[LOCKSUIT]

So AGI is all about Prediction.
And Recognition.
Although both are very similar.

Prediction is made more accurate mostly by frequency in big data. It increases probabilities. It tells you what usually happens when you see a prior or what usually came before something that entails it.

For example:
aab
aab
aab
aac
aac
abd
What letter usually follows if we see aa? B.

Prediction is also improved when Recognition is improved. Cat=Dog. So if we see 'my ct bit her momm' it can match 'his dog chewed very very hard on my dad' and get extra entailment frequency to mix in the set of candidate letters that entail.

So the idea is we match, and get candidate probabilities.

What else can we do to improve our 'GPT-2' here? Anyone? What more does GPT-2 do?

Another trick I'll add is energy remains in the net where recently activated nodes are, so that they get extra energy, because they are likely to be said again.

AGI just has to know how to get, what it wants. We could add reward to the prediction to boost certain answers. Careful though! Bliss is ignorance!

[yotamarker]

hmm
how would this apply to
detective work algs for example ? where you use clues to make a suspect list
than end with an exact culprit ?

[HS]

I could never figure out what a want, or reward is for AGI, how it might be made, but once its got a "want", it should learn to use appropriate data for performing the kind of prediction it wants to do. For weather predictions, it should decide to ignore influences which have small effects because there are too many of those. So that probably means down regulating certain predictive mechanisms within itself, when the AGI predicts that they will not be well suited for the upcoming prediction.

[LOCKSUIT]

Yes my code does that, higher frequency has exponentially more weight.


Primitive AI is trying random motor actions until one moves it forward to trigger the accelometer, which tells it to keep these actions it just tried and tweak them in the future to improve.

But this has too little context. A thinker like GPT-2 can plan using long context about many things while considering many tings that did or may happen or are related to things. All you do is plant/ hardwire/ mark nodes in its brain hierarchy and make it ask questions like 'how will i get sex?' or 'how will i get food?', and it can leak these reward chemicals to similar nodes like 'food=money'....artificial reward creation! Games etc do that. RL for thoughts. This is the way forward. It makes it talk about things it loves. The next word to predict is likely frequent and desired. All hobbies stem from sex and food, so get with it folks! Golfing, sailing, painting, home construction, dresses, etc. Sail to bring food, paint to share data, sight see to collect data, etc. The only reward you can't change in life is the food and faces you like - you can't turn gay or make carrots tasty. Not yet!

[yotamarker]

I could never figure out what a want, or reward is for AGI, how it might be made, but once its got a "want", it should learn to use appropriate data for performing the kind of prediction it wants to do. For weather predictions, it should decide to ignore influences which have small effects because there are too many of those. So that probably means down regulating certain predictive mechanisms within itself, when the AGI predicts that they will not be well suited for the upcoming prediction.

just program some drugs theory

[krayvonk]

Even if the bot is a great predictor,   whats it going to do with it?   Could be quite evil things - hard to say from this blurry unexperimentalled position.
How are u supposed to force it to do something without it becoming illogical and hampering it severly, im imagining.

<edit>
one more thing,  the computer "understands" things if u type it into it,  if you have come up with an equation,  this is what computers understand.
If you want to understand english,  you need an equation for english,  then it performs it.

[LOCKSUIT]

So many easy questions...

> Even if the bot is a great predictor,   whats it going to do with it?

Anything that text/images say. As long as it knows things fall, cats run, and food is money, then it can know the maze of the Earth!! Now all it needs is the scent, a reward question "How will I get the cat to leave the bed and poke mom's leg?".

> How are u supposed to force it to do something without it becoming illogical and hampering it severly, im imagining.

Needs excellent recognition and prediction to know how to go from one thing to the next while keeping on topic. And it needs milestone rewards recognized and an end goal to recognize that it got there in the story!

> one more thing,  the computer "understands" things if u type it into it,  if you have come up with an equation,  this is what computers understand.
If you want to understand english,  you need an equation for english,  then it performs it.

GPT-2 works nicely doesn't it?