I assume you heard of chatGPT, maybe played with it a little, and was imressed by it (or tried very hard not to be).
This framing reminds me of something I’ve suspected for a little while; I think that something about OpenAI’s products is memetic, and that playing with them for extended periods of time creates a memeplex which glorifies the product somehow. Playing with open-source LLMs doesn’t seem to evoke the same attitude, which makes me suspect either RLHF or the final layer of adaptive filtering and watermarking.
In short, to play [the guess-the-next-letter game] well, you need to understand the text, understand the situation described in the text, imagine yourself in the situation, and then to respond.
This is an unjustified claim, and without it, the entire idea of “AI-complete” collapses. There is no evidence whatsoever that humans understand text, and it seems like the typical human response might skip any sort of understanding, imagining, modelling, or other thought. (Previously, on Lobsters, we examined whether humans cogitate when they write.)
I see that I previously did not leave an example. Here’s a common one in the USA: somebody says “sixty-nine” and somebody else says “nice”. The second speaker does not necessarily understand the situation; they may be acting memetically, not emotionally or rationally.
I remember the sex jokes all kids threw around in primary school. No one actually knew what this really was about, apart from the fact grown ups did it and something about reproduction. But everyone threw them around (like swear words), because that gave a reaction from the surrounding people. So it was “fun”.
This is a good article by Yoav Goldberg (an NLP researcher) on LLMs, relating back to some early writing of Shannon and a bit of discussion on the universality of next token prediction. Then it goes into a bit of detail on the limits of LLMs.
This framing reminds me of something I’ve suspected for a little while; I think that something about OpenAI’s products is memetic, and that playing with them for extended periods of time creates a memeplex which glorifies the product somehow. Playing with open-source LLMs doesn’t seem to evoke the same attitude, which makes me suspect either RLHF or the final layer of adaptive filtering and watermarking.
This is an unjustified claim, and without it, the entire idea of “AI-complete” collapses. There is no evidence whatsoever that humans understand text, and it seems like the typical human response might skip any sort of understanding, imagining, modelling, or other thought. (Previously, on Lobsters, we examined whether humans cogitate when they write.)
I see that I previously did not leave an example. Here’s a common one in the USA: somebody says “sixty-nine” and somebody else says “nice”. The second speaker does not necessarily understand the situation; they may be acting memetically, not emotionally or rationally.
This is greatly evidenced by small children who are aware that the number 69 is funny or “nice” but have absolutely no context on why.
I remember the sex jokes all kids threw around in primary school. No one actually knew what this really was about, apart from the fact grown ups did it and something about reproduction. But everyone threw them around (like swear words), because that gave a reaction from the surrounding people. So it was “fun”.
This is a good article by Yoav Goldberg (an NLP researcher) on LLMs, relating back to some early writing of Shannon and a bit of discussion on the universality of next token prediction. Then it goes into a bit of detail on the limits of LLMs.