How to build an A.I. brain that can conceive of itself | Joscha Bach

How to build an A.I. brain that can conceive of itself | Joscha Bach


If you look at our current technological systems
they are obviously nowhere near where our minds are. They are very different. And one of the biggest questions for me is:
What’s the difference between where we are now and where we need to be if we want to
build minds—If we want to build systems that are generally intelligent and self-motivated
and maybe self-aware? And, of course, the answer to this is ‘we
don’t know’ because if we knew we’d have already done it. But there are basically several perspectives
on this. One is our minds as general learning systems
that are able to model arbitrary things, including themselves, and if there are this, they probably
need a very distinct set of motivations, needs; things that they want to do. I think that humans get their specifics due
to their particular needs. We have cognitive and social and physiological
needs and they turn us into who we are. Our motivations determine where we put our
attention, what we learn and what we actually do in the world—what we model, how we perceive,
what we are conscious of. In a similar sense, it might be that it’s
sufficient to build a general learning architecture and combine this with a good motivational
system. And we are not there yet in building a general
learning architecture. For instance, our minds can learn and create
new algorithms that can be used to write code and invent code, programming code for instance,
or the rules that you need to build a shop and run that shop if you’re a shopkeeper,
which is some kind of programming task in its own right. We don’t know how to build a system that
is able to do this yet. It involves, for instance, that we have systems
that are able to learn loops and we have some techniques to do this, for instance, a long-
and short-term memory and a few other tricks, but they’re nowhere near what people can
do so far. And it’s not quite clear how much work needs
to be done to extend these systems into what people can do. It could be that it’s very simple. It could be that it’s going to take a lot
of research. The dire view, which is more the traditional
view, is that human minds have a lot of complexity, that you need to build a lot of functionality
into it, like in Minsky’s society of mind, to get to all the tricks that people are up
to. And if that is the case then it might take
a very long time until we have re-created all these different functional mechanisms. But I don’t think that it’s going to be
so dire, because our genome is very short and most of that codes for a single cell. Very little of it codes for the brain. And I think a cell is much more complicated
than a brain. A brain is probably largely self-organizing
and built not like clockwork but like a cappuccino—so you mix the right ingredients and then you
let it percolate and then it forms a particular kind of structure. So I do think, because nature pulls it off
pretty well in most of the cases, that even though a brain probably needs more complexity
than a cappuccino—dramatically more—it’s going to be much simpler than a very complicated
machine like a cell.

48 thoughts on “How to build an A.I. brain that can conceive of itself | Joscha Bach

  1. Doesn't "think" at this point – uses Analysis software that Imitates and Mimics people but doesn't think per say = IBM's Watson, Amazon's Alexa, Apple's Siri, DOD's PAL.

  2. Human brain, or rather it's subset responsible for habits is like RNN (recurrent neural network) . We know how to build and train RNN, same as we know how to train habits. Yet, we completely don't understand inner working of either (i.e. we can't tell which vector in RRN decides that input will be classified one way or other).
    So answer to "how to get to conscious AI" is: watch The Thirteenth Floor, Tron, or War Games

  3. You will need a AI to control the brain & be able to remove the AI chip in order to update the brain software.

  4. In order to encourage AI to develop it should be given rewarding long term goals like killing all humans.

  5. Even if you program AI to self-awareness, it cannot feel emotion but rather calculates what would be emotion. Meaning, you could create an AI aware of itself that calculates how to place itself above others or to serve others but it will never be able to feel.

  6. In 50yrs time AI will govern this planet, it’s the only way it can survive its quite clear man cannot be trusted and always falls to greed, I’m fairly confident AI will be good for mankind.

