🚨 AI alignment and the questions being overlooked.
If you agree with this, please like and retweet to maximise reach.
1) What are we aligning to?
By now you probably have a rough idea of what people mean by AI alignment, and how it could help us to avoid a direct AI risk. The less talked about aspect, however, is what will shape our alignment process? What is our North Star?
Such a question is of huge importance, alignment defined by a single organisation, country, or individual, will no doubt lead to bias, or at least a limited view of a true human perspective, if such a thing exists.
Social media is a goldmine of human perspectives and opinions, possibly offering a way to give everyone the opportunity to have their say, but without a deterministic method for quantifying these opinions, using it for AI alignment could also pose a risk.
A quick thought experiment on this:
- Ask every human adult if they agree or disagree with the following statement, "thou shalt not murder".
- Answers are collected on a scale of 1 (fully disagree) to 10 (fully agree).
How would this help us?
- Alignment data structured in this way makes interpretability a much simpler task.
- Answers provided across different cohorts (nationality, age, gender, etc) can help us to remove bias of any specific cohort, giving us that true human perspective.
- Avoids alignment decisions being made by small groups of individuals.
Ultimately you need to ask yourself, do you have faith in humanity?
2) What are the indirect risks of AI?
Pop culture is filled with theoretical AI extinction scenarios, I have lost count of the number of terminator movies that have been released, and yet indirect risks are more likely to pose a societal risk before we even get close to direct AI risk.
The singularity will permeate everything, education, business, economics, and many more. While these are concerning non-direct risks, I feel the most critical right now is geopolitical.
Another thought experiment:
- Imagine a game of Chess, with a twist.
- Both white and black have a "mutually assured destruction" button, when pressed it denies victory for the other side, but guarantees destruction of both sides.
- Now imagine only one side has an "AGI" button, when pressed it lets the AGI take over control of that player's decision making in the game.
Now ask yourself this, do you trust the side without AGI to not push the destruction button when faced with defeat?
3) How much ��intelligence” is enough?
Let's compare artificial intelligence to human intelligence, how much is enough? 10X? 100X? 1000X?
Is there a limit that we are willing to accept? Even if we can solve all other risks, the endless pursuit of more will no doubt lead us down a path to unforeseen consequences.
According to @Keir_Starmer
Apparently I’m a ‘far right thug’
I want to follow fellow ‘far right thuggish’ ordinary UK citizens
Please like,repost and follow ❤️
#FarRightThugsUnite#TwoTierKeir
@thejessicatapia @cooltechtipz Discovered this in Montonegro recently and honestly, every other time I got an iced coffee without ice cream, I was disappointed.
@kimmonismus Damn, I’m glad I missed it 😂 OpenAI’s livestream though was beyond impressive, and makes me think an omnimodal GPT-5 agent will effectively be AGI.