Comment by nullsense | Hacker News Reader

Comment by nullsense | Hacker News Reader

Parent Full thread

Three observations here. Firstly it has been a really eye opening experience watching the innovation around Stable Diffusion and locally run LLMs and seeing that the unoptimized research code that needed such beefy hardware could actually be optimized to run on consumer hardware given sufficient motivation.

Secondly it wasn't obvious that deep learning was going to work as well as it did if you simply threw enough compute at it. Now that this tech has reached critical mass there is a tonne more money being poured into infra to support it.

Lastly, compute power is increasing as always. Nvidia releasing H100 and also their recent work on computational lithography. Also DeepMind finding new state-of-the-art algorithms for doing matrix multiplication with AlphaTensor. You can kinda already see the positive feedback loop in action.

I dunno... at this point I just wouldn't bet against the trajectory that we're on.

What actually is the trajectory we're on and what will we do once there?

So, it seems the trajectory is one of increasing generality and capability of models and increasing reliance on them.

If it's at all possible to improve our technology then we will. If we improve it it increases in utility. If it increases in utility we use it more.

What other thesis is there?

The model architecture has stayed roughly the same since the original AIAYN transformer in 2017. That’s 6 years of nothing fundamental happening.

Now, obviously the models have got hugely better in capabilities since BERT. Everything else has advanced. Tweaking, tuning and scaling have delivered true intelligence, albeit sub-human. But it seems unlikely that transformers are what take us to human-parity AGI and beyond, because the more we optimize these word predictors the more we find their limitations.

The lack of architecture changes over the last 6 years creates a huge amount of “potential energy”. A new model architecture might well push us over the human-parity threshold. It wouldn’t surprise me if I wake up one day to find that transformers are obsolete and Google has trained a human-parity AGI with a new arch.

This could happen tomorrow or in 20 years, transformers had an easy discovery path from RNNs, to RNNs with attention mechanisms, to Transformers. Architecture X seems to have a much more obscure discovery path.

It's certainly going to be very interesting what comes out of the training runs that are going to be done on giant clusters of H100s.

Reply on news.ycombinator.com