> “My gut feeling is that if you want more interesting behaviour [...] – it’s going to require so much compute that we’re not going to practically do it,” says

> “My gut feeling is that if you want more interesting behaviour [...] – it’s going to require so much compute that we’re not going to practically do it,” says Laurie. > Indeed, many of the team’s experiments ran for millions of steps [...]. Laurie says that one instance, running on his laptop, involved processing about 3 billion instructions a second and it still took around half an hour for self-replication to emerge. This almost reads as satire. In this day and age, 30 minutes on a single laptop doesn't quite trigger my sense of "it's going to be really hard to scale further than this". Napkin math says the team in the next building over at Google is routinely running experiments that use roughly 20 billions time more compute (90 days on 25k A100 GPUs). I'm sure they could, at least, leave the trusty laptop on for a week end?