Hacker News Reader

Comment by dreamcompiler | Hacker News Reader

This happened right before the first AI Winter in the late 80s: AI (in the form of expert systems) solved a number of hard problems and was hyped as being able to solve every problem. Reality set in when we figured out:

1. It didn't scale and

2. Getting 80% of the problem solved was easy, but getting that last 20% was very, very hard. Maybe several orders of magnitude harder than the first 80%.

Nowadays we don't seem to have problem 1 quite so much, but problem 2 is still there in a big way. Witness self-driving cars, where driving on an interstate highway in broad daylight is easy, but driving through a snow-covered construction zone at night is impossible. Or just dealing with a bicyclist on the road without killing them.

We're not going to have AGI any time soon.

View on HN

The first "AI winter" happened in the 60's. In 1949 Warren Weaver published a memorandum outlining his plan for universal machine translation. He likened the problem to cryptography, a field that had exploded due to WWII, and thought it would be solved in a few years. In hindsight, we know that his ideas were very naive but Weaver was an influential and charismatic person so they inspired lots researchers. They also attracted lot of funding from the US government, likely because of the envisaged military and political uses of machine translation. Then in 1966, the ALPAC committee setup by the US government published a report that concluded that machine translation were infeasible and thereby killing the funding.

Work on machine translation resumed at IBM in the late 80's, but then based on statistical methods.

It feels harsh to call him naïve because I'm sure I wouldn't have had any more insight, but it is surprising how little we knew about the science of language pre-Chomsky, considering that we have been learning languages for millions of years. Now that we have had several generations of computer technology fall short of the abilities of human translators, it seems obvious that you're probably not going to be able to write a machine translation program on a computer with a few megabytes of memory.

Did you mean post-Chomsky? It's my understanding that machine translation started to become semi-useful after abandoning traditional grammars (or anything Chomsky would recognize) and switched to statistics.

> It's my understanding that machine translation started to become semi-useful after abandoning traditional grammars...and switched to statistics.

What's curious about this statement (and forgive me if I am in error here; I am not extremely familiar with the machine translation field) is that the community chose to go down the route of "traditional grammars" and their rules, for the purposes of translation, rather than down the route of statistics to begin with.

Because as it was noted, this happened after WW2 and the advances in cryptography, which were by and large also driven by statistical analysis and other similar mathematically based advances; one would think that such insights would have carried over into the machine translation realm.

But they didn't - which is simply a curious historical footnote. It's also something we see often in history - particularly as it relates to technological advances; that there are certain paths that are taken that lead to virtual dead-ends decades later, but which the path that should have been taken was presaged by earlier work, yet for one reason or another wasn't pursued.

If we only had a way to avoid such "wrong turns", we could be much further along technologically - but of course, that would also be tantamount to "telling the future"...almost.

The downside of statistical models has always been understanding and explaining why they do whatever they do. Any statistical model will generate a result, either right or wrong, but it's impracticable to figure out how it came up with that result: the only knob you have is training data.

"Symbolic" techniques are the usual first choice because in order to get them to work reasonably well (i.e. beyond the 60%-80% threshold), you have to understand what they're doing.

"It's my understanding that machine translation started to become semi-useful after abandoning traditional grammars (or anything Chomsky would recognize) and switched to statistics."

The same thing happened in speech recognition. IBM started throwing statistical modeling and horsepower at the speech recognition problem, and folks like Kurzweil thought it was inelegant and not "true" speech recognition. However, it serves the purpose, and we know that humans use similar tricks when listening to augment pure recognition. It's harder to hear someone in a loud room if you can't see their mouth, or don't know the context of what they're saying.

I did mean pre-Chomsky. I was using him as a rough reference point for when linguistics started to get serious as a scientific discipline.

Because people have been informally studying linguistics for all of human history (by trying to learn new languages in adulthood), it's surprising to me that we ever thought that natural language processing would be easily solved by computers. To me, it seems quite obvious that this is a fundamentally very challenging problem for computers, like theorem proving or painting, rather than accounting or printing.

I was born with the benefit of hindsight here, but I find it strange that computability theory was formalised in the 1930s (before we could actually build powerful computers), and yet decades later, leading experts still failed to realise that language processing was fundamentally difficult, and intractable on the hardware they had at the time. This seems like an obvious consequence of a number of facts which they did know, like "describing an algorithm is very different to having a normal conversation," and "texts are open to interpretation and cannot be perfectly translated into other languages."

