In the context of modern neural nets I keep observing the "tough Siberian bloke" effect from the old joke.

Ha, look, they don't know left from right on the road; ha, look, they get confused about temporal details; ha, you can obviously tell it's a neural net!

Yes, you can. For now.

I've deliberately attached what art nets produced at the end of 2022 — exactly three years ago — next to fresh examples.

They didn't understand context; they couldn't even hold a line. They made a kind of mince across the whole frame, and even then it was a wow, because — well, it's cool, software generating a picture. Mixing colours, grasping some fun bits like wide angle.

I wrote this post on Facebook back then, because a strong wave of indignation had begun.

Three years on, we have what we have. Fresh models understand context in different languages, work carefully with references, and produce results you're not ashamed to show a client — in FMCG, for instance. They follow instructions attentively and are precise in the details. Some of it is already built into software, Photoshop for one. There's an API and thousands of sites for the truly slow, if you don't want to bother with settings. Certain tasks — restoring focus to a blurred image, say — work better and better. If you don't know how to compose a prompt, you can ask an LLM for help. Will Smith eating spaghetti does so quite convincingly and without artefacts, i.e. video too.

Over the past year of my work in video production there hasn't been a single job that didn't use neural nets in some way. Sometimes they handle almost everything. The speed at which all this happened is terrifying and impressive in equal measure.

Yes, you can still tell — if you know where to look.

And what happens in another three years?

If the question of "how" gets less relevant by the day, how do you get back your understanding of what you want?

And do you even want anything at all in this new world where everything is possible in two clicks?