Unless I'm wrong, an AI can't just talk to itself during file compression evaluation to get extra training data, to therefore compress better. Because every next letter it predicts would be already known with probability for any context that predicts it. Ex. you know the memory 'walking', you see 'walking' and predict 's', you store now 'walkings', not found in the dataset but from your brain. Problem here is, great, but next time you see 'walking', you predict 's', whether stored walkings or not. If you see 'walkings', (and you know 'walkingsz' after talking to self further ahead, as you already know kingsz prior), you predict 'z' only because you already know kingsz, ingsz, ngsz, gsz, sz, and z. So it'd be same if not had stored walkingsz.
Only if you do semantic discovery does your prediction get better without new data. Then, you can generate new entailment predictions, ex. you didn't know walking>fast, you see that the ing is recognizable and predict more randomly all that can come next, you don't know its likely that 'fast' comes after walking though, until you discover jogging and walking share predictions, now you bring predictions over to walk and know now 'walking>f' (for fast). So now, after seeing the two words share predictions, it's important you store over at the neuron walking this now > walking>fast. And so to do that you talk to yourself. Done. Made the sequence. Collected new thought/ insights/ DATA, for free. Data you love and data where you knew you are unsure of what letter to predict next. Because as said above, unless I'm wrong, storing self talkings is useless and doesn't improve prediction. You have to first take memories, get more out of them, then you can make new sequences. Another way to do it is sparse matching and delay match models. So you see 'walking', you know 'w_l_i_gs', so you predict walking>s, now you store 'walkings', and that's it, you talked to yourself. One problem with the aboves. You store the new predictions from translation and holed and etc matching, so they don't need to be computed again, though you stay between a trade off of mem and speed. But only when get new data. If you talk to yourself in brain, there is no need to do it, because if you try to get new predictions for 'walking' using a sparse memory, it would have already when saw walking, or the sparse memory like it, whichever came last you would have saw a similar match, already, when it came in. So we see the need to store sequences and use some the memory up, but why do we seem to do it in brain with no novel real world stimuli? Because we are still recalling real world data, we are still processing it, we don't ZIPADO done it in 10 seconds, we ramble on often, at least rarer pattern finder mechanisms do, that are not hardcoded but ran using memories. So, we are still processing the next letter, for all sorts of areas of the paragraph read recently or one that we love permanently ringing in our brain. In conclusion, our many rarer softcoded mechanisms take longer to improve prediction, and are actually still reading the last seen or loved contexts! The us talking to ourselves is the act of the softcoded mechanisms taking long and searching ****and running using sequential memories****, hence seem conscious. It would be great if we could hardcode these in AGI on the fly, the common ones at least. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tdd35fb7acb50d151-M4cc6229bc039da43ea2f233f Delivery options: https://agi.topicbox.com/groups/agi/subscription
