Comment by randomsearch | Hacker News Reader

Comment by randomsearch | Hacker News Reader

Parent Full thread

Can someone please explain what has happened in ML or AI that makes AGI closer? Whilst some practical results (image processing) have been impressive, the underlying conceptual frameworks have not really changed for 20 or 30 years. We're mostly seeing quantitive improvements (size of data, GPGPU), not qualitative insights.

ML in general is just applied statistics. That's not going to get you to AGI.

Deep Learning is just hand-crafted algorithms for very specific tasks, like computer vision, highy parameterised and tuned using a simple metaheuristic.

All we've done is achieve the "preprocessing" step of extracting features automatically from some raw data. It's super-impressive because we're so early in the development of Computing, but we are absolutely nowhere near AGI. We don't even have any insights as to where to begin to create intelligence rather than these preprocessing steps. Neuroscience doesn't even understand the basics of how a neuron works, but we do know that neurons are massively more complex than the trivial processing units used in Deep Learning.

Taking the other side for a moment, even if we're say 500 or 1000 years out (I'd guess < 500) to AGI, you could argue that such a period is the blink of an eye on the evolutionary scale, so discussion is fine but let's not lose any sleep over it just yet.

What I find most frustrating about this debate is that a lot of people are once again massively overselling ML/DL, and that's going to cause disappointment and funding problems in the future. Industry and academia are both to blame, and it's this kind of nonsense that holds science back.

I think the most accurate answer is that we just don't know. Since we really don't know how an AGI could work, we have no idea which of the advances we've made are getting us closer, if at all. Is it just an issue of faster GPUs? Is the work done on deep learning advancing us? I don't think we'll know until we actually reach AGI, and can see in hindsight what was important, and what was a dead end.

I do take exception to some of the specific statements you make though, which make it sound like the only real progress has been on the hardware side. There's been plenty of research done, and lots of small and even large advances (from figuring out which error functions work well ala Relu, all the way to GANs which were invented a few years ago and show amazing results). Also, the idea that "just applied statistics" won't get us to AGI is IMO strongly mistaken, especially if you consider all the work done in ML so far to be "just" applied statistics. I'm not sure why conceptually that wouldn't be enough.

It's _mostly_ hardware and data. There are some smarter steps in training etc., but most of the ideas have been around for decades; it's the scale that made the difference.

> I'm not sure why conceptually that wouldn't be enough.

This one is harder to refute. I guess it's because statistics doesn't involve understanding. Try considering something like LDA for topic discovery: there's no understanding of the semantics of the model, it just identifies them statistically. There's a huge difference.

It's funny that you mention Relu. People have recently trained Imagenet networks using sigmoid/tanh (e.g. the activations that were used decades ago) on GPUs and they train just fine. They train a bit slower is all. Not the breakthrough you're making it out to be. Relus were a very useful stop-gap in 2012 when GPUs weren't as fast.

Now that we know how to initialize the weights so as to have the layer activations be something like sane, yes, we can use sigmoid/tanh. If you don't know modern clever ways of initializing weights then multi-layered sigmoid/tanh causes your activations and gradients to die out fast in deep networks, and ReLU is a godsend.

The biggest advance that I've seen towards AGI is the work using reinforcement learning, e.g. neural nets that learn to play video games through trial and error. There is an impressive repertoire of _behavior_ that emerges from these systems. This, in my opinion, has the greatest potential to take us another big step towards -- but not necessarily to -- AGI.

That advance happened by 1992, with TD-Gammon [1]. Our hardware and software are clearly much better now, but this seems like a solid example of what the GP said: the conceptual framework has stayed the same for 25 years.

[1] https://en.wikipedia.org/wiki/TD-Gammon

You’re engaging in the time-honored tradition of dismissing progress with the term “just”. In the spirit of the article, I recommend you list and publish specific things that are too hard to achieve in the next five years. And then commit to not dismissing them post-hoc.

This is a really difficult question to answer, because humans are super-clever. Whatever I pick, humans can quickly work out ways to game the metric. Turing invents the Turing test, and rather than go for an AGI, we take shortcuts to spit out convincing sentences.

I think for a start you'd have to move away from things that can be gamed through statistics on large amounts of data.

