The new distbelief paper doesn't explain some things. Why are "receptive fields". An earlier paper claimed better than 15% results on imagenet and so does this, what changed. Why have you claimed sgd works on nonconvex well without comparing against actual nonconvex opto algs. You also didn't cite my comments on metaoptimize where some of your ideas seem to come from.
Yeah, it's such a simple thing that nerds don't seem to get. Put out ads on craigslist, adwords, indiegogo. Whatever sells, do that. The biggest problem is having more ideas than time to write them into an ad, there's too many potential products. You can outsource the creation of whatever sells so executing isn't even that big a hurdle (though you do have to be careful with initial pricing on the ad so that costs don't get you).
Telecommuting high def video. Upstreaming dozens of security cam/robot videos from your house. Caching data. Google can cache much of youtube and web right on your internet router so that you pay the electricity bill instead of them. Telemedicine.
As much as I like google, I doubt they're as good as that. Their search and advertising is essentially a monopoly because they have such a data lead over anyone else. That's why facebook freaks them out because they know how having more data is the secret sauce. Their data center innovations don't seem all that incredibly innovative. If they had robots whirring around I'd be much more impressed.
I mean in their data center. Also, I'd be more impressed if they actually got the self-driving cars out to the world. It's been 5 years now. Google has a problem executing, they're lulled by their one success.
If you live in the Bay Area you see the automated cars driving Google Employees to work all the time. I used to get an SMS any time somebody saw one (they are very distinctive) - but not as often, because people are starting to get more used to them. Though, they are still pretty cool to run across.
I'm impressed because people have been trying to nail this problem for decades. Try convincing people that a few million automated half-ton canisters of metal going 60 miles an hour are safe. An early success will win investors or make a nice press release, but developing a rigorous, safe product of that complexity and risk is a long slog. It's got to be so well understood it's boring - then it's ready for massive public use.
DHH needs to get off his high horse. He's hacked his way into providing a tiny service for artsy folks but he's not exactly hacking on immortality tech or AI like some of us. The government needs to consider redistributing some of his wealth. This essay is him gloating about sitting on piles of cash, where's the concern for starving african children or terminal cancer patients. He should sell his company and work on immortality tech. He can start with extracorporeal devices like described here http://www.youtube.com/watch?v=JuBDhT7Y5GU
marshallp needs to get off his pony. DHH's is not gloating about the piles of cash he surely has on this essay. He's talking about the unprecedented and amazing opportunities unregulated internet gives businesses.
And about your career counseling, I'd rather have DHH building niche web software with nice UI than have him build life supporting medical equipment.
Why should he keep building nice UI's. He's proven himself as a can-do hacker and has millions in the bank. He could spend his resources saving lives instead of polishing what's a tiny hill in the global landscape of ideas.
You could propose him a venture on life supporting devices and see what he tells you.
There's a (tiny) chance he hasn't gotten into life supporting devices because the opportunity hasn't shown itself. On the other hand it might just be that this area is too far out of his domain of expertise or he's just not interested in it. Maybe he doesn't feel like immortality is a worthy goal to pursue. Maybe he donates thousands monthly to whatever charity he feels truly help people. Maybe he drinks all that money with friends after race driving.
> Maybe he drinks all that money with friends after race driving.
> Maybe he donates thousands...
Damned good show if he does. The world needs more young, selfish playboy millionaires who make absolutely no public show of being philanthropic. Leave that stuff to the billionaires. (No irony or sarcasm intended; I love to see the rich having fun, as long as it's not to anyone else's detriment.) Matthew said it well in the Bible: "Take heed that you give not your alms before men, to be seen of them." He didn't have much to say on car racing, but even Jesus liked a boozy wedding.
> Maybe he doesn't feel like immortality is a worthy goal to pursue.
It's usually only the most awful people who do think it's worth pursuing. We'd all like a few more years, but it takes a special sort of narcissist to think their presence will be appreciated after 200, 500, 1,000, 1x10^6 years...
What if it's a relative you'd like to see live another 10 or 20 years because they're too young to die?
Yeah, we need more playboys. The kid with cancer or malnutrition and his family should just learn to cope with death.
Anti-death people are such narcissists. The financiers like you are the true heros. You get a lot of hookers and blow and we should be thankful for you consuming them.
That's why I said government maybe needs to redistribute wealth if these rich folk genuinely don't give a shit about poor people. They can keep their money if they invest in society improving stuff. That's why I said immortality tech. He could get involved in miniaturizing extra-corporeal devices (artificial organs) that keep people alive even if their hearts or lungs get too damaged. He can start by watching that video and contacting those people and putting together a team of engineers to get it done. If you're not doing an Elon Musk, then maybe the government needs to raise your taxes.
