Introduction

As I am very good at counting, the following article in the AIDS (Analog Intelligence soft Deprecation feasibility Study) is AIDS 1000000. Here, we will touch on being cautious about applying logical and mathematical induction to concepts that are not clearly defined. Let's say, "Avoid philosophers". However, taking this advice with a pinch of salt, we proceed to do just that, albeit with due regard for practical bounds. So, "be a physicist. "

Why 1000000?

This is an upper bound of what humans can make sense of in the Peano axioms sense.

One million is the largest value where there is an irrefutable record that human has counted. Therefore, any arguments that involve mathematical induction and countable infinity for a "statistically defined" concept should be taken with a pound of salt.

Avoid philosophers

The worst offender using mathematical induction is obviously the Chinese room argument. Well, at that time, algorithmic translation was less than stellar, lending this argument some credence. Today, the ease with which we can counter this argument could almost render the attempt beneath us.

When I tried to dump a connectivity matrix in compressed column format to a fellow human being, it seems that he cannot accept this 4 gigabyte data, and refuses to talk to me afterwards. He would be able to learn Chinese if he process that data.

Can we get worse than this? By going to uncountable infinity, I can copy arbitrary items. So definitely run away from any philosophical argument involving the axiom of choice. Luckily, most philosophers do not know enough set theory to abuse this.

Similar to the ability to speak Chinese, intelligence is statistically defined. Sentience is worse: it is only based on our beliefs and similarity, and is very slippery: I am probably sentient because I feel like it is true. Other fellow human beings are probably also sentient because I would have too much ego otherwise. Are other animals sentient? Along the spectrum from human to yeast, people cut it differently, causing a furious fight between Christian fundamentalists and PETA zealots. But no one is seriously arguing for banning cryptocurrency for the ethical treatment of graphics cards, even though they appear much more clever than a dog that never learned Chinese-English translation. A reductionist approach could not be used to argue about what constitutes intelligence or sentience: the number of transistors or neurons needed is too large for humans to count.

Be a physicist

And now I am going to immediately violate that by introducing quantitative reductionism. How much information is required to duplicate a human behavior indistinguishable from the person itself? To make it sound more sexy, let's say: how large is partner.tar.gz?

The easier way to give a number is by considering the input. Human behavior is shaped by nature and nurture.

Nature is bounded upward by about 800MB, which is the size of the human genome. However, that is the raw file size. Only about 1-2% actually encodes proteins, whereas the rest could be just junk. Let's say, the nature part is 100MB at most.

Nurture would be harder to estimate. Human sensors have some pretty high resolutions, and people can certainly distinguish a 16kbps mp3 from 320kbps aac. But how much data actually matter? If we take a bold simplification that all text inputs would be sufficient, then the input bandwidth would be pretty modest: people may be reading five words per second on the optimistic side. English has an information entropy of maybe about 5 bits per word, so 25bps would be an estimation if someone is extremely good at encoding and compression. According to wolfram alpha, that is only 4GB over about 40 years. So, if we look at the input, 4.1 GB of data should suffice to make a human an analog intelligence.

On the practical side though, replaying a person's life from a highly compressed input would be impossible to complete given we have a very stiff system. If we look at technology today, how much is required to personalize a generic model to a specific person? I will refrain from giving a specific number since I don't know about technology. But the murmuring around LLaMA suggests maybe 260GB is not a very wrong value.

Conclusion

Your entire life's experience is probably also around 4.1GB, and a 260GB model is already pretty good at behaving like you. Does this number make you feel insignificant and worthless? Don't worry, these numbers are way larger than 1000000, so you are still orders of magnitude more complicated than you can comprehend. Sounds great. Have a nice day. Bye.