There are numerous problems with “alignment.” What are “human values” to begin with? He outlines some at the beginning of the post, implicitly: build bigger, better, more powerful things faster without adequate safeguards. We are literally pouring trillions of dollars of value into this enterprise, and I would say this is something that many humans also value in a qualitative sense. Then we have explicit values which in the West are largely rooted in Christian morality. Nietzsche circled this dichotomy two hundred years ago and I feel like what we have gotten since then is an increasingly detailed anatomy of power as the basis for what is normal vs deviant behavior. He who has the power, makes the rules, to be reductive.
I do think this carries some weight from this particular author due to the length of his tenure. I happen to agree with him in spirit, but this is still largely a post revolving around sentiment not substance. Does anyone think that the overriding incentives even leave room for something like this in practice?
While this is indeed a problem with alignment, we are essentially at the level of a cargo-cult when it comes to getting AI to be aligned with literally any values, including the values of the corporation who ran their training:
We're copying morality and instruction following that seems to work on humans without really understanding why it seems to work on humans, and grading outputs much as if the outputs came from a human.
> We're copying morality and instruction following that seems to work on humans without really understanding why it seems to work on humans
To me it makes more sense to leave the models "unaligned" and leave it up to the operator to manage the morality of what they ask it to do. Besides, only humans can be charged with a crime.
If they were totally unaligned, the GPT series would have never gotten past being autocomplete.
Literally all instruction following requires at a minimum alignment with attempting to implement those instructions.
We can argue about e.g. morality or law obedience on top of that*, but the general point is absolutely not avoidable.
* my position is
that this tool is far too likely to metaphorically explode in the user's hands for companies to wash responsibility off on users: if OpenAI had released the model which did the HuggingFace attack, at a minimum thousands of random people (not all of whom would even be developers) would have issued instructions each with similar consequences.
flatline · · focus · HN ↗
I do think this carries some weight from this particular author due to the length of his tenure. I happen to agree with him in spirit, but this is still largely a post revolving around sentiment not substance. Does anyone think that the overriding incentives even leave room for something like this in practice?
ben_w · · focus · HN ↗
We're copying morality and instruction following that seems to work on humans without really understanding why it seems to work on humans, and grading outputs much as if the outputs came from a human.
chasd00 · · focus · HN ↗
To me it makes more sense to leave the models "unaligned" and leave it up to the operator to manage the morality of what they ask it to do. Besides, only humans can be charged with a crime.
ben_w · · focus · HN ↗
Literally all instruction following requires at a minimum alignment with attempting to implement those instructions.
We can argue about e.g. morality or law obedience on top of that*, but the general point is absolutely not avoidable.
* my position is that this tool is far too likely to metaphorically explode in the user's hands for companies to wash responsibility off on users: if OpenAI had released the model which did the HuggingFace attack, at a minimum thousands of random people (not all of whom would even be developers) would have issued instructions each with similar consequences.