3 ms·
We already understand how transformers work: their architecture can learn to approximate a very large class of functions-on-sequences (specifically continuous s
by logicchains 3y ago
We already understand how transformers work: their architecture can learn to approximate a very large class of functions-on-sequences (specifically continuous sequence-to-sequence functions with compact support: https://arxiv.org/abs/1912.10077 https://arxiv.org/abs/1912.10077). It can do it more accurately than previous architectures like RNNs because transformers don't "forget" any information from prior items in the sequence. Training transformers to predict the next item in sequences eventually forces them to learn a function that approximates a world model (or at least a model of how the world behaves in the training text/data), and if they're large enough and trained with enough data then this world model is accurate enough for them to be useful for us.
If you're asking for understanding the actual internal world model they develop, it's basically equivalent to trying to understand a human brain's internal world model by analysing its neurons and how they fire.
- theGnuMe 3y ago>Training transformers to predict the next item in sequences eventually forces them to learn a function that approximates a world model There is absolutely no proof for this statement.
- H8crilA 3y agoHave you used ChatGPT? It can do (at least) simple reasoning, for example simple spatial reasoning or simple human behavior reasoning.
- theGnuMe 3y agoI suggest you start here: https://ahtiahde.medium.com/limits-of-turing-machines-and-algorithmic-finger-prints-f33803ecbb43 https://ahtiahde.medium.com/limits-of-turing-machines-and-al... If you don't have a CS background, I would suggest reading the wikipedia entries referenced in the medium article as well.
- seydor 3y agoWe shouldn't assume that either of those tasks is impossible
- zamfi 3y ago> it's basically equivalent to trying to understand a human brain's internal world model by analysing its neurons and how they fire A major challenge here is that it's very hard/expensive to analyze these neurons, especially at any kind of scale within one human. Not so with LLMs.