zlacker

To be clear, he thinks that LLMs are probably a general architecture, and thus capable of reaching AGI in principle with enormous amounts of compute, data, and work. He thinks for cost and economics reasons it's much more feasible to build or train other parts and have them work together, because that's much cheaper in terms of compute. As an example, with a big enough model, enough work, and the right mix of data you could probably have an LMM interpret speech just as well as Whisper can. But how much work does it take to make that happen without losing other capabilities? How efficient is the resulting huge model? Is the end result better than having the text/intelligence segment separate from the speech and hearing segment? The answer could be yes, depending, but it could also be no. Basically his beliefs are that it's complicated and it's not really a "Can X architecture do this" question but a "How cheap is this architecture to accomplish this task" question.

>>lucubr+(OP)
This is wholly besides the point. The person I'm replying to is clearly saying the only people who believe "GPT is on the path to AGI" are non technical people who don't "truly understand". Blatantly false.

It's like an appeal to authority against an authority that isn't even saying what you're appealing for.