To be very honest you don't need Chomsky to realize how complicated languages are

People who think it's easy usually don't speak a 2nd language (amongst other problems in communicating with non-natives for example) so I'm going to bet on that.

Good point. I thought you were going to talk about Minsky and Papert and Perceptrons. That caused another mini-winter, at least w.r.t. connectionist models.

So true. The real danger with 'AI' at the moment is the premature deployment of half-baked parlour tricks or the over reliance on simple algorithms where there are real consequences. i.e. prison sentences and insurance eligibility and yes, autonomous vehicles.

> ...premature deployment of half-baked parlour tricks...

And the “fake it till you make it” attitude in SV only makes this worse.

You're always going to see #2. Use cases get marketable before they get fully solved so you'll always see pretty good but not great examples. In the mean time there are a lot of things that are solved by neural nets that we take for granted such as voice and face recognition.

Voice is solved? Then wider hall cunt my giggle nexus ever get thaddeus tribbleshits?

That's MISTER Thaddeus Tribbleshits to you.

AGI is not needed for self-driving.

None of what is mentioned above is a deal-breaker for self-driving car service:

- lidar at night works just fine

- plenty of cities with no or very little snow

- construction zones: blacklisting, remote monitoring & manual mapping, detection of cones, barriers, re-painted lanes

- self-driving cars with 360 degree view and plenty of patience and no distraction are safer for bicyclists than manually driven cars

If I just wanted to drive around a city I’d grab a smart car. I have taxis to take me home at night when I’m drunk and the busses have stopped running. This solves a problem only rich people have: traffic is tedious. The rest of us just take public transit and listen to podcasts or music to bypass it.

From my perspective the idea of a self driving taxi is the opposite of public transit in the worst possible way: pay more so fewer people have a job and you’re even more lonely and alienated than before. If I indulge I’m just making myself and everyone around me more miserable, and no amount of marketing will convince me otherwise.

The area where I’d appreciate this the most is the country where uber and lyft haven’t reached yet, and that’s precisely the environment where self driving cars will take the longest to reach. We’ll see.

Plus, I can’t wait to see the game of “beat the shit out of the corporate empire’s self driving car for fun when drunk”.

Self-driving cars may be safer than manually driven cars in certain contexts.

What's missing from AI is reliability. It's brittle, it works well in some situations and not at all in others.

Which is a problem, because you can never be sure how well it's working for you.

There's no anti-Dunning-Kruger-function to say "I'm sorry Dave, I can't handle this, you should take over" - partly because that's not something you want to experience driving into a hail storm at 70mph on a freeway if you're asleep, but also because it would require a level of domain-specific AI self-awareness that is barely on the radar in most domains.

Unfortunately any AI systems will encounter the verification dilemma, the more powerful an AI system becomes, the less verifiable it would be.

Verification is easy, just like with humans. It's called a driver's license. And since the AI does not tire and can probably be sped up, you can put it through hundreds of thousands of hours of driving quickly in a good simulator with adversarial and normal situations, then rate it at various tasks. Just like we should do with human drivers, but fail to.

Explanation is harder. But we probably shouldn't care, even in courts people cannot often explain what and why they did while driving, or they just lie. The thing is, for liability purposes you have to ensure it is not a series defect and that a good human driver would not be able to handle it

In the world of investments, that would be called buzzwordily as “back testing”.

But is it sufficient to answer _why_ the machine chose to act a certain way given a certain set of instantaneous input criteria?

No. But, as the parent post said, we don't always know why a human driver chose to act a certain way either. So that shouldn't be a blocker.

But we evolved to have empathy to help understand how humans act in the cases we don't have complete or good information.

This is why "crazy" people make us so uneasy. They don't fit our mental models for how a human should act. Would you be comfortable driving with a road full of unpredictable "crazies"?

I wonder if we'll ever have the same level of trust with AI as humans if it is still being used at a black box level.

Bad AI will suffer from the Dunning-Kruger effect and overestimate its abilities, while good AI will suffer from Imposter Syndrome and underestimate its abilities.

>- lidar at night works just fine at night

Yes, let’s rely on a single sensor.

> - plenty of cities with no or very little snow

But plenty of city have plenty of snow. So perhaps not a deal breaker in a limited set of circumstances.

> - construction zones: blacklisting, remote monitoring & manual mapping, detection of cones, barriers, re-painted lanes

Why should we add to already high construction costs just because the self driving tech isn’t up to snuff?

