The Cost of Computing for General Intelligence

The equivalent with DBT is much less to just run it to produce a few paragraphs, then to train it on all the data that made it whatever. The difference between inference and training is not something I've thought about a lot before when Chinese bio anchors like this for working out of timelines. But yeah, it sounds like you're very much on the side of okay, these aren't particularly comparable scale wise. There's an interesting thought here about how many experiments to take for us to find an architecture which is good enough to be general.