When AI meets ancient Rome https://share.google/5MWBXCl98CpEwLkyT
Another AI article....I'll let you decide
I think it's amazing how well the tech is coming on. Nicely rendered faces, in particular, I think. But, if the AI is trained on real images, why does it produce legionnaries is Hollywood style leather muscled cuirasses with pteruges?
The chariot procession is fun too. I can see three spears in the picture but only one is being held. The other two are out of reach of everyone. Blow the picture up to full size and you will see there seems to be a horse missing. One might say it's easy to nit-pick but surely if it's obvious to me, who is not a classical academic, why aren't the originators, one of whom is, spotting this and correcting it pre-publication?
Quote from: Erpingham on Aug 28, 2025, 06:08 PMI think it's amazing how well the tech is coming on. Nicely rendered faces, in particular, I think. But, if the AI is trained on real images, why does it produce legionnaries is Hollywood style leather muscled cuirasses with pteruges?
The problem is I suspect hollywood style armour appears in equally real images.
I wonder what the algorithm uses when selecting info and pictures...does it look at academic papers for instance
Quote from: Imperial Dave on Aug 29, 2025, 04:55 AMI wonder what the algorithm uses when selecting info and pictures...does it look at academic papers for instance
Well, the article says
"trained on research literature and ancient sources to generate scientifically sound images". I think academic papers would classify as research literature. I presume what they are keen to remove are Victorian versions, modern reconstructions and TV/film. I'd suggest they also excluded re-enactors, as I don't see any re-enactor influence in the images.
Quote from: Imperial Dave on Aug 29, 2025, 04:55 AMI wonder what the algorithm uses when selecting info and pictures...does it look at academic papers for instance
Whilst anything is possible, I'm not confident that 'AI' in general will be looking for answers in academic papers that much right now. Perhaps in a brighter future it may.
I know nothing about how it all works but I suspect that a commercial AI scours the internet looking for content to plagiarise into its imaginary pastiche (you have gathered I'm not yet a fan). It will find huge amounts of inaccurate content from Hollywood, video games and other media compiled by people who don't know or don't care. Whilst many academic papers are now available online many are not, and the same goes for books, but let's assume the AI does access all the academic papers in the world. What next?
Presumably, the AIs available to us commercially are intended primarily not to maximise human enlightenment but instead to maximise profit. I also presume that their 'analytical thinking' has been curated by their commercial creators to give their intended audience the greatest amount of content commensurate with the majority of customers' preconceptions and desires. The great commercial success of, for example, superhero movies suggests a very large paying audience prefers such fare over rigorous historical drama. Why would a commercial AI be intentionally biased, or even allowed to 'choose', to produce much less exciting historically accurate material which its creators suspect will not monetise anywhere near as well as a more fantastical or Hollywood version?
Coming back to this particular story (a non-commercial project at this stage), I note that "The AI tool doesn't produce definitive visualizations but rather invites users to experiment with different scenes." Whilst I will happily argue that we can never be 'definitive' about anything in the ancient world, I'm not sure that's what the creators mean here. I suspect it's their wriggle room to excuse the inevitable absurdities and inaccuracies that critics may level at it. The fact that the authors accept that inaccurate Roman sandals are generated doesn't suggest to me that the inputs are that effective yet. It may certainly be a better attempt than many others but it still cannot be relied upon. Importantly, I think it does seem to confirm that different AIs can be designed to pay more attention to some forms of evidence over others, and in that there is some hope.
In conclusion, I will now fall back on my prejudice. The AIs we are currently being offered commercially seem to be entirely derivative creating new pictures by copying existing pictures. They don't currently know what is real and what is false, they just copy with their virtual fingers crossed behind their virtual back. For the majority of their customers it will appear a marvellous innovation and a 'revolutionary' tool. For those with a little more insight it will disappoint. The danger for historical accuracy is that these AIs will further entrench by endless replication the inaccurate tropes that already exist in popular culture. Some may favour that, some may not.
This reminds me of an earlier thread (https://soa.org.uk/sm/index.php?topic=4597.msg60816;topicseen#msg60816) that discussed the Photoreal Roman Emperor Project which "using the neural-net tool Artbreeder, Photoshop and historical references, created photoreal portraits of Roman Emperors."
Indeed, looking at the guy driving the chariot, I am convinced the AI has derived his likeness from the Emperor Augustus who, as we all know, had his bust modelled by Vladimir Putin.
:P
Quote from: Adrian Nayler on Aug 29, 2025, 10:49 AMI know nothing about how it all works but I suspect that a commercial AI scours the internet looking for content to plagiarise into its imaginary pastiche (you have gathered I'm not yet a fan).
Sounds reasonable, especially as that's basically what human brains have been doing for millennia. Probably the major difference between AI and human brains is that AI is much better at accurately plagiarising than we generally manage.
The other thing to remember about AI is that it's an emergent algorithm. Someone doesn't sit down and design how it does things, instead they design a learning engine, and then are surprised at what it manages to learn how to do. Which is why AIs have the unfortunate habit of making racist and other inappropriate comments.
Quote from: Erpingham on Aug 29, 2025, 09:12 AMQuote from: Imperial Dave on Aug 29, 2025, 04:55 AMI wonder what the algorithm uses when selecting info and pictures...does it look at academic papers for instance
Well, the article says
"trained on research literature and ancient sources to generate scientifically sound images". I think academic papers would classify as research literature. I presume what they are keen to remove are Victorian versions, modern reconstructions and TV/film. I'd suggest they also excluded re-enactors, as I don't see any re-enactor influence in the images.
A lot of academic literature is effectively behind a pay wall. I suspect that will not be searched as it would involve cost and paying people for their work
The more I come across AI the less I find a need for it in my life
Indeed.
I have a work colleague who uses AI to write emails and it grinds my gears