Comment by jjk166
6 days ago
People have been saying, without any evidence at all, it's reached or about to reach a plateau for years now. We are clearly still seeing significant forward progress. While it's reasonable to think it will hit some plateau eventually, there's no reason to think that right now just happens to be as good as it's ever going to get.
Context is the plateau. It's why RAM prices are spiking. We're essentially throwing heap at the problem hoping it will improve. That's not engineering. It's not improving on a fundamental, technical level.
> Context is the plateau. It's why RAM prices are spiking.
Yes, context is the plateau. But I don't think it the bottleneck is RAM. The mechanism described in "Attention is all you need" is O(N^2) where N is the size of the context window. I can "feel" this in everyday usage. As the context window size grows, the model responses slow down, a lot. That's due to compute being serialised because there aren't enough resources to do it in parallel. The resources are more likely compute and memory bandwidth than RAM.
If there is a breakthrough, I suspect it will be models turning the O(N^2) into O(N * ln(N)), which is generally how we speed things up in computer science. That in turn implies abstracting the knowledge in the context window into a hierarchical tree, so the attention mechanism only has to look across a single level in the tree. That in turn requires it to learn and memorise all these abstract concepts.
When models are trained the learn abstract concepts which they near effortlessly retrieve, but don't do that same type of learning when in use. I presume that's because it requires a huge amount of compute, repetition, and time. If only they could do what I do - go to sleep for 8 hours a day, and dream about the same events using local compute, and learn them. :D Maybe, one day, that will happen, but not any time soon.
If a bridge girder isn't strong enough to support a load, you add material in the right places to make a larger, stronger girder. That is engineering. The idea that if you're not making fundamental improvements to your formulation of steel you aren't progressing is absurd. If adding RAM leads to improvements, and we have the engineering ability to add more RAM, then we are still making progress.
Regardless of how true your statement is (just adding metal to a structure is commonly not a way to solve the problem you stated, it just makes the structure heavier which means other systems have more to support) the point is that it isnt exponential/fundamental progress, which is the type that would be needed to avoid the plateaus folks are mentioning. Also adding RAM doesnt give you even linear improvements, its logarithmic.
3 replies →