← All posts

What Teaching a GAN to Play The Sims Taught Me About the LLM Era

  • #generative-ai
  • #gans
  • #llms
  • #perspective

In 2021, before “generative AI” meant a chat box, I spent a weekend teaching a machine to dream up characters for The Sims.

There was no dataset of Sims faces, so I made one the dumb, wonderful way. I opened the game’s character creator, wrote a script to hammer the “randomize” button for about an hour, recorded the whole thing, and sliced the footage into thousands of individual faces. Then I took a StyleGAN pre-trained on celebrities, pointed it at my homemade dataset, and watched. For hours the output was noise. Then smudges. Then, slowly, faces: game-accurate, uncanny, and new, faces that had never existed.

I still remember how it felt to watch the blobs resolve. Not because the result was useful (it wasn’t), but because I had touched every part of it: the data, the model, the training loop, the failure modes. When it broke, I knew why.

What the LLM era got completely right

Five years on, almost none of that ritual survives, and mostly that’s progress.

  • You don’t build the model anymore. You call one. The gap between “idea” and “working generation” collapsed from a weekend to a sentence.
  • You don’t build the data anymore. No more scripting a video game at 2am. The models arrive knowing more than any dataset I could scrape.
  • It’s for everyone now. In 2021, generating an image meant a Colab notebook and a tolerance for CUDA errors. Now it’s a text box your parents can use.

That democratization is real, and I don’t want it back the old way. Scale won, and it won for good reasons.

What the GAN era had that’s worth keeping

But something did get lost in the trade, and I think it’s worth naming.

When you build the whole pipeline yourself, you develop an intuition for where models lie. You learn that a generator will happily produce something plausible and wrong: a face with three-and-a-half eyebrows that your loss function is perfectly happy with. You learn that the demo always works, and that the edge cases are where the truth lives.

That instinct, a healthy suspicion of a confident output, is exactly what the LLM era needs most and teaches least. It’s easy to trust a system you never had to debug. The people I see using today’s models best aren’t the ones who prompt the fastest. They’re the ones who still ask “how would this fail?” out of muscle memory.

The takeaway

You no longer need to build a GAN from scratch to work in AI, and thank goodness. But the mindset the GAN era forced on you (own the data, distrust the output, understand the machine even when you don’t have to) didn’t expire. If anything, now that the machines are easy, it’s the only edge left.

I don’t miss scripting The Sims at 2am. I do miss knowing exactly why my model was wrong. So these days I try to keep asking.