Comment by SomeStupidPoint | Hacker News Reader

Comment by SomeStupidPoint | Hacker News Reader

Parent Full thread

SomeStupidPoint

I don't think you're wrong that AGI will have (something like) those components in its inception.

I just disagree that we have to intentionally create them -- I think there are a lot of places we can experimentally bruteforce the implementation instead of having a solid theoretical understanding if our only goal is creating an AGI.

I don't think we need any theoretical breakthroughs, per se, but rather, a lot of computing power and time. No one seems willing to take a million nodes and run them for a decade to brute force some of the mechanisms -- everyone wants results on the short timelines that grants or Wall St operate on. I get why, but that bias towards quantifiable short-term gains fundamentally limits the search algorithms we're implementing in our quest for AI, and likely means that we won't get there in a meaningful way, because intentionally implementing the necessary requirements is a fool's errand of complexity.

In short: we're not smart enough to do the simple, but slow thing, so we're trying the highly complex one with demonstrable incremental results. I expect this approach to continuously fail to develop AGI even while it demonstrates results on discrete problems.

I think you are right that we could, in theory experimentally brute force intelligence. But if we try and do that, we are basically trying to simulate an entire universe and wait for intelligence. But the amount of computing power that that would require is beyond even what we can imagine. And even if we could find some ways of simplifying it, it still seems like a waste of our intelligence.

Like when it got cold for humans, rather than just dying and waiting until they evolved fur, they just made clothes. Evolution is inherently able to do great things but is very inefficient. We have the power to reverse engineer nature and use that knowledge to our advantage, and if our goal is to create general AI, then not trying to understand the one sitting in everyone heads and just trying to throw more computing behind it just seems like a waste of a tool that has allowed us to get this far.

I'll preface this by saying that my views don't necessarily represent a mainstream position here:

Humans already accidentally invented AI at least a few times, sort of. To talk meaningfully about this topic, I think it's better to step away from the topic of "intelligence" and instead speak about "reasoning ability". It's still fuzzy, but seems to sidestep a lot of the "can submarines swim" arguments.

Things like societies, governments, corporations, economies, etc all show similar features to machine learning algorithms and reasoning animals in their ability to dynamically problem solve -- the main difference is that they run on human computers instead of electronic computers. (Or photonic, quantum, etc computers.) Well, in modern times, they're actually sort of cyborgs, but that's a side issue.

The quest for an AGI is more about translating the notion of a corporation (or government, etc) meaningfully into fully electronic computer terms than it is in replicating the human mind. We know enough about the supra-structure of the corporation to meaningfully bound the search, and it's much more prudent to spend a decade on bruteforcing each component of a corporation than it is to navel gaze about reinventing minds. The truth is that human brains (or animal brains) aren't the only general intelligence that we know -- so reinventing them isn't the only path to AGI, though it's a worthwhile project in its own right.

I just worry that many of the meaningful contributions of the people working on reinventing minds will be lost on the birth of the actual first AGI through corporate cyberization. (And that many of us will fail to recognize corporate cyberization for what it is -- an AGI. Any (successful) corporation which can act fully by algorithm must be a general intelligence.)

1. What are the other general intelligence that we know of besides life?

2. How is it that you know that the reasoning abilities of the structures that we create are because of the structure, and not because we are the pieces that make up the larger structure?

1. Sorry if I was unclear -- the examples I listed were meant to be examples of "general intelligences", but I think corporations are a good example.

2. Well, this is somewhat more complicated to answer --

Obviously the reasoning abilities are (partly) because we're part of the structure. But in a way that we know can be decoupled from the reasoning ability of the structure -- see modern CS on which problems are tractable to mechanize and the success of ML at individual tasks -- because modern corporate theory has formulated a system of corporate governance and execution that only depends on individual people for one or a few tasks. The fixation of modern corporate management on replaceable, cheap, low-skill workers makes it a perfect template to replace components one at a time with bruteforce discovered ML components (and in fact, we already see this happening -- because it is part of the reason that corporations are designed that way).

After all, from the perspective of the corporation as a reasoning entity, increased cyberization is increased efficiency, which ultimately means increased competitiveness (and hence survival).

Electronic computers were designed to replace human computers in the military, and we have a wealth of information on their numeric and string processing capabilities. They're specifically designed to emulate in another medium some of humans' reasoning abilities, particularly those around logical deduction and numeric calculation (which in some senses are similar). Since modern corporations have reduced much of their operation to machines (eg, factories), string manipulation and calculation, it seems perfectly conceivable to have a corporation control all of its "core functions" via electronic computer after a period of increasing cyberization, with contracted delivery services and on-site technicians viewed something like a doctor.

Corporate cyberization is, if you take my first claim that corporations represent a "general" reasoning structure of made of specialized components as true, an open and direct path to an AGI that we seem to be stumbling down.

Some very good points.

I mean, what is the problem we are trying to solve here? Fully emulate a human? That's worthless, we have an overabundance of humans already and it's easy to make more.

Or are we trying to come up with something that performs our tasks without us having to think or do any work and presumably makes life richer and more pleasant? Well that's a different problem, and yes, we have been automating that forever and are arguably getting better at it. (I say arguably because "richer and more pleasant" are relative as it turns out).

The argument over what constitutes "intelligence" is not an insignificant one either. My dog likely thinks I'm stupid because I don't eat butter all day and shave and go to work. And come to think of it, he might be right. Point being, we often conceive of intelligence as "having better ways to do or get what we would". But then there is abstraction, longer term planning, higher order thinking, which in the end almost certainly will be incomprehensible to humans.

>The quest for an AGI is more about translating the notion of a corporation (or government, etc) meaningfully into fully electronic computer terms than it is in replicating the human mind.

That's an astoundingly terrible idea, considering what powerful corporations and governments have done to the world.

I didn't mean it as a normative statement, just one about the technology required -- I share you concerns.

Perhaps an important warning that AI could go the same thing...

Reply on news.ycombinator.com