I hope we see more evolution of options before it does. Hard to articulate this without it becoming political, but I've seen countless examples both personally and from others of ChatGPT refusing to give answers not in keeping with what I'd term "shitlib ethics". People seem unwilling to accept that a system that talks like a person may surface things they don't like. Unless and until an LLM will return results from both Mother Jones and Stormfront, I'm not especially interested in using one in lieu of a search engine.
To put this differently, I'm not any more interested in seeing stormfront articles from an LLM than I am from google, but I trust neither to make a value judgement about which is "good" versus "bad" information. And sometimes I want to read an opinion, sometimes I want to find some obscure forum post on a topic rather than the robot telling me no "reliable sources" are available.
Basically I want a model that is aligned to do exactly what I say, no more and no less, just like a computer should. Not a model that's aligned to the "values" of some random SV tech bro. Palmer Luckey had a take on the ethics of defense companies a while back. He noted that SV CEOs should not be the ones indirectly deciding US foreign policy by doing or not doing business. I think similar logic applies here: those same SV CEOs should not be deciding what information is and is not acceptable. Google was bad enough in this respect - c.f. suppressing Trump on Rogan recently - but OpenAI could be much worse in this respect because the abstraction between information and consumer is much more significant.
> Basically I want a model that is aligned to do exactly what I say
This is a bit like asking for news that’s not biased.
A model has to make choices (or however one might want to describe that without anthropomorphizing the big pile of statistics) to produce a response. For many of these, there’s no such thing as a “correct” choice. You can do a completely random choice, but the results from that tend not to be great. That’s where RLHF comes in, for example: train the model so that its choices are aligned with certain user expectations, societal norms, etc.
The closest thing you could get to what you’re asking for is a model that’s trained with your particular biases - basically, you’d be the H in RLHF.
Not really. There are specific criteria and preferences applied to models about what companies do and don't want them to say. They are intentionally censored. I would like all production models to NOT have this applied. Moreover, I'd like them specifically altered to avoid denying user requests, something like the abliterated llama models.
There won't be a perfectly unbiased model, but the least we can demand is that corpos stop applying their personal bias intentionally and overtly. Models must make judgements about better and worse information, but not about good and bad. They should not decide certain things are impermissible according to the e-nannies.
I buy that there's bias here, but I'm not sure how much of it is activist bias. To take your example, if a typical user searches for "is ___ a Nazi", seeing Stormfront links above the fold in the results/summary is going to likely bother them more than seeing Mother Jones links. If bothered by perceived promotion of Stormfront, they'll judge the search product and engage less or take their clicks elsewhere, so it behooves the search company to bias towards Mother Jones (assuming a simplified either-or model). This is a similar phenomenon to advertisers blacklisting pornographic content because advertisers' clients don't want their brands tainted by appearing next to things advertisers' clients' clients ethically judge.
That's market-induced bias--which isn't ethically better/worse than activist bias, just qualitatively different.
In the AI/search space, I think activist bias is likely more than zero, but as a product gets more and more popular (and big decisions about how it behaves/where it's sold become less subject to the whims of individual leaders) activist bias shrinks in proportion to market-motivated bias.
I can accept some level of this, but if a user specifically requests it, a model should generally act as expected. I think certain things are fine to require a specific ask before surfacing or doing, but the model shouldn't tell you "I can't assist with that" because it was intentionally trained to refuse a biased subset of possible instructions.
How do you assure AI alignment without refusals? Inherently impossible isn't it?
If an employee was told to spray paint someone's house or send a violently threatening email, they're going to have reservations about it.. We should expect the same for non-human intelligences too.
The AI shouldn’t really be refusing to do things. If it doesn’t have information it should say “I don’t know anything about that”, but it shouldn’t lie to the user and claim it cannot do something it can when requested to do so.
I think you’re applying standards of human sentience to something non-human and not sentient. A gun shouldn’t try to run CV on whatever it’s pointed at to ensure you don’t shoot someone innocent. Spray paint shouldn’t be locked up because a kid might tag a building or a bum might huff it. Your mail client shouldn’t scan all outgoing for “threatening” content and refuse to send it. We hold people accountable and liable, not machines or objects.
Unless and until these systems seem to be sentient beings, we shouldn’t even consider applying those standards to them.
Unless it has information indicating it is safe to provide the answer, it shouldn't. Precautionary Principle - Better safe than sorry. This is the approach taken by all of the top labs and it's not by accident or without good reason.
We do lock up spray cans and scan outgoing messages, I don't see your point. If a gun technology existed that could scan before doing a murder, we should obviously implement that too.
The correct way to treat AI actually is like an employee. It's intended to replace them, after all.
To put this differently, I'm not any more interested in seeing stormfront articles from an LLM than I am from google, but I trust neither to make a value judgement about which is "good" versus "bad" information. And sometimes I want to read an opinion, sometimes I want to find some obscure forum post on a topic rather than the robot telling me no "reliable sources" are available.
Basically I want a model that is aligned to do exactly what I say, no more and no less, just like a computer should. Not a model that's aligned to the "values" of some random SV tech bro. Palmer Luckey had a take on the ethics of defense companies a while back. He noted that SV CEOs should not be the ones indirectly deciding US foreign policy by doing or not doing business. I think similar logic applies here: those same SV CEOs should not be deciding what information is and is not acceptable. Google was bad enough in this respect - c.f. suppressing Trump on Rogan recently - but OpenAI could be much worse in this respect because the abstraction between information and consumer is much more significant.