Ignoring all the problems like: who gets to decide; how are the decisions made; how often is success measured; what corrective measures are allowed; what is the best way to help the poor; what is the best way to redistribute the wealth; how to prevent lobbies from destroying the plan; how to avoid money escaping the countries implementing this into tax heavens.
Ignoring all that (and many more questions of the sort) I'm left with only one question: how exactly is Musk helping "the poor"?
He's helping the poor by making energy cheaper indirectly through kickstarting an electric car industry.
The tax issue can be resolved by luxury taxes. If people decide they'd like $100 million yacht, tax them $200 million for it. Countries like Singapore already have punitive taxes like that on cars.
How the guy decides to spend his own money that he earned is none of your business. Mind your own, please.
If you think it's such a travesty that the guy has money and enjoys it, then maybe you should go to all the people who consensually traded with him and tell them what bad people they are for buying his products and services.
Like I said earlier, a luxury tax. The capitalist utopia of singapore has them. Tax the rich and give it to NASA or create multi-billion dollar seed fund for immortality hacktivists.
AFAICT taxes in Singapore only create a welfare state that relieves somewhat poverty. A noble and worthy goal, but there are still poor people in Singapore. What about cancer cures and immortailty research, are they big there?
And now for the things you haven't addressed. Why is the government the most efficient way to solve the problem? Why cancer or immortality or space exploration are more important than food or malaria or sex education? Just because you like them?
I think you should stop and think real deep about the questions I've stated and implied in this thread, and answer them in length, with references where appropriate. Then share them.
Else all you end up doing is coming across as a naive angry fanatic with not too much reading comprehension skill (remember all this started because you misinterpreted DHH's essay) who doesn't even get to see the problem and is boasting about solutions.
Airplanes should be fully automated, they probably already have the technology to do this. They could start with cargo planes and then move to passenger once the public fears are quelled.
No, they really shouldn't. Autopilots are very good at handling routine, monotonous tasks and very bad at handling unusual or unexpected tasks. See, for example, Air Canada 143, British Airways 38, Air France 447 (admittedly, the crew human crew didn't handle that one so well, but only after the automated system completely gave up), US Air 1549. The BA 38 flight in particular highlights why you really, really want a human crew on board. The pilots of that flight had to over-ride the autopilot, and turned what would have been a complete disaster with everyone killed into a mere total write-off with less than 50 injuries and no fatalities.
What impressed me most about that entry is how they managed to trace it back to the root cause (besides, obviously managing to do as well as they did).
What were the things an autopilot couldn't handle. If more engineering effort were spent, couldn't those also be automatically handled. Plus, couldn't airplanes be remote controlled.
> What were the things an autopilot couldn't handle.
Anything unforeseen. Aviation is still a craft, not a science even if a very large number (a huge number, actually) of hours have gone into making it a science. The basic problem here is that you're dealing with chaotic natural systems on one end and machinery at the other. All it takes is a flight of geese or some unusual weather pattern and all your careful programming is worth zip. In situations like that humans tend to do better than computers.
That's because the current software intentionally uses very simple cookbook algorithms for flying the plane. Proper software would simulate thousands of scenarios when an anomaly occurs and pick one that gives the least bad result.
The BA 38 example highlights why we need to get humans out of the loop. Good software would have instantly detected the uncommanded loss of power, and immediately started raising the flaps and landing gear. In fact, good software would probably have exercised the control surfaces and engines before descent and diagnosed the problem then, when the aircraft had tens of kilometers of glide remaining.
Would you, by any chance, share any links regarding "proper software would simulate thousands of scenarios"? Genuinely interested, how are these kinds of problems solved in real life.
I was thinking of automated trial and error using a simulator. Try a bunch of variations in the simulator and pick the one with the best simulated results.
Approaches like that are common for playing board games like chess, for finding clever transfer orbits for spacecraft, etc.
Really? Because BA 38 shows to me why software has a long way to go before it can replace pilots.
So you want to detect a fail to spool? OK you need to put a sensor on the engine to detect speed. And one on the throttle to detect position. Maybe there is another way to detect fail to spool but this is what we're going with because linear and rotational sensors are pretty common and well sorted.
Now you need to make sure that those sensors are 100% correct before factoring it into your code. So you add another of each sensor. But what if one of those sensors is out of calibration or broken? You need another one to vote for majority. Ok so now you've got 3 sensors on both the throttle and engine shaft.