>- self-driving cars with 360 degree view and plenty of patience and no distraction are safer for bicyclists than manually driven cars

Citation?

You also have night vision and radar to detect metal objects. Plus sound which is still barely utilized besides Doppler.

Ultimately humans cannot match the senses of an automated car, the problem is in processing and integrating. Which we are pretty far away from solving in a general case.

I'm not so sure about that. For one thing, unless they've changed a lot more than I think they have since the last time I got to fool with one myself, the sensors have very limited range. Take a look at the screens displaying "what the car sees" in this self-promotion video from Waymo:

https://youtu.be/B8R148hFxPw?t=148

What the car is in essence doing is driving by GPS and maps while staring at its feet. It is not as bad as driving with your eyes closed, but it's not that much better.

Further, some of the infrastructure it is relying on, like lane markings, are very poorly maintained.

Why shouldn't surrounding infrastructure adapt a bit for self-driving cars? Look at everything that had to change for railroads, cars, and airplanes.

If railroads were invented today they would never be allowed. They can't stop for miles? You have to teach everyone everywhere to stay off the tracks? How's that supposed to work?

It's good that our safety standards are a lot higher today for new tech, but perhaps there are a few common sense rules that could be taught, rather than requiring perfection?

Honestly? Because I see self driving cars as expensive toys for rich people[1], and I don’t feel like spending tax dollars or making other concessions to subsidize that.

I think the analogy to railroads is a bit tenuous. There are immediate benefits to connecting two sufficiently separated points with a rail line. A single stretch of properly marked and instrumented roadway is relatively worthless[1].

[1] Trucking being a potential exception.

You do realize horseless carriages were also expensive toys for rich people when they were first introduced?

Yes, and the infrastructure we ended up building for them has made horrible changes to most cities and had extreme negative effects on the environment

> plenty of cities with no or very little snow

But what happens that rare day it snows? Hundreds of deads? Cars get recalled for much less than that.

I don't know if you live in such a city, but I can tell you that in mine the rare days it snows circulation is extremely chaotic, with people having slow paced crashes all over the place, and I'd wager AI is good enough to do that as well.

My understanding is that AI has different failure modes. If it doesn't see the road, it might not realise it doesn't see the road and just go off road. It's not necessarily going to react like a reasonable human.

I see your point, and it's indeed valid : I do remember a tesla crash where the autopilot simply followed an old line on the road instead of the regular one, thinking all was fine until it hit a column, killing the passenger.

But in this specific case of snow, I think this could be handled in a decent way by AI, as I would not expect snow conditions to be too hard to detect. Even something like "grip is not good enough for autopilot/snow detected on/around the road, stop on the side or slow down to the point of being impossible to have a fatal car crash until driver takes over" would be good enough.

There have been a bunch of prototype self driving cars driving in snow. I'm not sure it's fundamentally a much different problem to self driving in general. eg https://www.engadget.com/2018/05/08/waymo-snow-navigation/

Sure they may perform worse in snow but humans do too.

Drive it in manual mode like we do right now? I mean even an autonomous car that could handle 80% of normal driving just fine would be great. And by 80% I'm talking about driving on a freeway without crashing into a barrier in broad daylight.

Gonna suck for those people who don't own cars and are relying on the fleets of privately-owned pseudo-taxis.

A standardized QR code printed on cones to say "construction ahead do this" seems relatively simple compared to correctly interpreting the construction site. Would also produce automatically labeled training data

And then someone runs over the cone and 100 following cars crash into construciton workers at full speed. The core issue of automated driving isn't solving any of the 1000 simple driving tasks. It's creating an entire distributed system that is inherently robust against errors, failures or manipulation.

If navigating the world were that easy we wouldn't be (non-A)GI. The bulb on the end of our spine doesn't use 1/5th of our oxygen (or whatever huge amount it is) intake because it makes our foreheads look sexy.

An auto-auto (I'm using that term unabashedly for "self-driving cars", you can too) has to be able to perceive and understand that the car in the next lane with the mattress poorly tied down might, at any moment, suddenly become two large moving objects. And so on.

If we set it up right the whole fleet will learn from the experiences of each member, likely in near-realtime. More than just traffic conditions, this will help with things like emergency response.

(We should be making self-driving nerf golfcarts with a top speed of maybe 15kph. Duh! We could make those today and sell a million and then incrementally make f'ing sportscars and sh!t.)

