Hacker News Reader

Comment by mordymoop | Hacker News Reader

The problem here seems to be that you think the state of the art resembles “being really good at identifying objects”. This makes it clear that you are not keeping up with the frontier. I recommend looking up DeepMind’s 2019 papers, they are easily discoverable.

When you read them, you will probably update in the direction of “AGI soon”. It’s possible that you won’t see what the big deal is, I suppose. I personally see what Carmack and others see, a feasible path to generality, and even some specific promising precursors to generality.

It also helps to be familiar with the most current cognitive neuroscience papers, but that’s asking a lot.

View on HN

You're going to have to be more specific about what constitutes a major advance forwards. So far DeepMind's work (while impressive) has proven to be very brittle, and not transferable without extensive "fine-tuning". Previous attempts at transfer learning have been mixed to say the least.

I'm going to be pessimistic and say that AGI is probably decades away (if not centuries away for a human-like AGI). There are clearly many biological aspects of the brain that we do not understand today, and likely will not be able to replicate without far more advanced medical imaging techniques.

What are some of the highlights from DeepMind that gives you optimism for a path to AGI? I am not seeing it, personally.

I would like to know as well

Is there anything in the structure of the brain that makes you think "of course this is an AGI"? For me, the answer is no. That's why I think progress on narrow AI and AGI is going to be unpredictable. Nobody will see the arrival of an AGI until it's here.

Some also think that nobody will see the arrival of an AGI even after it’s here, because after arrival there will be no one left to see.

Various meta-learning approaches and advancements in unsupervised learning and one-shot learning.

Can you explain the general path in layman's terms in a few sentences? As far as I can tell AI is really good at analyzing large datasets and recognizing patterns. It can then implement processes based on what it's learned from those patterns. It all seems to be very specific and human directed.

Link to DeepMind's papers: https://deepmind.com/research?filters=%7B%22collection%22:%5...

Reply on news.ycombinator.com