Nothing is certain, but my $0.02 is that setting LLM-based agents up with long-running tasks and giving them a way of interacting with the world, via computer use (e.g. Anthropic's recent release) and via actual robotic bodies (e.g. figure.ai) are the way forward to AGI. At the very least, this approach allows the gathering of unlimited ground truth data, that can be used to train subsequent models (or even allow for actual "hive mind" online machine learning).