I haven’t read the whole paper, but my instinct after reading summaries and the first 12 or so pages is that this will be very difficult to implement with today’s (2025) hardware. So much is going on that is computationally complex. LLMs are all the rage because they are incredibly simple and incredibly scaleable. My hunch is that this will only be practical with a completely different computer paradigm such as neuromorphic computing and their highly efficient spiking neural networks. If we want AGI, money needs to be poured into research, not bigger and bigger LLMs and data centers. The payoff would be long term though, not short term, so I don’t think capital will be directed in a way conducive to this undertaking.
2025
“What’s cool is making them run longer … before giving their answers”
Right on time!