That bulb is there because of politics, and because speak sounds sexy. There are many animals capable of navigating the world even better than us (most fly, for obvious reasons) that don't have anything near the size of our brains.

We don't employ orangutans as chauffeurs. ;-P

On second thought I bet you could make a decent auto-auto out of a dog's brain... https://villains.fandom.com/wiki/Rat_Things_(Snow_Crash)

These are made up numbers, but I use this to explain the scale to people. If it costs $1Bn to reach 80%, it'll take another $100Bn to reach 90%. There isn't enough money to reach 100% for two reasons. 1) You can't get to 100% and 2) It will take so long people lose interest. You can't sustain the hype that long.

Right. There has been a lot of progress within the last decade in narrow AI, but AGI remains as elusive as ever.

The best self driving systems do not have trouble dealing with bicyclists. The one time it happened with Uber was a combination of user errors starting with a very serious misconfiguration of the system.

Also, AGI and self-driving cars have almost nothing in common. Driving cars is a very narrow AI task.

What GOFAI technique got even close to 80% of the way? Maybe pathfinding, if I'm being particularly generous?

These days it’s more 95% easy, last 5% hard.

>> We're not going to have AGI any time soon.

In your mind, how many inventions are we from GAI? 1? 1000?

Also, given we have our very brightest working in the problem space (MS, Google, FB, OpenAI ...) how long would each remaining invention take?

It seems plausible to me that we are only one invention from GAI, one novel combination of two or more existing solutions and that this may come from anyone and at any time.

Applying this method of thinking, is there anything that we aren't one invention away from?

I'm just trying to stay positive.

If you talk optimistically about AI (or big data for that matter) you quickly get downvoted these days. I'm curious as to why that is. I thought this was 'marica. Isn't this 'marica?

Edit: my apologies, that was condescending. I should have finished of with a wink of an eye.

There's optimism and there's being too vague to disprove. You might be able to dodge downvotes by adding more objective content to your comments.

History has played out a particular way quite a few times (https://en.m.wikipedia.org/wiki/AI_winter). If you can tell us why today isn't like all the other days using facts, I'll bet the downvotes will stop.

Large scale colonization of Mars would require multiple inventions. We know this because we have a good understanding of what a successful colonization of Mars entails, and have a pretty good idea of the main technical obstacles. AGI, on the other hand, might be one breakthrough away or two hundred breakthroughs away. We just have no idea what makes intelligence tick.

We've drawn a couple of circles and we want to draw the rest of the owl, and you're telling us there's only one step.

Yes. Because we don't know it is an owl we should draw. It's equally likely AI is just three circles combined.

exactly... we don't even have a definition of consciousness that is better than Jacobellis v. Ohio - For that matter, I'm not sure we even have it that good. I'm not sure I'd recognize a non-human consciousness when I saw it. I'm not sure anyone would.

I mean, sure, I can totally see the idea that consciousness is maybe some emergent property of complex systems... in which case, sure, we could stumble upon it by accident. That is totally possible; we're creating complex self-replicating systems all over the place.

But as far as intentionally constructing the thing that will come after humanity? we can't do that until we know what, exactly, it is we are constructing, and as far as I can tell, we're pretty far away from any idea of exactly what that is.

Is consciousness a necessary component for AGI, though?

>Is consciousness a necessary component for AGI, though?

I always thought so. Like, without volition, it's just a really good natural language tool, and not a general intelligence.

What would an artificial general intelligence without consciousness look like to you?

A really efficient slave. Essentially it does not get creative and is ultimately predictable, even in failure modes. It would also have no subconscious drives to do unexpected things.

Capitalistic dream.

without creativity... I don't think it would make for a very effective slave in a world that is constantly changing.

Is creativity possible without consciousness? I don't pretend to be an expert.

Probably possible, yes. Subconscious drives are a thing. You could have sentience without consciousness. (The difference is that consciousness can change or modify drives and recognize its own actions on the world - its own agency. It is related to sentience and the line between the two is thin and blurry.) I'm not sure it is possible to be conscious without having been at some point sentient.

The idea was that such an AI would never question its priorities or change them, not that it wouldn't be able to solve novel problems.

(Which is the definition of intelligence, duh. And shows how far we're from that. What we have is glorified pattern matching algorithms and some basic symbolic logic and clustering.)

