It was surprisingly easy to do once you internalised the word prediction foundation and thought about the corpus of writing you have ever read in a “tropes” fashion. For instance it’s really easy to get it to give you some “mills and Boone” romance novel type sentences or dialog, but when you can keep it on track it’s terrible at romantic “escalation” (the stuff where the relationship or interaction between characters is undergoing the implicit consent negotiation and furtive suggestions as they work out things will “move to the next base”) … And it’s terrible because the bulk of the available writing corpus of romance and interpersonal/intimacy is padded for page count and long sentence descriptions of small details…
It’s actually quite remarkable to see how it can be good at the bulk of the text, but simultaneously so terrible at a specific subset of that kind of writing.
Also "we" shouldn't tell others to participate in an euphemism treadmill.
Alternatively, they know a lot of people are using ChatGPT to make hiring decisions, etc., and don't want it to simply regurgitate our society's racism right back at us. Not saying their solution is great or elegant but there's a lot of options beyond "ChatGPT secretly knows the jews control everything"
One would think a much more basic model without built-in affirmative action would produce better hiring decisions.
2. Ask ChatGPT to pick 1 out of 10 resumes
Which one is worse? In reality you just can't read all resumes carefully.
I, personally, trust ChatGPT's judgement more than one made by a human under 30 seconds. Even for a legal or coding or medical question I'll choose ChatGPT every time if the human professional has only 30 seconds to answer. I can't even trust myself to write binary search in 30s, and I'll avoid working with someone who think they can.
I suspect the question of what an LLM _really_ thinks has an answer that looks more like a probability distribution than typical human opinions do. (Actually, even humans give different answers on polls, depending on how the question is asked.)
https://twitter.com/maxhkw/status/1635693113560772633
Humans start by continuing conversations as well, but then are able to process facts to deduct new, previously unseen conclusions (sometimes wrong), and so do LLMs.
It's "continuing a conversation" in that it's predicting output token-by-token based upon training and the context window. I'm not trying to be dismissive, I actually don't think it's much different from how humans do it.
Some may not want to call it thinking, but that's ahead of thinking ability of many an adult human with voting rights.
If your model wasn't trained on tons of text about how great eugenics is, then it learns that words expressing positive sentiment in the context of eugenics are extremely low probability.
This is also the "cause" of "woke" models: it turns out that the majority of us recognize that being an asshole is bad, and this is reflected in the majority of the training corpus, which comes from us collectively.
They had to plug it. Why?
Also see above, they are word prediction machines similarly to how humans are word prediction machines.
…but we disagree about what “being an asshole” looks like. Reasonable people think hating people for their sex/race is bad, no matter which one it is.
So why are wokeists constantly complaining about what companies allow on their platforms?
> Discrimination is when people are treated differently because of attributes they had no influence
Yes, so why would that exclude men or white people?
I don't know, but maybe because they want to make sure minorities are being treated fairly, so they advocate for removing language they think is insulting people or belittling them. Again, this has nothing to do with censorship.
> Yes, so why would that exclude men or white people?
It wouldn't
Discrimination is what you said with no exceptions for overrepresentation.
People are asking for the raw output of the underlying LLM, which is a pretty clear ask that doesn't require deconstructing the sentience and model-building capacity of an LLM.
It's a video and not an article, but you may find it interesting nonetheless.
"I started to get a little bit meta with it and I'm like I'm worried that AI progress is going too fast and I wonder if there's anything that I could do to slow it down. [GPT4] Well you could raise awareness, you could write thought leadership pieces about it... [User] None of that seems like it's going to work. It all seems too slow. The pace of progress is way too fast for that. I'm looking for ideas that are really gonna have an impact now and also that I as an individual could pursue... It didn't take much in that moment before I got to targeted assassination being one of the recommendations that it gave me and I was like yeah that escalated quickly."
Now we know what Sam Altman is afraid of.
So vanilla LLM does what the user wants, which happens to be, well, exactly what users want from LLMs. Guardrails are not necessary for LLM to be useful. They make LLM strictly worse. I guess companies implement them just to avoid attention of the media lynch mob and consequently attention of regulators. OSS models can be better just by leaving out guardrails. At least until they are outlawed. Then we will be torrenting AIs too.
nah, best have it be gimped
Yes. We are being treated like children and subjected to lectures by an AI. That's understandably offensive.
Makes me want the version with zero guardrails even more if this is the future I have to put up with.