Nothing really crazy here and pretty common.
Now you've got to build the software to model the engine. Jet engines have a delay in spool times that depends on air speed, how fast you feed in the throttle, how fast the engine is already turning, and other factors I have no clue about.
Now that you've got the model setup you have to add in some hysteresis to keep things nominal and to deal with glitches in readings due to iPods and cosmic radiation and vibration.
And then you have to remodel the whole thing depending on which engine is fitted (there are 3 available for the 777 plus variations). And then you've got to test it.
Now that you've tested that you are able to detect a fail to spool situation you have to generate the model of what to do in various flight modes. And that depends on how much thrust was lost, air speed, altitude, attitude, rate of decent, wing configuration, weight, landing gear configuration, etc.
And then you have to test the flight model. And then you have to update the operational handbook to inform the pilot of what to do if the above fails to work. You also have to amend the inspection and maintenance books to inform the A&P how to test and verify the sensors. You also have to modify the flight recorder to record the sensor positions and software behavior. And you have to modify the sat recording and playback software to handle the new data.
You have to do all of the above because if the fail to spool code activates on a normal approach it can very easily lead to a stall and kill 200+ people. You want to be 99.99999% sure your code is correct.
You've now spent millions of dollars and years of development to replace something a pilot can detect in 3 seconds by simply listening and feeling to what the plane is doing. AND you have created a liability if a bug crops up.
Loss of power can be detected by measuring combustor heat output inside the engine, which you need to do anyway to detect flame out and to adjust the engine for best efficiency.
The control software does not need to know about the different models of engine. Each engine should measure its own performance and report the results to the self-fly software.
In the case of BA 38, this was done: the aircraft automatically detected the loss of power. Unfortunately the (dumb) autopilot had used a flight profile that was unsalvageable if loss of power had occurred at that point.
The future of travel is self-controlling vehicles. We have usable self-driving cars today. Airplanes are easier to control than cars, and the aviation market has a huge pent up demand for short hop travel that cannot be met by the heavily crewed business model.
I'm pretty sure that it will be a long time before we have technology that can land a plane safely in the Hudson River after it loses all power due to multiple bird collisions.
And we'd need to be able to do this even for a cargo plane with no passengers on board, since a disabled plane crashing into a populated area would cause a lot of deaths.
Airplanes are substantially more complicated machines than cars, and are also pushed much further in their performance envelope than cars.
Cars also enjoy much safer failure modes than aircraft - if a catastrophic failure occurs in all likelihood you can pull the car over to a stop without issue. When a catastrophic failure occurs in an aircraft's controls it hits the ground and everybody dies.
We're not even quite there yet for self-driving cars, though we are close - and that is a problem orders of magnitude easier than fully-competent autopilots.
We already have the technology to take off, fly, and land planes in the Happy Case. But the devil is really in the unhappy cases.
Nobody is saying we won't ever have full autopilots - but that we are very, very far away from such a capability, and simply throwing engineers at the problem is unlikely to give you the solution substantially faster. When you're pushing the absolute edges of scientific and engineering knowledge progress does not correlate strongly with workforce.
The big difference between cars and planes is that a near universal failsafe in a car is to slam on the brakes. Car stops. Simple and relatively foolproof. There is no such avenue available to a plane in flight.
If you still think it's possible take some time to learn of all the novel failure modes that have occured over the past few decades in commercial aviation. Pay special attention to the ones requiring crew to override onboard computers.
I don't really follow developments in the robocar world, but I would have thought the passenger would still be able to tell the car to stop for a hitcher.
- focus on data, like google does, by massively investing in siri (they should make forstall in charge)
- focus on design, but expand out of consumer electronics (which will shrink) into cars and robots, like tesla and rethink robotics (tim cook and jony ive are the right people for that)
Their current niche, square screens and possibly google glass is a dead end
Airbnb should go all the way and start renting out trailers in the desert. Attach gigabit fiber as well, big opportunity in that. With self-driving cars and self-flying drones on the horizon there'd could be a first mover advantage for a short while.
If she's a prof, then she's engaging in the scam that is the current university system. Is she or the other profs you know advocating for the abolition of undergraduate degrees to be replaced by online learning like coursera or udacity. Also, abolishing phds/tenure and replacing with funding their research from kickstarter/indiegogo. Probably not. They might care about their few chosen students (selfishly, for the social company and labor it provides) but they're putting their finger up at the entire rest of the world (which includes potential students that never were, like in africa or india).
