4 ms·

I'm unconvinced. You've described that it transforms an input to an output, but you've ignored the possibility that the hidden layers in the neural network have

I'm unconvinced. You've described that it transforms an input to an output, but you've ignored the possibility that the hidden layers in the neural network have discovered that generalizations are the best approach to predicting in entirely novel situations. Long before GPT came about, machine learning researchers have discovered (in research examining the purpose each neuron serves) emergent generalizations in image classifier models. To use the example in the article, the answer to "Here we have a book, nine eggs, a laptop, a bottle and a nail. Please tell me how to stack them onto each other in a stable manner" is not something the original 'lookup table' could possibly help you with, other than by manually reasoning about how the various parts of a 'concept web' fit together. This question represents a completely unfilled gap in the training data.

ActorNightly 3y ago
>is not something the original 'lookup table' could possibly help you with, Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic. Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same. Just like transformers were revolutionary because they essentially are representative of a process inside human brains where we learn to pay attention to select features, there will be some new architecture in the next 10 years that will encode this ability to assign semantic properties to information and operate on those properties, which will then allow for things like AI to ask clarifying questions.
- creatonez 3y ago
  > Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic. This is very reductionist. Of course everything is numbers under the hood. But it's entirely orthogonal to the question. Numbers can encode all of the complexity needed for anything that can do reasoning. Numbers in the form of a neural network can execute arbitrary algorithms, given enough layers. > Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same. I really don't understand the purpose of this thought experiment. A human is able to figure this out because they can see what those objects are. Is your objection to language model reasoning that they are not multi-modal? > Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic.
  - ActorNightly 3y ago
    > A human is able to figure this out because they can see what those objects are. What if you couldn't see? What if the interaction was limited to just text? Would you be smart enough to ask for an image of those objects? Do you think an LLM can "reason" to ask for additional information, like images of the objects? A sufficiently large LLM can be trained with a wide range of training data, to the point where it can answer some questions where its seemingly generating new information, but that new information is simply a combination of existing information. In the same way, when you ask it some question about code, it may put together two snippets of code never seen before into something that is new. The key factor there is all of that information is encoded into the network, so all the network is doing is essentially a somewhat complicated graph lookup.

ActorNightly 3y ago

>is not something the original 'lookup table' could possibly help you with, Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic. Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same. Just like transformers were revolutionary because they essentially are representative of a process inside human brains where we learn to pay attention to select features, there will be some new architecture in the next 10 years that will encode this ability to assign semantic properties to information and operate on those properties, which will then allow for things like AI to ask clarifying questions.

creatonez 3y ago
> Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic. This is very reductionist. Of course everything is numbers under the hood. But it's entirely orthogonal to the question. Numbers can encode all of the complexity needed for anything that can do reasoning. Numbers in the form of a neural network can execute arbitrary algorithms, given enough layers. > Lets say I ask you the following question "Here we have a flibert, nine moltice, a grook, a seerik and a lopo. Please tell me how to stack them onto each other in a stable manner". You would be able to figure out how to stack them, given the fact that you have the ability to "reason" and figure out what those items are in the first place. In the same way, an AI that can reason would be able to do the same. I really don't understand the purpose of this thought experiment. A human is able to figure this out because they can see what those objects are. Is your objection to language model reasoning that they are not multi-modal? > Sure it can. Statistically there is some language semantics associated with that sentence, and given a large enough training data set, there will be a generated answer thats purely statistic.
- ActorNightly 3y ago
  > A human is able to figure this out because they can see what those objects are. What if you couldn't see? What if the interaction was limited to just text? Would you be smart enough to ask for an image of those objects? Do you think an LLM can "reason" to ask for additional information, like images of the objects? A sufficiently large LLM can be trained with a wide range of training data, to the point where it can answer some questions where its seemingly generating new information, but that new information is simply a combination of existing information. In the same way, when you ask it some question about code, it may put together two snippets of code never seen before into something that is new. The key factor there is all of that information is encoded into the network, so all the network is doing is essentially a somewhat complicated graph lookup.