That sounds more like a Communist dream. Capitalism doesn’t reward really efficient slaves.

Asimov's Three Laws of Perfect Slavery (Forgive me Isaac!)

If such beings were conscious they would be saints.

One of my favorite AI jokes:

Q: What's AI?

A: When the machine wakes up and says, "What's in it for me?"

It would be fetish to actually make humanoid servitor robots. IMO.

Like the intellectual equivalent of a beast of burden - good at solving open problems by (apparently) intuiting connections and making inferences, probably with the appearance of creative operation and experimental learning, but lacking any genuine desire or personal agency.

IMO this is much more plausible as AGI than some kind of mythical general meta-intelligence that has a virtual soul.

I don't see why consciousness or agency are at all necessary for AGI - except as a science fiction trope.

>good at solving open problems by (apparently) intuiting connections and making inferences, probably with the appearance of creative operation and experimental learning, but lacking any genuine desire or personal agency.

so, like a philosophical zombie? like something that appears to be conscious but isn't?

Besides which, consciousness and agency are two different things, damnit! I've got a github repo in which I train agents by policy optimization. I have no illusions that my Pytorch models are conscious.

I mean, we could call it an "artificial general intelligence" or we could call it a "generalized optimizer & problem solver"

So basically if we can have a system that can solve any problem thrown at it, that's essentially the AGI.

I don't think it's necessary for us to get a machine to reach "consciousness", whatever that is, before we realise it's overtaken our mental ability to combine two existing concepts in order to create a new one.

I imagine this is what's going to happen: we humans have drawn two circles (thank you dear fella for that analogy) and we are about to draw a third, combing two existing concepts.

At some point we will have drawn so many circles that the next one will not be drawn by us. It will be drawn by AI.

Who knows when we'll reach that level of machine intelligence and if it even requires them, the machines, to be conscious? But we will, without any doubt, reach a point where there have been so many circles drawn that they are now smarter than us at drawing them. Absolutely without any sort of doubt in my mind this will happen.

Tens of thousands of people comprehend Einstein's work. It used to be hundreds. We're smart as shit these days. Fuck me if it's not before I die.

>Tens of thousands of people comprehend Einstein's work. It used to be hundreds. We're smart as shit these days.

In some ways, I feel like this is the real singularity. This trend towards feeding and offering education to everyone. Like, it used to be that to be an intellectual, you'd need to be born rich to get the free time and education. I mean, being born rich still helps, sure, but a lot more people are getting a shot.

On the other hand, it's a process that levels off at some point, like everything else. The population is leveling off; as we bring education and leisure time to more and more of the world, eventually most of the people who have the capability to do this sort of thing will have done it.

I mean, I see that with computers, too... like with the leveling out of Moore's law; exponential processes, in nature, tend to not remain exponential.

If you subscribe to the theory of great and continuing, accelerating change due to technological improvements, advanced technological education for everyone is an expected ovservation during the decades or centuries it’s occurring :)

I may not know an owl until I see it, but I do know that's not it.

General intelligence? Does it even exist? I find the evidence that humans have it unconvincing, and I've seen nothing that makes me think we're "one invention" away. The direction of machine learning seems to be specialised tasks on specialised data sets; surely the very opposite of any kind of general intelligence.

So get whatever intelligence human have in computers. That was the goal. Call it general intelligence or call it x.

Human intelligence works well [0] from inside human shaped objects with human shaped biology, in human shaped societies, and to no small degree for doing human-style tasks.

I am not convinced that putting that human intelligence, whatever it is, into a silicon box will give us the results we hope for (or is even a sensible ambition, given the degree to which our intelligence is part of our biology).

[0] Actually, maybe it doesn't work really well - maybe it's a slow-motion car crash over millions of years. But it works well enough that we clearly want more of it.

Whatever human intelligence is, there's clearly some part of it that can do things like design computers and build houses. That part, at least, people would like to be able to match in performance using machines. It's 'intelligence' in the sense of being able to do things like think up clever strategies in novel environments that people want to put in a silicon box.

If that is the goal, it's certainly not what the current batch of machine-learning and hype-train riders are chasing. They (or at least, the ones doing it via huge training sets) seem to be working on the very opposite of novel environment situations. I wonder if anyone is seriously pursuing that.

That's exactly what they thought in the 1950s and 1960s 1970s and 1980s. And oh by the way, our brightest minds were working on it then too.

Reply on news.ycombinator.com