Why understand logic when matricies are there to save the day?
“ Mathematics is the language that God has written the universe in ”
Humans are fascinating creatures. We are so cool that we don’t even realize it. The progress that we have been able to achieve in just the 10000 years of our existence is nothing but mind blowing. From the development of ammonia to reaching the cosmos, we have done it all, and are just improving in every way we can.
Much has been debated about the intelligence of humans and where it has originated from, but for sure we can mimic mother nature through one of the most powerful languages of all time - Mathematics.
“ A scientist who states a law of nature with the aid of mathematics is taking an essence of the real world. ”
Artificial Intelligence is closer than ever to mimicking what nature has gifted us with, and of all the applications that could exist for AI, the one I am about to walk you through might be good…
We humans have been sending signals into the cosmos for a long time in order to search for answers to the universe. Assuming the validity of the Drake equation, there must be some form of extraterrestrial life, but so far all approaches have converged to a single result - we are alone in this universe. Being a space connoisseur and a big fan of Carl Sagan, I absolutely loved the Voyager missions and the Golden record they carried. These capsules of humanity have all the information about us, and are a remarkable message for human civilization. Thinking about all of the recent advancements in AI and these messages, I hypothesized a question(which I find truly fascinating)
“ Are there any methods to test whether the messages we send into the cosmos are understandable by other civilizations ? ”
Many people have pointed out the inaccuracies and “human-ness” of the golden record, but imagine yourself in the aliens’ shoes and try to see whether you can intercept an unknown message from an unknown sender that doesn’t speak your language, and is a little biased about the terminologies that they have adopted… It is by no stretch of the imagination a daunting(and a lot scary) task. What if we could find ways to communicate through some form of interlingua, and what other alternative is there than mathematics? Only if we could go inside their heads and see what is going on when they see this message… wait a second, we can actually model that :
Assuming no previous knowledge or anything about our pale blue dot, we can devise a mental model* of the alien’s interpretation of our data using a Multi-Modal Large Language Model. Let me explain in detail to the best of my abilities.
Multimodal Large Language Models are a step closer to understanding how we make use of our sensory organs to make an intelligent decision(Cognitive AI). I propose a class of multimodal language models - named XXXXXX, which makes use of Text, Video, and Sound data(similar to the one that was used in the Golden record(albeit in a different way)). We design these language models assuming different levels of understanding for the civilizations, and testing these models for interpretability using some model agnostic interpretation method.
The notion of interpretability allows us to discover meaning between our input data and attribution and the model outputs. We can make use of these metrics(for example, Shapley values) to better guide our inputs. Doing this has 2 benefits :
- It allows us to better design the messages that we broadcast for humanity.
- It allows us to have a systematic class of language models to classify alien species based on their intelligence levels(similar to Kardeshiev scale, but for Intelligence)
If you have come up this far, there must be many questions. Now I would try to explain my chain of reasoning by answering some questions that really guided me to this idea.
Think of mental models as someone’s thought processes or explanation of how something works in the real world. They are quite imperfect in the sense that there are many inconsistencies, but they are very useful for understanding. An analogy that helped me in understanding this goes as follows :
If you are a little familiar with Game theory, this will be a cakewalk. Game theory is a way to represent models of situations such as negotiations(read about Prisoners Dilemma), relationships, trust etc. In a sense, game theory is a mental model that encapsulates all of these situations, and we can use it to arrive at a solution(if it exists)
Similarly, a machine learning model(again, to some extent) can be treated as a mental model which helps us make sense of data and relationships between them.
In our case, a mental model is a large language model which tries to make sense of the data that is given to it. It is beneficial because apart from just reading all of that data, a large language model during pre-training tries to understand complex relationships between the data that might make them “understand” it.
Q2. How do we know whether these mental models have learnt something or not? Is there a way to explain it?
It’s time to learn about a good machine learning concept - Interpretability. Interpretability is a great way to understand what exactly the model is learning and justifications for it. The objective function with which the model is pre-trained, cannot accurately capture the real-world costs associated with model decisions.
In our case, we can draw parallels between the model’s interpretability and the advanced civilization’s interpretability of the data we send. There are several standard ways to deduce model interpretability such as using Shapley values, or LIME. This interpretability data can go a long way in understanding how we can deduce what data to send, and maybe can tailor different versions of the same data for different levels of civilization.
Q3. Were the compression techniques used in the Golden Record too much? We might have had a better chance with images, texts, and audio?
The encoding techniques(converting all of the images, text, audio into audio signals) for the golden record are not really generalizable techniques to represent information, at least in my opinion. What if the ‘advanced’ civilizations are in their stone age era, and the best use of the golden record is perhaps using it as a weapon or a plate(however bad that it may sound, that could be one of the possibilities). We can’t just send them a USB drive for the same reason. How we should send those data is not something that this project aims to explain. I am working on interpreting what the other civilization might interpret our data, so that we can send actually useful data into the cosmos.
Maybe having a mathematical representation of these data is a step to understand the mental model of the other civilization.