For example, show a child a single object, it can then recognise instances of that object all over the place with almost perfect recall (in the statistical sense). I think a computer would find this a hard task. Eliminate the advantage of big data.

Or perhaps turn it around and put the emphasis on the machine to invent its own test for intelligence, allow the machine to come up with something that is convincing - make it argue for its own consciousness with an argument that it creates entirely for itself.

But... I'm sure someone would find a way to game these examples. That's because humans are very smart. We've outsmarted Turing, so I don't hold much hope for my snap ideas in a five minute HN post :-/

A very interesting opinion combo:

1. Of course we should be prepared for the existential threats of AGI and ASI. 2. BUT the threat isn't imminent, so we should prepare later.

The article (and I, mostly following its lead) is trying to encourage people to concretely answer the question "Okay, if not now, when? How will you know?"

The problem is, most people aren't answering based on a model ("If we can solve problem X, we have 50% probability of AGI within X years.") Instead, they're using the difficulty heuristic, and the insufficiently-impressed heuristic. ("This is really hard right now, and I'm not impressed by what I've seen so far. Therefore, 100 years.")

Your concerns about gaming are only a problem if the notes were to form the basis of an argument. I was suggesting you have them for yourself. (The act of publishing them is to encourage thinking about them now, not to be a gotcha later.) So you'll know what you mean, you won't be arguing with yourself over definitions, and you'll have thought hard about what looks dangerous to you. It's about being honest with yourself.

Incidentally, the child example is misleading. Children spend literally years understanding things like depth perception, object permanence, etc. A child is already a highly trained agent; that training comes from daily interaction with the environment. You show a baby an unmoving object somewhere and to my knowledge there is no evidence that the baby will identify it as a separate object, much less recognize it in a different configuration.

While part of me agrees with your analysis, I'd like to point out what I think could make this wave of ML/AI more serious. You are absolutely correct that deep learning is not very biologically accurate and that what today's models do seems a long way from AGI. However, in my opinion, the most fundamental aspect of intelligence is the ability to form useful abstract ideas to model reality. To make that more concrete, as a rather extreme example, consider the invention of numbers. The process by which people developed the notion of abstract quantity separated from any particular real experience is, to me, the most archetypal example of what it means to be intelligent. Of course, deep learning can't invent abstract math, but it seems to be able to mimic this process in a very rudimentary way. It's not a faithful representation of real neural networks, but perhaps it has just enough of the right ingredients, scale, depth, non-linearity, hierarchy, such that it is able to demonstrate a spark of that magic, hard-to-define process of intelligence. When a deep net learns MNIST, it seems to come up with an abstract notion of what a handwritten 9 looks like and it's hard to argue that there isn't something very mysterious and special happening.

Your example actually runs counter to the idea that we're seeing a massive breakthrough. Deep nets on MNIST for recognizing numbers were done 20 years ago.

> it seems to come up with an abstract notion of what a handwritten 9 looks like

It does indeed - it comes up with features that indicate what a handwritten 9 looks like. But it doesn't develop the concept of what 9 _is_. It doesn't say "well, that's a concept I can apply to lots of places. Hey, I wonder what nine nines look like!" It's doing pattern recognition on pixels, which is cool and no doubt what we do to some extent, but it doesn't have that higher level of reasoning.

Agree. Deep learning does not bring us closer to AGI. It might get us closer to other proxies of "mechanical intelligence" that will be very productive.

I now believe we are 3 years from building an AI that writes Python well enough to build itself, based on some experiments I did recently: http://sparkz.org/ai/program-synthesis/2017/10/12/self-hosti...

Most technical people will understand the difference between programming and AGI. The general public might not.

The useful thing out of AGI discussions, is that they engage the general public.

>> I now believe we are 3 years from building an AI that writes Python well enough to build itself, based on some experiments I did recently

Why 3 years? Can you elaborate on the timeline? What should happen in 1 year, what in 2, what in 3 etc?

>ML in general is just applied statistics. That's not going to get you to AGI.

I don't see how we can rule it out. The size of the statistical models we use are still dwarfed by the brains of intelligent animals, and we don't have any solid theory of intelligence to show how statistics comes up short as an explanation.

We can learn concepts rapidly from much less data than statistical methods require.

One shot and zero shot learning also use statistical models.

Reply on news.ycombinator.com