|
Owlofcreamcheese posted:The two things you have mentioned is that you are a factory worker and work in food safety. Why should I accept either of those things make you know anything about drones? Why do you think talking about people's jobs is a normal way to interact on internet forums at all? Solkanar512 posted:The only person who's employer is known is mine. Quit loving lying. If I see either of you quote each other's posts again I will ban you. This pseudo doxxing poo poo is creepy and at the next inkling of anything approaching it I'm going to ban the offender and gas this thread.
|
#
?
Jan 15, 2019 23:32
|
|
|
|
| # ? May 8, 2024 12:17 |
|
|
luxury handset posted:as i've repeatedly said, because any sort of a passenger scale quadcopter is inherently going to be lighter than a passenger scale helicopter, so any advantages you get in automatic response could be easily outweighed by, uh, weight Well if you want to compare something compare a single rotor RC copter to a multi-rotor copter, which one is lighter and which one does better in wind? I'll give you a hint, it's not the heavier one.
|
|
#
?
Jan 15, 2019 23:33
|
|
|
ElCondemn posted:Well if you want to compare something compare a single rotor RC copter to a multi-rotor copter, which one is lighter and which one does better in wind? I'll give you a hint, it's not the heavier one. nobody cares if a toy crashes in high wind though everyone cares quite a lot if something with people in it crashes to the ground, which is why i'm trying to talk about large multirotors and not toy single rotors. but since nobody has yet upscaled a multirotor to economically carry a person, i have to discuss it using theoretical examples of smaller designs and their flight characteristics. but at this point i feel like i'm explaining too much of my own argument to you
|
|
#
?
Jan 15, 2019 23:42
|
|
|
luxury handset posted:nobody cares if a toy crashes in high wind though I didn't want to make the comparison, you're the one that insisted on making the comparison. I'm just trying to explain why it's technically reasonable to think these things could see actual use in the real world. luxury handset posted:everyone cares quite a lot if something with people in it crashes to the ground, which is why i'm trying to talk about large multirotors and not toy single rotors. but since nobody has yet upscaled a multirotor to economically carry a person, i have to discuss it using theoretical examples of smaller designs and their flight characteristics. but at this point i feel like i'm explaining too much of my own argument to you Well, I would suggest you do some google searching then. There are plenty of passenger sized multi-rotor aircraft companies out there with working prototypes and thousands of hours of flight time. Not saying anything is settled but they're well past the "will this crash into the ground and kill everyone" point. ElCondemn fucked around with this message at 23:48 on Jan 15, 2019 |
|
#
?
Jan 15, 2019 23:46
|
|
|
we are going in way too many circles if i have to point out the difference between technically feasible and economically practical again. never mind, it's almost dogmatic
|
|
#
?
Jan 15, 2019 23:50
|
|
|
Edit this is the dumbest thread
|
|
#
?
Jan 15, 2019 23:50
|
|
|
as i pointed out in my first post on this subject three pages ago, which we have now returned to - having a working prototype means very little for continued feasibility. it's just a milestone, not a demonstration the technology is useful
|
|
#
?
Jan 15, 2019 23:58
|
|
|
luxury handset posted:we are going in way too many circles if i have to point out the difference between technically feasible and economically practical again. never mind, it's almost dogmatic So if that was your point why did you get into a technical discussion? BENGHAZI 2 posted:Edit this is the dumbest thread This helps! luxury handset posted:as i pointed out in my first post on this subject three pages ago, which we have now returned to - having a working prototype means very little for continued feasibility. it's just a milestone, not a demonstration the technology is useful Seems like you don't understand the difference between proofs of concepts and prototypes!
|
|
#
?
Jan 16, 2019 00:00
|
|
|
ElCondemn posted:So if that was your point why did you get into a technical discussion? the obvious impracticality of the technology was not obvious to you, so i attempted to explain it using technical terms. i was not successful regardless, the world is full of inventions which have no clear use or economic purpose. this is one of the constant truths of human invention, we like to create things just because only for them to have no utility -especially personal flying devices and other forms of transportation https://www.youtube.com/watch?v=ONwe96StEpA&t=2002s
|
|
#
?
Jan 16, 2019 00:03
|
|
|
luxury handset posted:the obvious impracticality of the technology was not obvious to you, so i attempted to explain it using technical terms. i was not successful Oh come on, you could have at least posted the idlewords talk. Feasibility vs. practicality: "4 lawnmower engines strapped to the edge of a marine was dumb in the 1960s" isn't informative about anything except whether quadcopters were practical vehicles in the 1960s. We have better hardware and manufacturing now, maybe good enough to make cheap reliable flying taxis, maybe not, but looking at early cold war toys doesn't contribute much to the discussion.
|
|
#
?
