Something very weird would have to happen for it to be that hard.
Humans have been optimizing for information transmission between individuals for a long time. Not just our hardware and neural nets have been optimized, but the structure of the information itself.
It is quite possible that you need such a large set of compex interactions to develop intelligence that intelligence can only develop in reality. It is possible that the process for doing this inherently inolves a chaotic system that makes duplicating or paralleling that development impossible. I don't see any reason why these artificial intelligences with divergent chaotically emergent capabilities would have to be able to teach each other things as easily as we do.
I've long been of the opinion culture / "social software" is vital for producing knowledge for humans and that this social software has been fine tuned and iterated upon for a very long time (and possibly co-developed with the evolution of our brain hardware so that both are co-specialized for each other). This software is likely very specialized to our brain structures and might not generalize well to other intelligence substrates. Thus even if we do develop neural network structres that can run on silicone and match human level intelligence, those entities would still need a lot of interaction with reality to develop their own culture that would allow them understand the world at the level we do. Edit: Is it even possible to efficiently run the necessary theories of human mind on non-human brains to operate effectively in our epistemic landscape? Communication requires huge amounts of inference using theories of mind to resolve ambiguity and add implicit information.
It is possible that we can "recompile" human knowledge to run on other types of intelligences, but I suspect this will be far more difficult that commonly expected. Just look at how hard (nigh impossible) it is today to even boostrap a new general purpose computer without depending on the output other general purpose computers.
We simply do not understand cognition or semantics well enough to know the answers to these questions adequately or with any degree of certainty.
Even if it does take an enormous investment and many years to train a computer to do a task, you can make as many copies as you want and everyone can have one. Educating people doesn’t have that kind of payoff.
That’s why you see billions spent in the effort to make a self-driving car. But it turns out that real-world experience isn’t enough, so far. Waymo has collected > 20 million miles of driving experience.
It’s not enough because machine learning is very inefficient at learning from experience. That could improve, though.
The basic idea is that, the smarter you are and the faster you can think, the more information you can extract from weak signals. It involves creativity, and building models, and some minimum amount of in-built bias, but the upshot is that the amount of information a super-intelligent AI could infer from very small amounts of experience might absolutely stagger the imagination.
Do you think everything taught in an oral form (as opposed to hands on) is completely useless?
Are all military strategy books used by military academies a waste of time, since you can only learn by leading a real army?
You can't understand those books or information without having a lifetime of experience to contextualize them. You can't convert them from rote knowledge into skills without trying them out.
In AlphaZero for example, there were 44 million training games total for 700,000 steps of training for the full 9 hours.
Turning that to human-scale numbers, 44million games with on average 60 moves, at 1 second thinking time per move,
> 44000000*60/60/60/24/365 = 83,7138508371 years of training experience in 9 hours
The whole field of Reinforcement learning has agents training and playing games for many orders of magnitude more time than a human ever will. In-fact, we can scale this to over 100k of actions per second, in a single machine:
https://github.com/alex-petrenko/sample-factory
Then, there is also distributed Reinforcement Learning, where hundreds of agents can play at different machines and share experience, see AlphaZero, LeelaZero, R2D2 agent, R2D3 agent, Apex, Acer, Asynchronous PPO.
> but the data isn't useful without the context of experience
The experience is the data in Reinforcement Learning.
> and all processing power can do it overfit model without experience.
That is wrong, the agents perform what is called exploration to avoid getting stuck in simple strategies.
> Even if we put AI into an army of robots running around and experiencing things, there are still scaling limits to encoding and communicating knowledge and understanding.
True, but machines scale better because they speak the same language, or they can learn to tune their language to get their message across.
> Human organizations are a great example of the scaling limits of intelligence.
Human organization is a testament to how far we can get with something as limiting as the commonly used language. The language that we use to communicate is subject to misinterpretation due to our subjective experiences, this limitation is not shared by machines.
> The experience is the data in Reinforcement Learning.
This is very true, and the critical problem. Data about how reality responds to an AI's actions is very sparse right now.
AIs do have a potential advantage in communications efficiency, but at some level of scale compression will happen, locally "irrelevant" data will be discarded and simplified approximations replace it. None of this will change the "big O" of the scalability of intelligence, just the constant factors. There is no exponential kickoff point.
The difference I can see is that there is no one explicit objective function, but this doesn't stop generally capable agents [1], and doesn't imply that inverse RL is not possible.
> The ability to play go very well doesn't enable an AGI to better understand reality.
I disagree, model based RL constructs a model of the agent's reality and can use it to plan ahead, train the agent, or do some form of monte-carlo tree search. The latter is something very similar to how we imagine and think about the future.
[1] https://deepmind.com/blog/article/generally-capable-agents-e...
Finding out the consequences of an action is a lot more expensive in reality than a simulation of a game.
Even if we put AI into an army of robots running around and experiencing things, there are still scaling limits to encoding and communicating knowledge and understanding. Human organizations are a great example of the scaling limits of intelligence.