I agree that they are mainly talking about old ideas; I’m curious about whether you think important progress has been made that isn’t included in Sam’s post. (A short reply with some links would be very helpful!)
Since Jan hasn’t answered himself: I just saw this tweet by Jan, which explains the differences between the original simulators frame and his current model of LLMs in depth. He also provides three relevant links here.
There’s some newish material here — they include and cite Nostalgebraist’s the void, which is recent and from a cyborgism or cyborgism-adjacent author. They also cite some alignment preparing data, like Tice, Radmard et al, which is quite recent. There’s also actually new material: they managed to objectively prove the existence of what I call the Puppeteer in Goodbye, Shoggoth: The Stage, its Animatronics, & the Puppeteer – a New Metaphor in their Empirical Observations section: the “authorial voice” that alignment training puts into the model: the model tends to simulate aligned users coin-flips.
I see this as covering the entire range of the topic from stuff that’s 3-4 years old, to recent, to some original ideas and results. It’s like writing a text on Evolution in Biology — yes, some of it goes back to Darwin, but the ideas and their consequences have been further developed since, and to explain them you need to start with the fundamentals and then build on consequences, elaborations, and remaining open questions.
I agree that they are mainly talking about old ideas; I’m curious about whether you think important progress has been made that isn’t included in Sam’s post. (A short reply with some links would be very helpful!)
Since Jan hasn’t answered himself: I just saw this tweet by Jan, which explains the differences between the original simulators frame and his current model of LLMs in depth. He also provides three relevant links here.
There’s some newish material here — they include and cite Nostalgebraist’s the void, which is recent and from a cyborgism or cyborgism-adjacent author. They also cite some alignment preparing data, like Tice, Radmard et al, which is quite recent. There’s also actually new material: they managed to objectively prove the existence of what I call the Puppeteer in Goodbye, Shoggoth: The Stage, its Animatronics, & the Puppeteer – a New Metaphor in their Empirical Observations section: the “authorial voice” that alignment training puts into the model: the model tends to simulate aligned users coin-flips.
I see this as covering the entire range of the topic from stuff that’s 3-4 years old, to recent, to some original ideas and results. It’s like writing a text on Evolution in Biology — yes, some of it goes back to Darwin, but the ideas and their consequences have been further developed since, and to explain them you need to start with the fundamentals and then build on consequences, elaborations, and remaining open questions.