do you really think that about deep learning? it has zero application or effect in my job (i don't see that changing for at least the next decade) and in my personal life, it only serves to produce "better ads".
That's not to disparage with your comment, only to say that while it might feel like a negligible effect on your life, I would say that it is non-zero.
My completely uninformed and lay persons opinion is that we might not yet know what we can really do with deep learning yet. For example there were many years between the discovery of electricity and the invention of the microprocessor. The latter certainly required the discovery of the former, but did not logically lead to it. I use this example because Andrew Ng called AI the new electricity.
Web page translations (at least in Google and earlier) didn't do "deep learning" for ages.
Plus most of those things, web page translation, image search, speech assistants, are pretty much either niches or gimmicks. Even if they worked perfectly, there would not be much to write home about.
It's not like people think "my life was changed by the Alexa". GUIs, on the other hand, made lots of things possible that weren't, including whole new professions (heck, the web IS a GUI thing).
That varies a lot by language. I keep finding its attempts at Polish painfully bad, and I only know a little Polish myself.
Yes, absolutely. To put it bluntly: how many jobs do you think are affected by being deaf and blind?
Computers are now able to see and hear, and on top of that they have become better at most classification tasks than humans. Those are game changers, this goes much further than advertising.
When I hear claims on how many jobs will be replaced because of deep learning (in contrast to the base rate of the long, steady slope of automation over the last 175 years), it's often about an abstract and simplified notion of the complexities of jobs outside the claimant's area of expertise. In such cases, it's useful to ask the claimant whether their profession is soon to be made obsolete. Will AI replace the pundit anytime soon? The economist? The enterprise programmer? The AI researcher?
Why not. You can go to any major news outlet, and expect a certain bias/comfort in their news, and a definite bias/comfort in their overall editorials or those of individual editorialists. Why, it's almost like they're programmed to reliably write under those biases.
We already have the beginnings of auto-written news articles. There was a system featured on HN a few years ago about weather news articles at, I think, the LA Times.
One reason some people stay logged out of Google is so their search results aren't colored by what Google thinks it knows about them. It's not that large a leap to systems reading all the opinion letters its owners receive, and maybe all the printed letters that other outlets receive, turn a knob or two and collect max-Laffer curve revenue, based on where it's set on the bias curve.
Let's review this in a decade or so and see what happened.
In terms of level 5 autonomous driving, the most common retort I've seen recently is "well, we're going to need special roads with sensors built in, and then it is a tractable problem." There are more than 4 million miles of road in the US; at any given time maybe 2% are under construction. The US Interstate Highway System was built during a boom and the cold war (where the system learned from Eisenhower's WWII experiences and learning from the Autobahn's military utility.) [1] If we're lucky, maybe the interstates will get sensors. But the complexity of re-doing an exit or interchange, which involves dozens of on-the-fly lane changes and cutovers, seems impractical merely to support self-driving cars.
I would love autonomous vehicles for the sake of my children, but I suspect that at best, maybe my grandchildren will benefit.
[1] https://www.fhwa.dot.gov/interstate/brainiacs/eisenhowerinte...
Nobody involved closely with self driving cars thinks we're anywhere close to a level 5 system. Why do you think we are?
That point in time has definitely come for self driving cars, we're not 'close' but we are much closer than we were even 5 years ago, there is legislation on the books permitting self driving cars from participating in regular traffic instead of on private property only, there is special hardware geared towards such systems and so on.
It's an idea whose time has come, now we have a bunch of engineering problems to solve and they're far from simple but I would not at all be surprised if self driving cars are going to be equal partners in traffic 10 years from now.
Lots of hard work between now but I think we'll see self driving cars long before we will see an operational fusion reactor.
I think it's obvious that machines can, for example, weld better than humans can. But machines can't decide what to weld. For now in that arena, things are mostly still human controlled.
I think your argument gets significantly weaker when you push truck, car, and plane driving onto the conversation stack.
Even the best of self-driving AI automation isn't nearly as good as bad human driving. Yet.
But the absolute best human weld joints are no better than the absolute best robotic weld joints.
I think you're talking about two entirely different domains here when you conflate welders and couriers. One of those needs a human decision to place the components to be welded next it in front of a robot that basically does one thing. The other is an actual adaptive AI that's trying to make decisions on it's own.
Yes, it's obvious that the lowest-wage and lowest-skilled workers are going to be displaced by robots the soonest. That's like arguing that the sun is going to rise. And as a society, we need to account for that.
But arguing that the lowest hanging fruit of automation is going to displace taxi drivers and gardeners in the next 10 years is pretty nuts, even for HN.
I feel like this fits the description in the comment above:
> When I hear claims on how many jobs will be replaced because of deep learning (in contrast to the base rate of the long, steady slope of automation over the last 175 years), it's often about an abstract and simplified notion of the complexities of jobs outside the claimant's area of expertise.
