I get that Grok has more credibility than Elmo at this point, but stuff that a chatbot says is no more newsworthy than stuff said by a parrot.
Or to put it another way, LLM’s are advanced chatbots. Their purpose is to generate credible sounding text, not accurate text.
But, like a human, it mostly tries to stick to the truth. It does get things wrong, and in that way is more like a 5 year old, because it won’t understand that it is fabricating things, but there is a moral code that they are programmed with, and they do mostly stick to it.
To write off an LLM as a glorified chatbot is disingenuous. They are capable of produce everything that a human is capable of, but in a different ratio. Instead of learning everything slowly over time and forming opinions based on experience, they are given all of the knowledge of humankind and told to sort it out themselves. Like a 5 year old with an encyclopedia set, they are gonna make some mistakes.
Our problem is that we haven’t found the right ratios for them. We aren’t specializing the LLMs enough to make sure they have a limited enough library to pull from. If we made the datasets smaller and didn’t force them into “chatbot” roles where they are given carte Blanche to say whatever they say, LLMs would be in a much better state than they currently are.
I wouldn’t say that precipitating a statistically average response from a primordial soup of training data is really following a moral code or “trying to stick to the truth”.
Programmers and researchers can try as much as they want to get LLMs to behave as expected, but they’re black boxes by nature.
If Elon had a parrot that constantly said “Elon is a Nazi”, it would be in the news.
Something tells me that if AI took over the world, we’d actually be okay.
Kinda like how self driving cars are still safer than the average driver, ya. Do they make mistakes? For sure, although the bigger annoyance is just how slow they are to turn sometimes. AI would be so so at leading but man is the bar low with Americans.
