back
17 comments
This is fantastic.

With the developments in AI it won't be long before you can just drop into ancient Rome and have the entire VR world generated on the fly with all the NPCs acting as they would based on the AI's understanding of the era taken from every contemporary document it has access to.

I predicted a few years ago that as AI improves we'll move closer and closer to being able to build an (inaccurate but close) VR time machine. The alarming thing I've discovered with AI is that you don't need a huge corpus of prior communications to reasonably emulate somebody's personality.

None of us are as unique or amazing as we think we are :)

(weirdly I got an almost identical output to the writer in the article even though I modified the prompt he used and asked for a hyper realistic photograph: https://imgur.com/a/uizM07g )

And we will believe all the lies AI will tell us about the past because we have no way to verify - what an amazing future.

Alternatively, we will only understand anything in ancient Rome if we understand Latin and other languages because the NPC stay true to that.

The biggest issue would be that Romans viewed certain things so different to us, so our ability to understand would need heavy support to work.

> And we will believe all the lies AI will tell us about the past because we have no way to verify - what an amazing future.

That's already the case for most of history. Many events were only written down centuries after the fact by biased historians. Even further back in time we can really only make vague guesses from archeological evidence.

Whether it's AI or a human coming up with historical fiction, it will never be exactly as it was. On the other hand an AI can be trained on everything we know about a given period, which is very unlikely for a human fiction author.

> That's already the case for most of history. Many events were only written down centuries after the fact by biased historians. Even further back in time we can really only make vague guesses from archeological evidence.

And that's assuming you spent significant effort learning history on your own, acquiring knowledge from contemporary and past historians. If you're relying just on history lessons from schools, then that has an extra layer of political and cultural bias on top.

That extra layer is fortunately the easiest one to deal with - it peels off once you start diffing your history books with equivalent ones from a different country (or even from your own country, but couple decades earlier). But it still requires some investment (time, knowing another language), and hardly anyone tries it.

Yes, which is why it is important to understand the gaps and biases in sources. Extrapolating into them (AI or not) is tricky.
> And we will believe all the lies AI will tell us about the past because we have no way to verify

How is it different from the popular history you have now? Unless you've spent years on studying actual facts, most of what you know is the opinionated interpretation by someone else who did. And even if you did, most sources about Rome are either secondary or of dubious quality.

Not that different for a lot of people.

Btw., sources on Rome and everyday life in Rome are actually plenty and from the time.

Aw man, this is gonna be like Civilization with hyper-aggressive Gandhi all over again.
My own curiosity, what did they view so different? While I have read a number of translated texts I was unfortunately never exposed to stark differences.
Their moral views were quite different from ours. For example, they would have trouble viewing depictions of something as "bad"/moraly corrupting (the corruption would be in the mind always, not in the picture which is passive), so our concepts there would be quite alien (thinking about the Late Roman Republic).

Similar for violence in some areas - much more opportunitistic there (not quite the right word, they were kind of hard-nosed pragmatists in a way).

>I predicted

Star Trek was a few decades before you.

Every morning we will get a fresh menu of new episodes of HBO-level historical re-creations.

And today I'm feeling like... Chernobyl Day Three !

I have my concerns about using it for historical recreation, but for fantasy worlds, that's actually a great tool

You could make a very big, very rough world based on a define set of core ideas and then deliver it to real authors to make it a cohesive and fun background for a new book or movie series, for examples.

Like giving a big marble piece in the rough shape of a human so that real human sculptors can make a suitable statue out of it.

An actual realistic scene in Pompeii would feature a lot more horse shit all over the streets. Of course, the idealized images are what saturate media, so even specifying "true to life" in the prompt won't change that.
Has there been any breakthroughs or sign of a breakthrough in term of LLM training efficiency and inference cost? Right now it seems like the capabilities of the models are the focus with more accurate image generation, videos, less hallucinations, etc. But it always at the cost of computing power. Average guy can't afford a $600 GPU to run Stable Diffusion at home at full capacity...
Depends on where the average guy lives, even in Eastern Europe $600 is something you can save up for if it's your hobby. A used 3090 is around $550 now and it's only going to get cheaper.
Make it work, make it right, make it fast.

Most players are still in the make it work part or before that. Only OpenAI is in the make it fast part.