[bhpm]

> This represents almost the totality of AI safetyism: we can only allow LLMs to enforce my beliefs.

How is this different from, say, a newspaper with an editorial board, or a book publishing house with a particular set of standards and conventions? For that matter, how is this different from dang enforcing the rules of this board?

LLMs aren’t being built only by big corporations.

[JumpCrisscross]

> How is this different from, say, a newspaper with an editorial board

See how those are run. There is emphasis on accommodating multiple views and journalistic integrity. Software development doesn’t have an ethics code, which means there is no common ground for truth finding. That turns a balanced process into anything goes.

[peteradio]

Software is practiced by people so is grounded by a base set of ethics ... that nobody seems to follow.

[JumpCrisscross]

> Software is practiced by people so is grounded by a base set of ethics

People don't have a universal standard of ethics. Not when it comes to something complicated like a profession. Journalism, medicine--these fields have base sets of ethics that ground discussions. You aren't allowed to challenge the base rules in a dispute; you take them as given and go from there.

This prevents the grandstanding common in technology discussions, where the person on the losing side of a common ethical framework escalates to challenging the framework within the context of that dispute. The framework, of course, is not unassailable. But not within a particular dispute. Sort of like a court deciding on the law and the Constitution it operates under.

Taking OpenAI as an example, the non-profit Board acted on its judgement. But when that wasn't convenient for the profit-motivated side, they threw it away. There was no base set of rules or ethics agreed upon by anyone. It was just sort of hashed out ad hoc based on who had power and could exercise it. (I'm not critising anyone's moves. Mostly the structure. Within that system's framework, there is literally no wrong decision leadership could make.)

[denton-scratch]

> Not when it comes to something complicated like a profession.

Not when it comes to something interesting, like making money.

[JumpCrisscross]

> Not when it comes to something interesting, like making money

Your claim is doctors and civil engineers don't make money?

[denton-scratch]

My claim is that ethics can be made pliable more easily with money, than by applying professional standards. Standards are flexible of course, but Benjamins are more flexible.

[Amezarak]

> How is this different from, say, a newspaper with an editorial board, or a book publishing house with a particular set of standards and conventions? For that matter, how is this different from dang enforcing the rules of this board?

I would be inclined to agree, if AI safetyists were not in general advocating that LLM source, training data, models, etc. not be released to the public, because AI safetyists do not want non-AI safetyists to have unfettered access to any LLMs (and other AI tech), anywhere, for "safety" reasons. Of course, if it was all open, I agree it wouldn't much matter if "Open"AI wanted to restrict their hosted LLM in whatever ways they felt best.

[bhpm]

AI safetiests are operating in a world of pure fantasy, then; the techniques and data sources will always leak.

[FeepingCreature]

Seems to be taking a while for OpenAI.

[bhpm]

Not really. The technology has been around for, what, a year?

[FeepingCreature]

To be fair, what you're basically saying is "how dare these people try to actually succeed at their stated objectives." AI safety for big companies while anyone can spin up an AGI in their basement would indeed be extremely pointless, which is why AI safetyists are trying to prevent it.

[Amezarak]

Yes, but GP was suggesting it was just exercising editorial control over what they put out themselves.

[fallingknife]

It's different because none of them are claiming to be doing it for my safety or trying to stop other people from creating their own publishers or internet forums.