Transformers have exactly two strengths. None of them is "attention". Attention could be replaced with any arbitrary division of the network and it would learn just as well.
First true strength is obvious, it's that they are parallelisable. This is a side effect of people fixating on attention. If they came up with any other structure that results in the same level of parallelisability it would be just as good.
Second strong side is more elusive to many people. It's the context window. Because the network is not ran just once but once for every word it doesn't have to solve a problem in one step. It can iterate while writing down intermediate variables and accessing them. The dumb thing so far was that it was required to produce the answer starting with the first token it was allowed to write down. So to actually write down the information it needs on the next iteration it had to disguise it as a part of the answer. So naturally the next step is to allow it to just write down whatever it pleases and iterate freely until it's ready to start giving us the answer.
It's still seriously suboptimal that what it is allowed to write down has to be translated to tokens and back but I see how this might make things easier for humans for training and explainability. But you can rest assured that at some point this "chain of thought" will become just chain of full output states of the network, not necessarily corresponding to any tokens.
So congrats to researchers that they found out that their billion dollar Turing machine benefits from having a tape it can use for more than just printing out the output.
PS
There's another advantage of transformers but I can't tell how important it is. It's the "shortcuts" from earlier layers to way deeper ones bypassing the ones along the way. Obviously network would be more capable if every neuron was connected with every neuron in every preceding layer but we don't have hardware for that so some sprinkled "shortcuts" might be a reasonable compromise that might make network less crippled than MLP.
Given all that I'm not surprised at all with the direction openai took and the gains it achieved.
First true strength is obvious, it's that they are parallelisable. This is a side effect of people fixating on attention. If they came up with any other structure that results in the same level of parallelisability it would be just as good.
Second strong side is more elusive to many people. It's the context window. Because the network is not ran just once but once for every word it doesn't have to solve a problem in one step. It can iterate while writing down intermediate variables and accessing them. The dumb thing so far was that it was required to produce the answer starting with the first token it was allowed to write down. So to actually write down the information it needs on the next iteration it had to disguise it as a part of the answer. So naturally the next step is to allow it to just write down whatever it pleases and iterate freely until it's ready to start giving us the answer.
It's still seriously suboptimal that what it is allowed to write down has to be translated to tokens and back but I see how this might make things easier for humans for training and explainability. But you can rest assured that at some point this "chain of thought" will become just chain of full output states of the network, not necessarily corresponding to any tokens.
So congrats to researchers that they found out that their billion dollar Turing machine benefits from having a tape it can use for more than just printing out the output.
PS
There's another advantage of transformers but I can't tell how important it is. It's the "shortcuts" from earlier layers to way deeper ones bypassing the ones along the way. Obviously network would be more capable if every neuron was connected with every neuron in every preceding layer but we don't have hardware for that so some sprinkled "shortcuts" might be a reasonable compromise that might make network less crippled than MLP.
Given all that I'm not surprised at all with the direction openai took and the gains it achieved.