Rainbow Attacks
For those not working on cybersecurity the “rainbow attack” might suggest something linked with the unstoppable rise in the use of rainbow flags in this pride month. It has not really anything to do with it … or not.
But first thing first: what is a rainbow table attack in the domain of cybersecurity ? It is a generic attack on keys that uses a table storing the keys and their hash. Indeed in many systems keys are not stored in clear — it would be too easy to crack— what is stored is the hash of the key. The hash is obtained by applying a secure hash function that is supposed to have “good” properties. One of those properties is the resistance to the so called “pre-image” problem. Suppose that f is your secure hash function, such that f(k)=s. The value s is called the “hash of k”. The problem of pre-image is like this : “fix a goal value g, now find a value v such that f(v)=g”. If f is a secure hash function, this problem should be “difficult” (there are tons of things behind this word but I won’t delve into the rabbit hole here). A rainbow table attack is the dummiest attack you can imagine: you compute all the hashes of the most common passwords and store the pre-image, like an inverse phone book. Then to crack a system you steal the file containing the hashes of passwords and look if there is an entry in your rainbow table.
The AI created the world in its own pre-image
Another desirable property of hash functions is to turn arbitrary long inputs into fixed size outputs. Now when you think of it, it is the exact dual of what chat GPT does. GPT takes a small input, called a prompt, and creates a long output. It is like the LLM is performing a pre-image computation of some unknown hash function. The immensity of the language model is reminding a huge rainbow table. The strange thing is that the space of what your preferred AI will produce is smaller than the space of prompts. By definition a prompt is a small text. This small text is the index of a much larger text. Typically when you request the model to produce a story from few ingredients. So, by definition, what can be produced via those AI is very poor. This is the first rainbow attack: the world of what is possible has been reduced to a skeleton. Using AI for artistic expression amounts to lock oneself up in a room and throw the key through the window.
LGBTQI2A+ army
When showed the ever expanding soup of letter behind the pride month, a mathematician of the 60s might imagine that he is contemplating the output of a hash function. June has become the pride month. The deadliest of the deadly sins is celebrated by the ruling class. The rainbow table attack in the political arena amounts to steer all political debates around a very narrow scope of subjects. Your sexual preferences is one of them (Ukraine, climate change, vaccines etc.). Just like AI that restricts the space of.your experiences to the space of prompts, there is a fixed size area open for political debate. This is more restrictive than the Overton window. The Overton window restricts the field of what can be talked about. The rainbow table attack is a positive force. It transforms everything it touches. It is rather like Midas who turns everything he touches into gold. You can only look at reality thought the lenses of the hash function used. Everything you talk about is turned into a rainbow argument.
I think there is something deep between this feeling. I am not able to figure out it clearly yet —and I am using this substack to make progress. The world appears to be fixed in advance just like the set of passwords you check in a rainbow table attack. Maybe this pride month will bring me additional thoughts about that.