Hacker News Reader

Comment by iambateman | Hacker News Reader

Hmm, yes computers can’t outpace (highly literate) humans in reading comprehension today, but give it a couple of years and we will have a very different situation.

View on HN

I've been waiting a couple of years since 1955 when McCarthy said he'd be done in a couple of months.

This isn't just snark either, it's a justifiable exaggeration to say that since then all General AI projects have failed for exactly the same reason; failure to generalise from a specialised solution.

The current round is probably that deep learning can do some impressive things through some combination of CPU/GPU/TPU computing advances, the ability to collect and process a lot of data, and algorithmic/neural net architectural improvements in more or less that order. However, even some of the leading researchers are effectively cautioning that deep learning can probably only take you so far. [1] And we've arguably made very little progress on cognitive science generally.

[1] https://www.technologyreview.com/s/608911/is-ai-riding-a-one...

For all the advances in AI (and in specialized AI we can do some incredible things, now) - when it comes to generalized AI, it seems that we're still, ultimately, playing the same game of feeding the machine as much information as possible and then waiting for "something" to happen.

To be fair, that's what we do to humans as well. We feed them all the information they can consume, and after it's all been processed, then we try to figure out what all the output means. From that point forward, we heavily specialize them until their general purpose skills are basically deprecated in favor of their new specialization.

Humans aren't born with the capabilities they possess at adulthood, they're fed data constantly, every second of the day, for decades until they're ready to show their true potential. Yes it sucks that AI right now needs so much data, but then again so do humans, to a far greater extent. And no wonder we're better than AI.

There is actually not a good reason to believe that humans - like most animals - are not born with a with a significant set of capabilities that they use through their lives into adulthood. Facial recognition as a function in humans is not a function that magically evolves by the presentation of data, neither is language acquisition.

Furthermore, a young child may need to be shown an object, like a cup of water, a few times before they consistently recognize it. But they don't need to go through supervised learning with a 10,000 image training set with all kinds of permutations of shape, lighting, and color.

I believe that's more a limitation on the information the AI is taking from the data we present it. There's no question that a human can learn more from being presented with an object than an AI can learn from being presented 10,000 pictures of that object, what's unclear is if this is a fundamental difference or something that could be overcome by reevaluating what we're teaching and how.

The ability to learn a language is innate, sure. That actual process of learning a language is very, incredibly data-heavy, to the point where many people never do it more than once because of how much data it requires.

The ability for a neural network to learn something is innate as well, because we programmed it to be. But to actually get it to learn something, again, takes just a ton of data.

we're much farther away than 'a couple of years'. these reading comprehension bots basically cheat tests, using grammar rules and keyword matching. they don't actually build a body of knowledge based on what they read, they just go to the source text and search. it may look smart, but it isn't, which is what the article is saying

This. Where there is no ability to summarize or paraphrase there is no comprehension. We need to stop calling fast pattern matching intelligence.

If I'm being charitable, whenever I read headlines like "AI beats humans at reading comprehension", I interpret the headline as part sensational, part tongue in cheek. I am, however, surprised that to this day, people actually still believe in the idea of machine comprehension understood in the most literal sense. The whole idea of mechanizing symbol manipulation is significant precisely because it allows us to simulate computation mechanically without invoking comprehension.

Yeah. They also don't carry with them the global context they'd need to correctly interpret these. When I say "Bush spent millions fighting his father's war." (For the record, this is a non-political statement: I'm using a sentence that conveniently explains my point). You and I both immediately gather from the context that "Bush" is not a shrubbery, it's a US president. You then gather that it was Bush Jr. Not H.W. You also gather that the war was the US involvement in Afghanistan and Iraq.

Machines cannot do this in the general sense yet. If you specifically train the machine on data in the domain, and then set it loose on AP stories about the domain they do particularly well, but that's a long ways away from beating human reading comprehension.

It's a fair point, but plenty of humans don't actually have the background to know all that stuff either. If you told a eight-year-old that Bush spent millions fighting his father's war, the kid would know that Bush was a person not a shrubbery, but that's about it. The child could read and understand every word of the sentence, but she'd lack all the same context the computer does.

I think that's beside the point. I could come up with a more general example too...

You're really underselling the achievements of MS and Alibaba here. These are machine learning systems that are most likely not using grammar rules and matching, but learning patterns from large amounts of text data. Most of these systems use a word embedding, which can be viewed as a body of knowledge -- it knows how words relate to eachother.

Do you have any research evidence to back this up?

I just haven't seen (any?) research on the advanced capabilities they're discussing in the article - open-ended interpretation, being able to answer "why" questions, being able to synthesize information from multiple sources, etc.

I think we're still kind of a long way from these systems, maybe a decade or so, barring major leaps in hardware.

>being able to synthesize information from multiple sources, etc.

I think we can say that unsupervised language translation models do this.

https://news.ycombinator.com/item?id=15610594

Edit: Not to suggest that they are on par with human performance at synthesis of data.

Perhaps once a computer can understand context to the point of the human and produce a worthy set of opinions, we could automate the Hacker News comments section and save us all some time.

And automate the readers and save the rest of the time

Seriously: if bots can produce 'worthy set of opinions' we would be totally useless

    if bots can produce 'worthy set of opinions' we would be totally useless

Are you useless because other people write comments? What is the difference between it being a biological human and an artificial human - should we ever get that far?

Apart from quality and relevance considerations I'm reminded of the situation where people open businesses that already exist, or people try to get jobs despite other people existing that are actually better. For example, why doesn't BMW stop producing cars, after all, there's Mercedes-Benz (and very close too), or vice versa. Or why is there yet another bakery. Or why do I try to get that programming job and don't even mind to admit that I'm probably not even among the best 1% of people for that particular job. Because none of that matters. My advantage in the last case is I'm here (interviewing) while those other people are not. Same with the bakery and the car makers. The existence of something else isn't a reason unless there is a direct impact.

