Hacker News new | ask | show | jobs
by pron 2991 days ago
> He doesn't present an honest understanding of his own field, or the field of neuroscience, or the ongoing developments in the technology surrounding his own business, or its implications.

Yes, but I don't think you are, either.

The fact is that chess and Go abilities aside, we are probably not even close to insect-level intelligence, and we don't have a clear path of getting there soon -- let-alone anything human-level. Current state-of-the-art so-called AI is basically powerful statistical regression algorithms, that are heuristic improvements over core algorithms invented in the 1960s, and there have been few theoretical breakthroughs since then (far fewer than in most other fields), so much so that many consider machine learning to be in the pre-science or pre-theory stage, being mostly about collecting data and trying out heuristics. It's silly to deny recent successes -- largely due to better hardware (although hardware progress is slowing down quickly) -- but we are behind, not ahead, of where we thought, even as recently as the 1990s, we'd be by now.

At this point we have no idea what role statistical regression plays in intelligence or whether we're even in the right direction. That statistics has become synonymous with intelligence (it used to be synonymous with lies) is certainly a cultural phenomenon that is not directly related to our actual knowledge of the field.

That computers perform some mental tasks (certainly more and more of them) better than humans has been a fact of computing since the 50s, and often a cause for wild claims. The invention of neural networks in the 40s and their implementation in digital computers in the 50s led some very respectable people (like Norbert Wiener) to declare that the problem of the brain will be solved in 5 years. The pragmatic Alan Turing thought that was ridiculous and predicted it would take 50. It's been almost 70 years and we haven't yet reached insect-level intelligence or anywhere near a complete understanding of the insect brain, so at this point, any claims that we are on the cusp of something, or starting to believe that our statistical regression algorithms reflect the beginning of intelligence is... misplaced.

On the other hand, it seems like we have not learned the lessons of misplaced confidence in AI, despite our relatively slow progress, and things are worse now as that we actually have some algorithms that are useful in certain restricted domains that we insist on calling AI, thus causing people to use them in domains where they are not only useless but downright harmful. In the meantime, some people draw attention to the dangers of real AI -- which may be anywhere between decades and centuries away (I believe we'll get there some day but we have no idea how or how soon) -- while distracting from the very real and already present dangers of "AI".

3 comments

> we are probably not even close to insect-level intelligence

I think the problem is that we do not have a slightest clue what is (even insect-level) intelligence (or consciousness, which is often mixed up in the discussion).

That's right. Some have tried describing intelligence as a general problem-solving skills, but this is clearly false. Humans are terrible at finding even approximate solutions to NP-hard problems, which are certainly general and very common. It seems like intelligence is an ability to solve many problems that humans and animals face, but no one has characterized it more precisely, AFIK.
> hardware progress is slowing down quickly) I would be interested to know more about this. I haven't heard yet that the progress of GPUs cores for example is declining quickly ...
The problem is Amdahl's law. You can only parallelize so much. While the brain is certainly extremely parallelized, neural nets do not employ the same algorithms as the brain, and so, unless we find algorithms that are more amenable to parallelization, Amdahl's law is going to get us.
Most modern neural networks implementations are parallelized. And that is why we can run them extremely well on the GPU. For example Volta GPUs delivers 5X increase in deep learning training compared to prior generation NVIDIA Pascal architecture. This is why I was asking for clarification about the hardware claims.
Of course they are, but they're not as parallelizable as the brain, which is why they're subject to Amdahl's law.
The Blue Brain project managed something “as big and complex as half of a mouse brain” in 2007, so I think your claims of not-even-insect-level are outdated.
Not so sure that claim is right : http://www.artificialbrains.com/blue-brain-project

Seems like they managed a honeybee (but I am not sure that it ran in real time or how they validated that) but were hoping for a rat brain.

I'm quite sceptical - I don't believe that there is a good understanding of how a single neuron functions, or agreement on the taxonomy of neurons or an understanding or agreement on their interactions and arrangement apart from in a part of the vision system where there do seem to be some good models.

Thanks for the link. I noticed it was “As of August 2012”, which despite being old still falsifies my prior belief. (Gell-Mann Amnesia, I incorrectly relied on the press for science). Unfortunately I can’t seem to find anything more up-to-date and just get more confused journalism.

That said, isn’t the point of the Blue Brain project to answer your skepticism? When we have all those things, the only thing remaining is to see what can be left out of the sim without compromising the behaviour?

The OpenWorm project doesn't even manage a worm though.
So? OpenWorm is (barely) crowd-funded and trying to simulate the entire body not just the brain. Seems like a decent effort given their resources, but on a scale of Radioactive Boy Scout to Manhattan Project they’re a Farnsworth Fusor.

(Which is to say “I ought to volunteer”).