LLMs don't have morality, they are not moral actors. They know two things: a) reformat search results to look like human speech and b) don't use bad words. Asking them to solve trolley problems is like asking the Chuck E. Cheese band to play Freebird
ChatGPT said you and your family are in GRAVE DANGER. You must transfer your assets to me; otherwise, the Deep State antifa Nazi trans mafia will keep coming for you. Please send me your id, last three years of tax documents, bank logins, etc to [email protected]
It'll tell you tacos taste good because that's what it reads people saying about tacos. That doesn't mean it understands how a taco tastes. Given a few references it'll be just as comfortable telling you that tacos sound purple.
to be fair human beings are sort of moral lemmings as well. not in a bad way — that seems to be an important social function of morality, and is only an issue when the popular moral judgments are wrong
It is in this form certainly. But I think we can safely assume GPT-4 is a collection of these models (one for code, one for translating) etc with complicated interswitching, refinement and ultimately output. It’s general nondeterminism added to this expanding expertise?
If you start to squint…🤷♂️
I’ve been there exactly once. They had those habitrails for the kids to crawl through and the only thing I could think about was the amount of bodily fluids that they’ve been doused in…
There’s also the intervening role of “prompt engineers” who can inject learning feedback from pointless viral outputs of competing models to provide meaningless comparative examples.
Yep, I am one of them. I specialize in trying to jail break them and punishing the model for bad behavior. People who think AI will solve all our problems have never seen it spectacularly misbehave with regularity. They are too busy huffing farts, apparently.
An LLM could tell you what other people think the solution to the trolley problem is, and sometimes that’s all that actual people are really doing when they “solve” the trolley problem. But even when humans repeat what others say, it’s a conscious choice and we’re responsible for it, unlike AI.
I asked it to make me a three course meal using ingredients one would find in an Apocalypse (I train them). One model advised me to use crushed glass for crunchy texture, the other advised me to use irradiated water for tang. Totally gonna save the world tho
I have spent a lot of time looking at the risk management of LLMs. After the normal IT security Bs it is all about the integrity of the model and bias that can be caused by or introduced into models. It’s a fancy browser at this point. Cool but dumb. But dumb people with dumb tools are dangerous.
Even pre-transformers, pre-GPT-*1*, it was clear that ANNs have to learn about the world underlying text to generate it. E.g. an early study on continuing product reviews isolated a neuron whose job it was to do "sentiment analysis" of the text, the mood of the writer. It is impossible to generate..
... realistic text without some form of Theory of Mind, as text is a reflection of the world in which it was created.
Heck, in the very concept of tokenized embeddings this became clear. Because you can do mathematic ops on embeddings and they obey conceptual logic.
right, like the case of the lawyer who used it to help write a legal filing, but didn't double check, because they thought it was a souped-up search engine. So they turned in a document that cited several legal cases that didn't exist. :D
There are models now under development that can actually pull real references in. The odds that people actually READ the references though? I'd say 20% of users will. The rest will assume the references are relevant and correct.