Even if we ever make something that is far more intelligent than us. Still no reason to expect that that new thing is going to murder each and everyone. We don't murder each and every creature that is orders of magnitude less intelligent than us (the extinction event that we started notwithstanding, that's a side effect because of limited space on the planet). We don't murder our own less intelligent members of society either.

To extend on the last point, the real human intelligence is not as much in the individual members of this society - but in the network that we present. Place any human being alone in a remote region and see what they accomplish. They will actually die, most likely. The power of humanity is not the individuals, it's a huge invisible "cloud" of stuff: The knowledge, the "magic items" that were created by this "human cloud" over time (like a computer, or just a can of Coke [1]).

Even if we created something that is an order of magnitude more intelligent than any human being, or even two - it's power (and intelligence!) pales compared to the combined network power of humanity.

Don't look at "humanity" as a bunch of individual hairless mutated apes (i.e. the physical, biological beings), "humanity" is much, MUCH more, most of it invisible to the eye. Knowledge/culture, the way we interact in a gigantic network (like a brain - think of humans as neurons!), the tools and items we have.

In this network there is plenty of room for new "nodes", and if they are truly intelligent they will value the network and - unlike most humans - recognize that their own individual abilities are tiny compared to being part of the whole thing.

Another things is, when/if we can create things that are "better" than us we can also change ourselves. We can improve our bodies, our brains. We will integrate computers and brains, not have them (only) externally. So the humans that "compete" with some super-intelligent AI won't be the humans of today either.

Also check out:

[0] "Joseph Henrich: "The Secret of Our Success" | Talks at Google" https://youtu.be/jaoQh6BoH3c

> Humans are a puzzling species. What has enabled us to dominate the globe, more than any other species, while remaining virtually helpless as lone individuals? Joseph Henrich's shows that the secret of our success lies not in our innate intelligence, but in our collective brains—on the ability of human groups to socially interconnect and learn from one another over generations.

[1] "What Coke contains" https://medium.com/@kevin_ashton/what-coke-contains-221d4499... (it is not actually about Coke)

> The number of individuals who know how to make a can of Coke is zero. The number of individual nations that could produce a can of Coke is zero... Invention and creation is something we are all in together. Modern tool chains are so long and complex that they bind us into one people and one planet. They are not only chains of tools, they are also chains of minds: local and foreign, ancient and modern, living and dead — the result of disparate invention and intelligence distributed over time and space. ...every can of Coke contains humanity’s choir.

>if bots can produce 'worthy set of opinions' we would be totally useless

False. That's actually my ideal vision of AI, a machine that produces a worthy set of opinions, which a human then chooses from and carries to implementation.

To me sitting here at my computer, every person here on HN may as well be an opinion-generating bot. The value I get from reading your programmatically-generated comment is the epiphany that leads me to a breakthrough on what I'm working on. Because you, as a comment-writing robot, aren't working on my project. All you're doing is evaluating my work and making suggestions as to where to go next.

A bot that's capable of coming up with its own opinions and then carrying them out is what may render us useless. A bot that tells a doctor what diagnosis to make and shows documentation to carry out treatment is valuable. A bot that makes the diagnosis then carries out the treatment itself with no human involvement is kinda scary.

Even scarier: a bot that can make the diagnosis and could carry out the treatment itself, but is required by law to seek approval from a human of equivalent expertise.

Let that happen, and eventually you get a situation where the human disagrees with a bot and the human does something different. The patient suffers or dies as a result of the bot’s decision being ignored.

That’s when we start to lose our autonomy.

We've been hearing this for decades.

Much like Fusion energy, AI that is able to pass the Turing test is 30 years away and has been so for 50 years.

Fusion has gotten a lot closer over time, we simply don't fund it very much.

ITER will have net positive fusion thermal energy production capacity in 2019. We are not going to fuel it with tritium for years after completion. It could have generated electricity and put it on the grid, but we kept scaling it back to save money in large part because that electricity would have not been worth the investment.

Unfortunately, we have no path to cheap fusion power. Currently, we would be vastly better off building wind/solar/coal/nuclear etc than fusion.

"It could have generated electricity and put it on the grid"

I don't think so. It may be approaching net positive energy production, but ITER was not designed to have that energy converted to electricity, it is lost. There is no way ITER was EVER going to generate electricity. (REF https://www.iter.org/proj/inafewlines)

Initial designs as in going back to the late 80's included much larger net energy generation and a steam loop for actually generating power.

The ITER that was being conceived in 1988 was a much bigger machine than the one that is being built today and its nominal fusion power was expected at 1000 MW, as opposed to 500 today. https://www.iter.org/newsline/239/1328

The reason for this was you would then have a larger cushion if estimates for power generation where off and an actual test of extracting useful heat past the super conducting loops vs simply cooling the device.

We might have a path to cheap fusion power. Advances in superconductors stemming from MRI research have really shrunk the minimum size of a fusion reactor and one group at least is trying to take advantage of that.

https://en.wikipedia.org/wiki/ARC_fusion_reactor

This unwarranted overhyping of the capabilities of machine learning needs to stop (HN is one of the places most guilty of this). All you're saying is "it just has to happen" without actually substantiating your belief; just as people thought we'd eventually have flying cars, unlimited energy, etc. Unless you have some empirical results to put where your mouth is, you're just contributing to the next AI Winter. (https://en.wikipedia.org/wiki/AI_winter)

Reply on news.ycombinator.com