Jan 16, 2019 00:12
|
|
I CANNOT EJACULATE WITHOUT SEEING NATIVE AMERICANS BRUTALISED!
|
Main Paineframe posted:I don't understand why you're talking about point-to-point transport through the air as if it's a new technology. Helicopter taxi services have been around for literally half a century and aren't even all that expensive. The limiting factor isn't that you need to pay a pilot - it's the fact that in the urban areas where transport demand is highest, alternative forms of transportation already exist that can get you far closer to your destination for a much cheaper price. Replacing the pilot with an AI doesn't change those fundamental economics. And no, the primary limiting factor in helipad placement isn't the size of the helicopter. Had to look that one up. Yikes! https://learning.blogs.nytimes.com/2012/05/16/may-16-1977-helicopter-accident-on-top-of-pan-am-building-kills-five/ Didn't realize rooftop helipads were banned in NY after that, and the one exception is going to be... Amazon.
|
|
#
?
Jan 16, 2019 00:12
|
|
|
ElCondemn posted:So if that was your point why did you get into a technical discussion? I'm sorry I decided a post was going to contribute to a bunch of bullshit and edited it out, I'll check with you next time you bootlicker
|
|
#
?
Jan 16, 2019 00:19
|
|
|
suck my woke dick posted:We have better hardware and manufacturing now, maybe good enough to make cheap reliable flying taxis, maybe not, but looking at early cold war toys doesn't contribute much to the discussion. well the guy is saying "but there's a prototype!" which really doesn't mean much. the first time we had that argument i posted a bunch of pictures of prototypes that never went anywhere the quadcoptor is a pretty simple idea that is feasible on a small scale, but there are real issues around the power output of electric engines, battery size and weight, etc. that aren't going to be mitigated by real time computer control of the rotors but if all that one has is faith, then faith must be enough i guess personally i think that quadrotors are old enough to where if they were feasible to implement on a human-size scale, they would have been implemented on a human size scale already. looking at startups to sort of innovate the idea into fruition is some tech booster nonsense, if the us military and various other legacy flight firms passed on the idea then there's probably a good reason. the idea that small, innovative firms are coming up with the best hardware is one that's generally too adjacent to cold fusion and "one weird trick big oil doesn't want you to know" for my taste 
|
|
#
?
Jan 16, 2019 00:19
|
|
|
Main Paineframe posted:I don't understand why you're talking about point-to-point transport through the air as if it's a new technology. Helicopter taxi services have been around for literally half a century and aren't even all that expensive. The limiting factor isn't that you need to pay a pilot - it's the fact that in the urban areas where transport demand is highest, alternative forms of transportation already exist that can get you far closer to your destination for a much cheaper price. Replacing the pilot with an AI doesn't change those fundamental economics. And no, the primary limiting factor in helipad placement isn't the size of the helicopter. Changed my mind, I'm in favor of it now
|
|
#
?
Jan 16, 2019 00:20
|
|
|
haha the flight time on the airbus design is 15 minutes yikes! luxury handset posted:when you start looking into dead end technolgies and other discarded ideas, there's just a giant heap of personal flying vehicles and jetpacks and all kinds of that nonsense. when large companies design these things they're usually just doing it as a concept exercise to create buzz, in airbus' case, or in the case of uber who had their own silly flying car concept it's to keep investor storytime perpetuating. journalists are hungry for optimistic sci-tech news for geeks to drool over and generate clicks, development on these things have been going on for decades and how many flying cars are there in the skies? in the absence of any recent breakthroughs in battery technology the entire idea is still wildly impractical Mr. Fall Down Terror fucked around with this message at 00:30 on Jan 16, 2019 |
|
#
?
Jan 16, 2019 00:23
|
|
|
luxury handset posted:nobody cares if a toy crashes in high wind though I don't know what you're talking about here, there are a wide variety of multirotor designs out there, and all sorts of toy drones have been made out of those as well. It's not a mystery as to how many of the designs would perform but there is very good reason that passenger designs stick to just one or two rotors for the most part - there's not much real benefit to be gained by having to maintain all the extra rotors at scales suited for passenger and freight carrying. It's not like people haven't tried it, sticking to just one or two rotors tends to be much more efficient. Failson posted:Had to look that one up. Yikes! https://learning.blogs.nytimes.com/2012/05/16/may-16-1977-helicopter-accident-on-top-of-pan-am-building-kills-five/ It's a bit more complicated than them being banned. All sorts of buildings have and continue to be built with helipads in the city, but they're restricted almost entirely to emergency uses. Buildings up by the rivers or otherwise with a wide amount of clearance available are permitted to have frequent non-emergency helipad usage, but few bother to meet the standards so they're restricted to one off usage (which tends to be fine). Even the Amazon helipad will be for a building up against one of the rivers, and is supposed to be used for no more than 10 landings per month, roughly equivalent to the restrictions most other rooftop pads that are still used are working under. luxury handset posted:haha the flight time on the airbus design is 15 minutes There's also no particular reason that the copters should even be electric to be honest, so I don't know why these latest plans and stans keep insisting on electric.
|
|
#
?
Jan 16, 2019 00:38
|
|
|
luxury handset posted:haha the flight time on the airbus design is 15 minutes So basically a quick 10km hop, attendant swaps out the battery pack kind of deal.
|
|
#
?
