[mrdoops]

I wonder to what extent the data being fed to these models are the issue. Or rather the problem is the systems that generate these data-sets and how representative of reality they are. If we make an app that involves humans and that data is used in a model - to what extent does user experience and other factors warp reality?

Maybe our existing methods are good enough given enough compute to reach AGI but our datasets are too low fidelity and non-representative of the problem space to reach desired results?

[MAXPOOL]

The problem is not the data. The problem is the need for high quality data. Current ML is data driven statistical learning. ML tries to learn a model that describes the distribution. It's impossible to get similar performance as the best reference implementation (human brain) using this approach. https://i.redd.it/kvvgv6zzhtp11.png

Think of 16 year old human:

* it has received less than 400 million wakeful seconds of data + 100 millions seconds of sleep,

* it has made only few million high level cognitive decisions where feedback is important and delay is tens of second or several minutes (say few thousand per day). From just few million samples it has learned to behave in the society like a human and do human things.

* Assuming 50 ms learning rate at average, at the lowest level there is at most 10 billion iterations per neuron (Short-term synaptic plasticity acts on a timescale of tens of milliseconds to a few minutes.)

Humans generate very detailed model of their environment with very little data and even less feedback. They can learn complex concept from one example. For example you need only one example of pickpocket to understand the whole concept.

[machiaweliczny]

This pickpocket example seems like symbolic? relations reasoning.

I think we need simulation of other agents outputs as primary tool for reasoning. That seems to be how intelligence emerged in evolution.

Something like this: choose desired action > simulate other agents outputs based on future state after performing action > check reward for this action after simulating outputs of others > perform action or not > update all agents models and relations in "world" graph model

I think world could be modeled as simple graph and each agent as NN.

Then based on graph we could conduct symbolic reasoning and very fast learning (by updating edges)

I think these models need also need good physical simulator and good understanding of competitivness.

Is anyone aware of such trials of building AGI as I described?

Humans have natural language as big competetive adventage (easy way to compress parts of world graph and pass it to others - ambiguous. I think with aftificial machiness can be done more efficient). Another advantage is knowledge storage - also easy to do with machiness.

If we can build insect AI building human AI should be easy.

[machiaweliczny]

I also think it would be great if we could just do "world_model = fold(books)" instead of simulation.

Is anyone aware of such efforts/results?