switch to dark theme

My take on AI safety

I am writing this page because of my concern about AI safety and regulations. I would like to share some useful resources to learn more about the issue and spread the word about it.

Note that this is my point of view. While I try to rely on reliable sources and verify my own knowledge, it might be biased or incomplete.

00 — The issue

The issue with AI is that very little is known about its inner workings, and that the world seems to be rushing towards better and better AI systems.

In the following, I will refer to AIs as models, because this is what they are: mathematical functions that, given an input, yield an output that is supposed to maximize a certain human-defined score.

The way an AI is trained doesn't dictate the inner mechanisms that enable it to achieve its goal. They actually emerge from the training process. This is because it consists of choosing a starting point in a certain parameter space and exploring this space in the direction that minimizes the global error of the model. Each point in this space represents a specific configuration of the parameters of the model.

Therefore the main concern is being able to control the behavior of any AI. When the training process does not introduce any control over the way the model achieves the goals it was given, this becomes a significantly tougher task. The underlying issue here is that it is hard to formulate a purely objective goal that takes into account all possible negative outcomesbpskph. This is called the alignment problem.

This alignment problem is what makes AI dangerous. Currently, companies and states seem to be rushing towards better and better AI models and as explained before those are not perfectly aligned. Even though imperfectly aligned AIs don't pose immediate threat to humanity, they might become hazardous when reaching the stage of artificial general intelligence.

An artificial general intelligence (AGI) is an AI that can perform as well as a human on all tasks one could perform from their computerqveifp. This is where AI starts being dangerous. It might make many jobs redundant but more importantly it could start designing better AIs which would in term design even better AIs, and so on. This is what could be called an intelligence explosion and could lead to an artificial super intelligence or ASI which is an AI that can perform better than all humanity unitedqveifp.

Those hypothetical super intelligences might become so intelligent that they would develop projects out of human comprehension. And if humans are in the way, they might not consider protecting them a necessity, much like how human expansion led to the extinction of many species.

This is why some think continuing the frantic race towards more and more intelligent AIs is risking human extinction if no action is taken against it in a very close future. This is quite concerning in my opinion.

01 — The solution

The solution lies in three important stepsnlcdqu:

  1. Set up strong and global AI safety regulations, which should be done as soon as possible
  2. Ensure the resilience of the regulatory system in order to avoid losing control (again)
  3. Solve the alignment problem to build new safe AIs after the first two steps are secured

Here is how you can make things change and contribute to step 1:

10 — Resources and references

I strongly recommend looking into those youtube channels and watching these videos as a short and informal introduction to the subject:

Those are some interesting websites which explain the issue (much better than me) and help take action individually, here are some of them: