Today’s AI, thousands of years from now…

AI in its current form is intelligent knowledge meant to outlast millennia

It’s been a few years since OpenAI released GPT 3.5 and shook the world. And while it’s still early days to perceive a trend, there’s still an interesting pattern emerging — there has been a constant stream of new AIs popping up. In fact, the rise in AI intelligence is release just a string of new AI model releases — GPT 3.5 to GPT 4, Claude 3 to Claude 3.5, Gemini 1.5 to Gemini 2.5, Llama 3 to…well, to Llama 3.3.

Here is a chart outlining some of the most important language models to have been released since GPT 2.

A chart I made with AI — after going back and forth a dozen times, and having to edit the file personally in the end to fix some blemishes. ‘AI is taking over the world!’

And we are still very close in time to these models since the moment they got released. Case in point, the oldest open source Large Language Model appeared at the end of 2019, when GPT-2 came along. Yes, I was in Uni at the time and so you might call that a different life, but in societal years, it’s still fresh out the oven. Put another way, 91% of people that live today were also alive in 2019 — so most of you know exactly what I mean.

But imagine 100 years from now — where that percentage of people that were alive in 2019 will be much closer to zero? When the state of the art AIs will likely be a force to be reckoned with, what purpose would the AIs that we’ve built today serve — the self-contained Large Language Models? What about 1000 years from now, or even later?

Now, if you’re not so hot and bothered about Large Language Models, a crucial piece of information about all of the open source models* released so far is that they are more or less plug and play. In other words, all of the heavy lifting has already been done during the training phase before the model got released — all you have to do to run it now is to have a beefy enough graphics card, and you can just start up the Large Language Model and start asking it to count the numbers of ‘r’ in strawberry…now or a million years from now. You don’t need the internet, an antenna or anything else really.

If you have these much LEDs in your computer, chances are you can run an LLM on it

And so, back to what purpose could current models serve in the far future. There could be many possibilities, but one that’s staring me in the face is that we’ll use it as sort of a time capsule. People 100, 1000, 1 million years from now could prompt any open source large language model of today with: ‘could you tell me how the vast majority of people were speaking english in 2022?’ and the model will go ‘sure, there was the preppy Cali style of talking and here are some others…’ and so someone could learn about how the vast majority of people spoke in that time. OUR time.

AIs that are released today could be the time capsules of the future…

To realize just how valuable today’s AIs could be for people in the future, imagine if we had this for people that lived 1000 years ago: An AI that knew as much about our society as any open source LLM out there today. Oh, how many questions we could ask! How much could we learn and know about that society; About our past! The condensed intelligence of billions of people surfing the internet of year 1000 would be absolutely priceless. A precious gift to our world.

That is why the open-source community is so crucial when it comes to these models. Because if models are open, they can also be preserved. And when a model is open source, the burden of preserving it suddenly goes from a company to the whole of society. We have no idea if OpenAI’s o3 model will ever make it outside of OpenAI for example, because it’s a closed source model. But we know that if humanity survives, there is a high chance that the Llama 3 model from Meta or another popular open source large language model will live on, keeping the ‘wisdom of the ancestors who made it’ alive.

AI — intelligent knowledge meant to outlast millennia

Bonus idea:

While I was writing this article, it came to mind that there could be also another use for today’s self-contained LLMs, something much more sinister: AI could be a social renaissance kickstarter in case our civilization got lost, provided that we give our forebearers the means to access it. To illustrate what I mean, I made this video with storytveller:

*Many models are technically ‘open weights’, which means that the files which contain their information are given away freely, but they may be under a license that legally precludes companies or other parties to use them, given certain conditions. But no need to get technical…