The Perfect Information Problem

Thanks to Large Language Model's ability to read large amounts of text in infinitesimally small time compared to humans, soon the world will enter an era where perfect information is truly available for all decisions. At first, it will be painful for most to ignore their own opinion in the face of an AI summary of thousands of lines of documents, but over time, I believe it will become the social norm to appeal to AI summary as a means of argument justification.

A single article, post, or message from a random person on the internet is almost guaranteed to be erroneous in some way, but a sum weighted average of all the entries on the internet can distill the truth by democratizing away the risk of bias. Its not a perfect system, but this is how humans efficiently absorb information today on a much smaller scale. We absorb the opinions of others and second hand accounts as positive examples of the existence of phenomena. "Did you hear about the lady in Florida that let an alligator eat her husband?" Every account opening our minds to the possibilities of the universe that would otherwise be unthinkable. Imagine an epidemic of husband eating alligators in Florida. After one month of daily national news reports, the idea of an alligator eating a person would become commonplace.

This thought experiment reveals the locality bias in any system trained on only positive examples. The accuracy of your answer to the question, "How likely is an alligator to eat a human?", depends on the weighted average of the number of times you have been exposed to the story "My husband was eaten by an alligator."

Lets say I want to be more certain about the answer to the alligator question. I'll have to go read all those articles myself and uncover details about each case and decide for myself if there were extenuating circumstances. It turns out each story has a link, all the attacks happened in the same neighborhood and the alligator was somebodies pet. Suddenly it seems to me alligator attacks aren't so common. My research done, I can rest peacefully knowing I've done my due diligence on the problem and I am ready to regale guests at my next cocktail party with how the alligator stories are actually overblown and there's no need to worry.

My party guests would listen to my second hand account of someone else's second hand account and be relieved to hear from a reliable source, an economist and computer scientist no less while we're appealing to authority 🎉, that they have nothing to worry about. The question remains however, "Do they have nothing to worry about?".

Further Reading: Thinking Fast And Slow

Your brain is privy to the competency of the person reporting information and choses to absorb information from reliable sources more readily than those that are unvetted. If we go back to the example of the alligator question and think about how much reading and research I had to do just to be a little more certain about the possibility of alligator attack, it becomes clear doing that for every question I have everyday quickly becomes intractable. Because we have limited energy and time, Humans evolved to be social creatures first and information processing machines second. We use the accounts of other humans who have first hand specialized domain knowledge of phenomena as an approximation. 😏

WARNING: OPINION BELOW how ironic

I don't really see why anecdotal evidence gets such a bad rap. Because we engage with anecdotal evidence at basically every choice its essentially the stock and trade of the human condition. People are pretty keen on when its appropriate to appeal to authority/anecdotal evidence in general. You wouldn't say , "Because my friend Bill thinks so" in the context of a PhD dissertation. We should be less verbally critical of people who propagate anecdotal evidence and instead just be aware that it is anecdotal and that we live with finite time and not everything needs to be 100% certain.

You know the story by now. In come Large Language Models and suddenly it becomes reasonable to perform all that research for the alligator problem in less than 10 seconds. In the same way Google search propelled humanity forward by helping them find relevant documents, Natural Language models of all types will allow humans to shed even the need to read those relevant documents. Its as easy as reading 128,000 words in 10 seconds.

It will undoubtably be a long and painful process letting go of our firmly held and biased beliefs. At the risk of sounding like a shaman, I do believe it will offer a type of healing society needs right now. Over time, our reliance on anecdotal evidence has been developed into a tool of oppression and obfuscation against those who lack access to opportunities. The very nature of the perfect information problem is that the less you have seen about the world the more you are likely to believe any evidence that is purported as credible.

Through the technology of democratized aggregate summary, we will be able to move the idea of perfect information out of the realm of philosophy and into the hands of millions.