Hacker News Reader

Comment by danaris | Hacker News Reader

The throughline from what software did 20 years ago to what it does today is pretty clear.

There's absolutely no throughline from the current ML projects to an AGI. The kinds of things they're doing don't even resemble conscious or abstract thought. They are probabilistic engines trained to mimic one thing at a time, only responding to specific input prompts, with moderate success.

Just because you can say that the computational structures being used resemble those in a human brain doesn't mean it's capable of the same things.

This "AGI is unfalsifiable" "either you believe or you don't" stuff is religious zealotry, just with a different focus.

View on HN

Nah. I don’t think you understand consciousness. You claim there’s no connection but you couldn’t describe how consciousness works let alone how the models work. Don’t put words in my mouth, I never talked about the architecture.

There is no throughline from software 20 years ago because software was written deliberately then. All the experts said we wouldn’t be anywhere near where we are today by today because of this. If you can’t see the paradigm has shifted then you are simply wrong.

Proof should be required to continue forward. We shouldn’t need proof to do what’s safer. Especially when the consequences are so extreme.

I understand consciousness well enough to name two necessary-but-insufficient prerequisites for it, which I call continuity of input and continuity of learning.

The former means that it has to be always-on, with some form of continuous input stream, like our senses, as distinct from the current ML systems that sit there idle, waiting for discrete input and giving discrete output.

The latter means it also needs to be, effectively, continuously using that input to re-train itself, as distinct from the current systems that are trained once or multiple times in discrete batches, then used over and over again with that static model.

Reply on news.ycombinator.com