"10% of the way towards AGI over ~5-10 years" (depending on how you measure it) would be a terrifying statement even if the rate of progress weren't also exponentially increasing.
Right now, all thats being done are LLMS being RL trained on specific tasks, which are usually drawn from human curated data sources.
This is why you can ask an LLM questions and it can write code, but you can't just leave it running with a complex task that would involve research and expect it to solve the problem.
Its obvious that part of AGI will be the mapping of language to latent space, and then transformations on either responses or actions. But the issue is that until the model can self learn to become better at coding without going through manual RL tasks designed by humans, you will never get AGI.
And if a model can do that, that means that its possible to have basically the "genesis" code - i.e something that can be run on a computer, with no training on english language, but aware of existance of things it can do, can basically figure out how to train itself to the standard of a current LLM and beyond.
And in figuring that part out, we don't even know where to begin. It very well may be a case of running LLMS in a genetic algorithm and simulating evolution to figure out what LLM+Agent harness works the best.