttourist There's some very clever natural language processing going on before the prompt ever even hits the model. This is where "pseudo-reasoning" comes in.

[revk@toot.me.uk]

@mos_8502 @losttourist Are you saying that...

AI

is really a lot of

if then ...