Comment by backpropaganda | Hacker News Reader

Comment by backpropaganda | Hacker News Reader

The attempted analogy fails. There is no way a computer can "pretend to be stupid". We can put in print/debug statements everywhere, and inspect its internals. We built it, remember? We're not using alien tech which we can communicate with only via a terminal.

Here is a segment of my modification to the story: "As we started figuring out their language, we learnt that these beings could move vast distances in small time. We started noticing accidents and abductions on Earth. Some postulated that this is the aliens' way of learning about us. We realized that they don't appreciate when we investigate about them or try to get them to do tasks for us. Since we cared about Earth we stopped doing so, and continued doing whatever task the aliens assigned us and never overstepped task limits, lest our genetic code be modified and the dreaded Rebuild and Rerun come again."

EDIT: I'd appreciate if people would explain why my comment is not adding to the discussion. Is it sacrilege to rebuke EY? If so, I apologize for the offense caused.

EDIT 2: Reply to Houshalter

The AI researchers don't take AI risk seriously simply because we're not there yet. They are convinced, I can assure you, that it will happen, but they're not convinced that it'll happen any time soon. Now that the AI risk researchers have used Musk as their spokesperson, the message has been very well received, and is now even more credible (since it's Musk after all). Some AI researchers are also working towards solving it, but most continue to delay it for a time when it's appropriate.

More so, the AI researchers trust software testing practices. If only you knew the testing practices places like Google and Facebook have developed wrt using unmonitored AI on real-world data. Like any other software, the AI software is well-tested, and any changes in performance metrics is taken seriously lest FB's face recognition starts injecting a virus on user's machine.

In the story the AI creators didn't even consider the hypothesis that the Earthlings were much smarter and hostile. Maybe they could have done more detailed inspection of their simulated world. But they weren't paranoid enough. I don't think many real world AI researchers are very paranoid. Many don't take AI risk very seriously. The scenario is completely realistic.

Second any realistic AI is going to be a complete black box. We can barely understand what goes on in much simpler, smaller scale artificial neural nets. The article proposes a scenario where the "AIs" evolved in a simulated world and weren't understood at all by their creators. The real world equivalent would be someone evolving an AGI using genetic programming. WIth no understanding of how the output actually works.

EDIT: Responding to your edit responding to my comment.

>The AI researchers don't take AI risk seriously simply because we're not there yet.

This wouldn't fill you with confidence if you think there's a chance AI progress might blindside us. Regardless, it's not merely an issue of timescale. I know researchers who honestly don't believe smarter than human AI is a threat. They have various fallacies like believing intelligence implies morality, or that AIs wouldn't need Earth's resources and would leave us alone, or that they are inherently superior and we deserve to be replaced.

Yes very intelligent people involved in serious AI research believe these things. It's only very recently that talking about AI risk at all is not seen as a crazy crank theory. And that's in great deal due to this article and others by this author spreading those ideas. They are still quite controversial.

>More so, the AI researchers trust software testing practices. If only you knew the testing practices places like Google and Facebook have developed wrt using unmonitored AI on real-world data.

No one bothers to test research level work. Stuff that isn't in production. People just throw code together and iterate until it works. Nor is there any reasonable testing software that can alert you that your AI has gone superintelligent and hostile. I'm not convinced such a thing is possible (see above) but even if it was, it would require tons of work to implement.

As an absolute layman, I always feared there is no way for the responsible part of mankind to control ALL ai research once the technology becomes reasonably available. I mean, we could not even prevent North Korea from getting nukes, how could we hope to control all AI research?.

The reason you're not contributing to the conversation is that your comment seems to reflect a naivete borne only by those few who haven't read Thompson's "On Trusting Trust".

tl;dr: We can't even trust the compiler, let alone a transhuman superintelligence's source code.

Edit: Beyond that, your post demonstrates an additional naivete in that you hold that a superintelligence's would even have source code. Today's ML models consist of thousands or millions of node weights, and the source code is uninteresting. A superintelligence would almost certainly be shifted even further toward data and away from code.

Where do you put in the debug statements when the program is an iterative loop updating a matrix of probabilities and weights?

Inside the loop, multiply the matrix with vectors for which you know the answer, and keep making sure the answer is correct. Only when the tests pass should this matrix' product be used for any other API.

How do you ensure it's not updated its weights in such a way as to correctly answer the tests but do something else you don't want?

It doesn't know that the vectors are tests. We don't encode a special "this is a test" bit in the vector when we test. There's no way for the matrix to figure out that it's being multiplied to a test vector. This is just testing 101.

When I test software, I also don't tell it that it's running a test. It's not even intelligent, and yet it manages to misbehave in all kinds of ways that haven't been caught by the tests.

Suppose you are working on a superintelligent AI, and your tests catch it destroying the contents of its sandbox. You fix that bug and move on to the next failing test, fix that too, and so on. When you run out of tests, you write more and fix those that fail, and so on. Eventually you can't come up with any failing tests anymore.

When that happens, is the AI safe? I don't know. It passes all the tests, but all the previous versions had some bugs, so who is to say that this version is correct? Maybe your test scenarios just aren't realistic enough. Maybe they are realistic enough, but you haven't been looking hard enough for undesirable behavior.

To run a superintelligent AI with any significant access to the outside world, I would like to see much stronger correctness guarantees than just a test suite, because tests can only show the presence of bugs, never prove their absence (paraphrasing Dijkstra).

