Today’s most advanced AI models have many flaws, but decades from now, they will be recognized as the first true examples of artificial general intelligence.
There’s a lot of papers which propose adding new tokens to elicit some behavior or another, though I haven’t seen them catch on for some reason. A new token would mean adding a new trainable static vector which would initially be something nonsensical, and you would want to retrain it on a comparably sized corpus. This is a bit speculative, but I think the introduction of a token totally orthogonal to the original (something like eg smell, which has no textual analog) would require compressing some of the dimensions to make room for that subspace, otherwise it would have a form of synesthesia, relating that token to the original neighboring subspaces. If it was just a new token still within the original domain though, you could get a good enough initial approximation by a linear combination of existing token embeddings - eg a monkey with a hat emoji comes out, you add tokens for monkey emoji + hat emoji, then finetune it.
Most extreme option, you could increase the embedding dimensionality so the original subspaces are unaffected and the new tokens can take up those new dimensions. This is extreme because it means resizing every matrix in the model, which even for smaller models would be many thousands of parameters, and the performance would tank until it got a lot more retraining.
(deleted original because I got token embeddings and the embedding dimensions mixed up, essentially assuming a new token would use the “extreme option”).
There’s a lot of papers which propose adding new tokens to elicit some behavior or another, though I haven’t seen them catch on for some reason. A new token would mean adding a new trainable static vector which would initially be something nonsensical, and you would want to retrain it on a comparably sized corpus. This is a bit speculative, but I think the introduction of a token totally orthogonal to the original (something like eg smell, which has no textual analog) would require compressing some of the dimensions to make room for that subspace, otherwise it would have a form of synesthesia, relating that token to the original neighboring subspaces. If it was just a new token still within the original domain though, you could get a good enough initial approximation by a linear combination of existing token embeddings - eg a monkey with a hat emoji comes out, you add tokens for monkey emoji + hat emoji, then finetune it.
Most extreme option, you could increase the embedding dimensionality so the original subspaces are unaffected and the new tokens can take up those new dimensions. This is extreme because it means resizing every matrix in the model, which even for smaller models would be many thousands of parameters, and the performance would tank until it got a lot more retraining.
(deleted original because I got token embeddings and the embedding dimensions mixed up, essentially assuming a new token would use the “extreme option”).