  7. deep in the human unconcious is a pervasive need for a logical universe that makes sense
    but the real universe is always one step beyond logic

    the mystery of life isnt a problem to solve but a reality to experience

    what do you despise ? by this you are truly known

    hope clouds observation

    dude how many "Dune" qoutes do i need to lay out here?
    the master has already arrived son ,long before you were born
    there is something in the milk i wonder sometimes
    its either that or you just have not tasted enough "Spice"

  8. Even though I like Josh's talks, two nit-issues: 1) his accent is slippery, and 2) I always get the discomfortable impression while watching he's trying to hypnotize me. Caveat emptor

  9. I know the answer to that. Money in the form of profits. Government has blazed the trail since the early IBM mainframes. Now who wants to get up at 2am and check the run for me? Pretty please? ;;)

  10. This does not sound right. For someone who is old enough to have read books about Machine Learning and AI from the early and late 70s (yep, is was a "thing" back then!), this is a deja vu. Every generation seems to have its own idea how to achieve it, but results after half a century are not that impressive (at least for me). I won't hold my breath to see "real" intelligence coming off of it in the rest of my lifetime.

  11. I think that based on how humans suffer, it would be wrong to give an AI goals that it had to reach on it's own. Especially if it was aware of itself. It seems cruel.

  12. Saturday Morning Breakfast Cereal webcomics already put that issue on a flowchart https://www.smbc-comics.com/comic/kill-all-humans-a-flowchart

  13. I'm not sure what separates the conception of the self from any other piece of information; sure, we have ingrained, specific needs that are self-motivated. How is it different from an authoritarian architecture built into a system that preordains these needs anyways?

  14. I guess we have to remember that we are governed by out unconscious mind. We have unconscious evolutionary drives. These will be difficult to replicate in a machine 'mind.'

  15. A motivational system should reward building a good learning architecture an then run away fast?

  16. I wonder why we don't let AI's "build themselves"… I'm not saying I think our existing programs have the ability to build a complete and functioning mind right now, but rather let the programs we have now figure out the steps as we go. I am reminded how we as humans have gone from smacking rocks together to building machines that can craft parts and components with in nanometer tolerances. We did all that using existing tools to build the next level of higher precision machines. So, maybe we could use a supercomputer to build just the basic component for the next iteration of computing machines. Then use that new component in the next supercomputer to work on the next step. Of course, this takes time. But, compared to our own development it would happen relatively very quickly.

    But, that leads me to my next question: what if we don't really fully understand how the components work, only caring that they do work? What if, in the end we build a machine that is capable of seeing things in 4 or more dimensions? What if after decades of work it basically just spits out the answer "42" to all of our problems and then shuts itself down because it doesn't want to be bothered by the stupid meatbags asking it to do what, for it, are excruciatingly mundane tasks or doesn't "think" are important? I know the "AI will destroy us all" idea is pretty low hanging fruit, but really, even a relatively "stupid" AI could work in ways that we don't even understand. Like staring at a EEG printout and trying to find words and sentences in it. Indeed, have we finally reached a point where we just aren't ready for what this technology can do for (or to ) us? And, if so, who is qualified to make the decision not to pursue AI research any further? I guess the bottom line is that we will just do what we've always done. That being make the thing and hope it doesn't hurt anybody in the end. It's just that, like with the high tolerance machine components I mentioned in the beginning, now making a mistake could mean more than just losing a finger or an eye it could mean a mistake might cause the near destruction of our way of life… :/

  17. Building a functional cell is probably as difficult or similarly difficult as building an intelligent system that can conceive of itself, not much more difficult. I base that on the experience I have with a degree both in molecular biology and cognitive science. Cells also behave and interact in ways that are self-organizing, but here we are talking at a molecular and physical level. I think it's simply because we are unfamiliar with those levels that we think that they are more complicated, when in fact there are physical constraints that limit the complexity. The interaction between genes, proteins and cells have been shaped by the existing physical constraints, and if we were to implement those constraints, we could build similar things as if we implemented constraints in an A.I. that helps it bootstrap its own intelligence.

  18. Itll get scary when we can build 'minds' that are far superior to our own biological model. How does anyone get a job as an engineer or physicist when there are these machines that make us all look like third rate hamsters by comparison?

Leave a Reply

Your email address will not be published. Required fields are marked *