But were you optimize the Boeing manufacturing process for machines, the way car factories are optimized, you could utilize strengths of machines while sidestepping their limitations.
(Why this does not happen often is I believe a combination of robots having high initial costs, coupled with products starting as human-assembled prototypes, and companies just optimizing that process incrementally, instead of redesigning it around machines.)
> I think it's obvious that machines can, for example, weld better than humans can.
Ok. That wasn't all that obvious not all that long ago.
> But machines can't decide what to weld. For now in that arena, things are mostly still human controlled.
Yes.
> I think your argument gets significantly weaker when you push truck, car, and plane driving onto the conversation stack.
What's so incredibly special about truck, car or plane driving/piloting that you feel they are immune to automation?
> Even the best of self-driving AI automation isn't nearly as good as bad human driving. Yet.
Precisely. Let's give it 10 years and look again. I wouldn't bet against it.
> But the absolute best human weld joints are no better than the absolute best robotic weld joints.
It's not about 'best'. It is all about repeat accuracy and consistency. That's what makes the robot better. Maybe the best human welds are better than the best robot welds. But to get a human to consistently outperform the robot or to consistently deliver a specific quality level is not going to happen.
> I think you're talking about two entirely different domains here when you conflate welders and couriers.
We consider welding to be a highly skilled job and yet welding robots are a thing. Courier on the other hand 'merely' requires a driving license, something we hand out to 16 year olds after a perfunctory inspection.
> One of those needs a human decision to place the components to be welded next it in front of a robot that basically does one thing.
Which takes a couple of years training for a human.
> The other is an actual adaptive AI that's trying to make decisions on it's own.
Yes, it's a harder problem for a computer. But it was mostly harder because in the past computers could not see the way we can. But deeply layered convolutional neural networks and their descendants have changed that, dramatically. What was SF 5 years ago is now commonplace. And computers now have access to senses that we don't have, such as radar and lidar.
That takes care of one of the harder parts of the problem.
There is still plenty left, but one of the most formidable obstacles to self driving vehicles is now behind us.
> Yes, it's obvious that the lowest-wage and lowest-skilled workers are going to be displaced by robots the soonest. That's like arguing that the sun is going to rise. And as a society, we need to account for that.
Yes. But I wonder if we're going to be ready when the software is ready. In fact I doubt we will be ready. It will be the industrial revolution all over again, only this time it will play out in a decade instead of in 100 years.
> But arguing that the lowest hanging fruit of automation is going to displace taxi drivers and gardeners in the next 10 years is pretty nuts, even for HN.
I don't think that was called for, that's pretty rude, even for HN.
That hasn't been my experience. But perhaps we should define better?
Computers are faster at doing the classification than a human, but not as accurate. So it depends what variable you're optimizing for when you say "better."
I'd be careful with that.
Like autonomous cars or security cameras or robotics.
We're not yet at a 'general vision is a solved problem' level and this may take a long time still - if it is ever solved. But many problems can be reduced to the point that they become tractable even absent 'general vision'. Affordable LIDAR is a huge step, so is radar. Those two reduce the problem of vision with two crappy stereoscopic cameras set 5" apart to something much closer to the domain.
As far as I can see the race is on.
Our GPU kit has got 6 times better in 2 years.
Exciting times.
I started from the dataset bias position myself - billions of family snaps and selfies can't provide the reference for arbitrary images of stuff from arbitrary angles. My reading of the results that we got is that the claims of lower level features extracted by the cnn are correct, and that a network trained on massive public data can be the basis for specialised tools trained on more constrained proprietary datasets.
Your milage may vary though...
Garbage in, garbage out. When was it any different?
The question is: can a human do much better on that garbage?
> Public Datasets have style of photo that represent it's own bias.
That I readily agree with.
Scoring is also critical, you want to gracefully degrade classification systems not getting the correct dog breed is insignificant vs calling a dog a sofa. Further, people base real world classification on stereoscopic video footage. Training robots to classify photos is a handicap for building robots to operate in the real world.
Different tasks have different error tolerances. When an MI tool can get within the acceptable margin of error for certain tasks, then there is a possibility for the MI to take jobs.
As long as accuracy is critical to a classifier, in many cases it's better to pay for human eyeballs to look at the criteria.
We're talking about a huge spectrum of applications here. Most of us probably don't care that much about the difference between human and MI error rates about classifying images used in broad-scale advertising on the internet.
We do (or perhaps should) care very deeply about human vs. MI error rate in terms of granting or not-granting things like parole.
Add non-standard background noise, and it also fails on hearing tasks.
That was exactly true of Xerox PARC in the early '70s.