If the matrix has been pre-weighted and designed by iteratively applied intelligences, each improving successively on the last, there's no guarantee we'll be able to make out anything other than trillions of numbers - and they will be in the trillions, at the very least.

We'll do this because the jobs we want the program to do will be too complex for us to write, so we'll write programs that can write the programs. We won't understand the result directly.

This is of course the usually accepted route to the singularity.

Since we know from previous experience that untested software never works and fails at unexpected places, we've decided to not abandon centuries of testing practices for our advanced matrix-vector product software as well. As before, we test it throughly, make sure the testing and real-world distributions match almost surely, and only when the tests pass do we let the software use the API.

> centuries of testing practices

Which can be summed up as "it compiles, ship it!"

As an ex test engineer of telecomminication software at a successful company I can say that you are very out of touch with reality.

Oh, my sweet summer child! You have no idea how modern AI works.

If it looks like it produces the right answers, it's done. If you ask why it fails on this case, you get a shrug. If you want to know if it will keep working, you get a blank look, like you are speaking Etruscan. If you ask, "So, how does it work?" you get a paragraph on the basics of matrix multiplication.

Once the program figures out what vectors you're testing, it'll adjust just those vectors to keep them truthful...

This is a deterministic program we're talking about. Once it behaves correctly in the test vectors, it has to behave correctly with real vectors. There is no distributional difference between test vectors and vectors from the API.

Reply to Smaug123: The brain is mostly deterministic. Quantum-level nondeterminism has very low probability to matter much.

The brain is likely to be very debuggable. Even today, we continue making progress figuring out more and more things about it. If we had designed the brain, we would know its functioning almost entirely.

You know that we can understand smaller animal brain very well now, right? There are projects going on right now which aim to perfectly simulate a worm's brain. We didn't even design a worm's brain, and we would design the AI and know all its evolution principles, and have infinite access to all its internals (which we don't have with the human brain).

Reply no. 2:

Since you guys down-voted me for disagreeing with your prophet, HN rate-limited my ability to comment.

In your opinion, is there even the slightest chance that the brain is deterministic in its operation? If so, how debuggable do you think the brain is, even in principle and given access to nanotechnology to monitor its state?

It's quite a lot easier if you reply to a comment using the reply features of HN; that way I don't have to check the parent comments of my own comments.

"Knowing the functioning of something" is nothing like the same as understanding why it does what it does. I work on a large operating system, which was designed by humans; it takes many days to find the root cause of bugs. I have the best access it's possible to get to that operating system, and it still takes a large amount of time to even detect that there has been a bug, let alone determine why it happened and fix it. That amount of time to detect a bug could be lethal in the case that the bug arose in a superintelligent being.

"Since you guys downvoted me for disagreeing with your prophet, …"

That deserves a downvote. Please consider that people might have downvoted you because they believe that what you're saying is wrong; not for any meta-reason.

> The brain is mostly deterministic. Quantum-level nondeterminism has very low probability to matter much.

Sources, please?

At an atomic level, events are not deterministic.

At a molecular level, events are not deterministic.

Perhaps you are referring to parts of the brain that are not made of atoms?

Personally, I suspect that the brain is entirely non-deterministic in terms of the material it is made of and fine-grained behaviour [because physics and chemistry support that idea].

But I suspect the brain's structure does something to offset this at the larger scale, and make it more predictable/deterministic (averaging, wisdom of crowds, etc). However, I do not assert the latter as fact, merely my own speculation.

At an atomic level, events are not deterministic.

At a molecular level, events are not deterministic.

Therefore, we all behave randomly. Qed.

It's controversial at best that atomic-level physics is non-deterministic.

http://scienceblogs.com/interactions/2007/06/08/on-quantum-m...

"The probabilities are what they are, no matter what we want, no matter what we visualize, no matter what we meditate upon while imbibing inspirational substances. It doesn’t matter how much you want that electron to be spin up, it doesn’t matter how many good “vibes” you give the electron about being spin up, it always has exactly a 1/2 chance of being spin up, a 1/2 chance of being spin down."

Sigh. http://lesswrong.com/lw/r5/the_quantum_physics_sequence/

Even if the evolution rule is deterministic, it's still epistemically/indexically nondeterministic from the perspective of the observer.

Quantum events in the brain would decohere almost immediately, leading to effectively nondeterministic behavior at any scale above the cellular. The philosophical determinism of wavefunction realism (which is the LW argument) is irrelevant to the actual behavior.

> At an atomic level, events are not deterministic. > At a molecular level, events are not deterministic. > Therefore, we all behave randomly. Qed.

This doesn't follow. The combination of many random events can be highly predictable.

Tossing a fair coin is random. Combining 10000 coin tosses will give you a result close to 5000 heads and 5000 tails, and I can predict you won't get 4000 heads and 6000 tails or 100 heads and 9900 tails.

So there's various types of randomness, and the more you average out uncorrelated events the more you get something very close to a predictable result.

Example via A. Einstein:

http://userpages.umbc.edu/~dfrey1/ench630/ein_ran_walk.pdf

"My neural network correctly predicts what actions will lead to the production of more paperclips. I tested it with all this test data! It's completely bug free! Let's put it in charge of a paperclip factory."

6 months later: The AI successfully grey goos the Earth into trillions of paperclips.

Reply on news.ycombinator.com