Can we train an LLM based on brainwaves rather than written text? Seems to be closer to how we actually think and thus should enable the LLM to learn to think rather than just learn to mimic the output.
For example, when writing we have often gone done many thought paths, evaluated each and backtracked etc, but none of that is left in the text an LLM trains on today. Recording brainwaves and training on that is probably the best training data we could get for LLMs.
Getting that data wouldn't be much harder than paying humans to solve problems with these hats on recording their brainwaves.
On the other hand, the main practical feature of a language is its astronomical SNR, which brain waves lack, to say the least. This allows LLMs to be trained on texts instead of millions of live people. Just imagine the number of parameters and compute resources required for the model to be useful to more than one human.
For example, when writing we have often gone done many thought paths, evaluated each and backtracked etc, but none of that is left in the text an LLM trains on today. Recording brainwaves and training on that is probably the best training data we could get for LLMs.
Getting that data wouldn't be much harder than paying humans to solve problems with these hats on recording their brainwaves.