re: #477775

In short: wisdom can always be gained from the past, and while the concepts behind certain events seem to resemble those of the past, the specific events and patterns that sit on top of the concepts are usually novel.

In other words, as Mark Twain said, “History never repeats itself, but it does often rhyme.”

Specifically, this challenge comes from @kr, who adds: “This came to my mind as I was thinking about data collection in AI and whether there is value in collecting more historical data if it does not improve the performance of an AI model”. (thanks to @kepford for asking for that context!)

I’ll start with AI, and bring it back to history. And then get overly meta and weird.

LLMs can’t rhyme. Well, they actually do a much better job now than they did about a year ago. But I’m not talking about word rhymes anyway. I’m talking about rhyming concepts.

It has been shown in multiple experiments that while ChatGPT can predict patterns, it does not understand the concepts behind the patterns. As soon as the concept is presented in a novel way that falls outside of its training data, it fails miserably.

I think that’s because, as this article describes: “LLMs learn in the opposite direction from humans. LLMs start out learning language and attempt to abstract concepts. Human babies learn concepts first, and only later acquire the language to describe them. So LLMs are doing it backward. In other words, perhaps reading the internet might not be the correct strategy for acquiring intelligence, artificial or otherwise.”

And I would argue that LLMs never actually learn or understand the concepts.

Therefore, it makes sense that just giving an LLM more data about history won’t necessarily help it better understand the underlying concepts that would meaningfully inform the future.

But more knowledge about history can help us humans, if our goal in both learning history and attempting to predict the future is to absorb concepts over specific events (insert bitterness about memorizing names and dates).

Think about coding alongside an LLM. If you’re building something new(ish), it’s likely that the pieces of the thing you’re trying to build aren’t necessarily new by themselves, but the concept and arrangement of those pieces is new and creative. The LLM will be helpful only to the extent that it can build the individual pieces, but if you ask it to build the entire thing to fit a new context, it will fail. (I think that’s why senior engineers who can easily picture the entire architecture of a project get more use out of AI than junior engineers who need to work through the creative process in order to develop the holistic view.)

My conclusion is, that’s exactly what trying to predict the future is like. The future almost always involves familiar concepts in a new context, which will inevitably establish a novel pattern of events that is unfamiliar to AI (it “falls outside of the training data”). That's why I would argue that all history is useful if we learn the concepts.

Here are the meta/weird/probably annoying thoughts:

Perhaps the reason history never truly repeats itself is because predicting the future is paradoxical in the first place. As soon as you know how a specific event will play out… isn’t there a high likelihood, if not certainty, that your knowledge alone will change the outcome? Perhaps that’s why our human attempts to predict the future tend to fail anyway.

Tenet is an awesome, super underrated movie imo. As soon as I saw the OP (#477775), I thought of the quote: “ignorance is our ammunition”.

If you haven’t seen the movie (or have, but don’t understand it), the basic plot is: They’re fighting against people in the future, and the more they know about their own strategies and desired outcomes, the more likely that knowledge will change the outcome and/or become known to their enemy in the future.

That’s still probably confusing but… the basic idea is this: In order to overcome the problems of the future, the protagonist specifically has to not predict the future, but rely on human instinct, core concepts, and “faith in the mechanics of the world.” 


All that to say:

Obviously future plans should be incorporated into our lives, but at the end of the day, we have to factor in the truth that we simply don’t know if those plans will succeed. Today is all we have access to, and it's is the only time we’ll have access to it. Maybe our goal should be “weaving another [better] past into the fabric of the mission” instead of being anxious about tomorrow.

Idk, just thinking out loud here. I might look back at this in a couple of weeks and smh.

Thanks to @kr for the OP and @kepford for asking for the context, this was super fun to think through.