Jan 16, 2019 00:38
|
|
|
Owlofcreamcheese posted:You bring it up a lot. It's really weird. Helicopters can operate in class B and C airspace because they are equipped with the necessary equipment and operate in accordance with a clearance to do so. That system cannot handle automation, nor can it practically handle an order of magnitude increase in the amount of air traffic in those areas. It would require a wholesale re-write of a huge number of regulations, a massive improvement to the technology involved in air traffic control across the board (both on ATC's end and in every aircraft currently flying), and it still doesn't address the other practical concerns regarding air taxis including cost, maintenance, safety regulations, weather minimums, and the lack of suitable legal landing areas. I mean, if your argument is that UberAir is going to operate like current helicopter charter services with similar restrictions, I guess I concede you're technically correct, but that still involves fairly limited utility overall and a great deal of expense. It's hardly going to be something readily available to the middle class on a regular basis. Nothing about quad-copter designs changes that, and at that point the expense of a pilot versus having a fully autonomous system is practically a rounding error in the overall economics of the situation.
|
|
#
?
Jan 16, 2019 02:16
|
|
|
fishmech posted:There's also no particular reason that the copters should even be electric to be honest, so I don't know why these latest plans and stans keep insisting on electric. when you have more than a few rotors involved that means more than a few motors and combustion engines are more complicated to operate, keep powered etc. than electric ones. so if you're going to do lift with a bunch of little rotors then it makes sense to have a bunch of little engines, since it's easier to power them all with wires than individual fuel lines or drive shafts. but if you're going to have one big rotor then it's better to have one big engine, and when it comes to big engines with big power combustion is better than electric electric is superior to combustion for small designs. but combustion is clearly better for big, lifting designs. the question is if one can upscale a small electric design to that middle, grey area large enough to lift a couple people. as i've said itt, i doubt it, but the reason they're going electric is they're trying to evolve the lightweight multirotor design to be as large as possible suck my woke dick posted:So basically a quick 10km hop, attendant swaps out the battery pack kind of deal. having all the battery pack infrastructure laying about the landing site definitely precludes just fitting pads in here and there wherever you have a spare 100x100ft area
|
|
#
?
Jan 16, 2019 02:33
|
|
|
luxury handset posted:when you have more than a few rotors involved that means more than a few motors and combustion engines are more complicated to operate, keep powered etc. than electric ones. so if you're going to do lift with a bunch of little rotors then it makes sense to have a bunch of little engines, since it's easier to power them all with wires than individual fuel lines or drive shafts. but if you're going to have one big rotor then it's better to have one big engine, and when it comes to big engines with big power combustion is better than electric Isn't this why someone invented range-extended hybrids? The powertrain is all electic but includes a gas engine to charge the batteries. Is anyone trying that, or is there too big of a weight penalty?
|
|
#
?
Jan 16, 2019 16:44
|
|
|
Baronash posted:Isn't this why someone invented range-extended hybrids? The powertrain is all electic but includes a gas engine to charge the batteries. Is anyone trying that, or is there too big of a weight penalty? The advantage of that in automotive applications is that you can have an engine running at peak efficiency to charge the batteries, instead of directly driving the car. The idea is that you can overcome the efficiency loss by using an engine that can run at peak efficiency the entire time. This is less of an advantage in aerospace applications, where we already have some pretty efficient gas turbine engines to provide constant power in their ideal operating window and propel the aircraft directly, before discussing the weight penalty involved.
|
|
#
?
Jan 16, 2019 16:50
|
|
|
Baronash posted:Isn't this why someone invented range-extended hybrids? The powertrain is all electic but includes a gas engine to charge the batteries. Is anyone trying that, or is there too big of a weight penalty? i think if you were going to bother having a gas generator on board an aircraft to power an electric engine, you might as well just simplify from there and have a gas engine. the advantage of gas over electric is that it is way more energy dense, so you can travel farther per weight of fuel
|
|
#
?
Jan 16, 2019 16:56
|
|
|
PT6A posted:The advantage of that in automotive applications is that you can have an engine running at peak efficiency to charge the batteries, instead of directly driving the car. The idea is that you can overcome the efficiency loss by using an engine that can run at peak efficiency the entire time. I've wondered this for years: Why not use a small gas turbine to charge a car's batteries? AFAIK that was the problem with the Chrysler Turbines, they ALWAYS ran at the same clip (and generated tons of waste heat), whereas that's a benefit in a hybrid. Guessing there isn't enough improvement over small cylinder gas engine setup?
|
|
#
?
Jan 16, 2019 21:59
|
|
|
Failson posted:I've wondered this for years: Why not use a small gas turbine to charge a car's batteries? AFAIK that was the problem with the Chrysler Turbines, they ALWAYS ran at the same clip (and generated tons of waste heat), whereas that's a benefit in a hybrid. probably yeah, using a combustion engine to charge batteries to power an electric motor is like, well, you have a combustion engine on board, so why not just use that when it's more efficient to use a combustion engine rather than an electric one? this is how pre-nuclear submarines worked, you'd have a diesel engine that would power the boat on the surface and charge batteries to drive an electric motor when submerged, since you don't want to run a noisy, gassy diesel when you're underwater as you'll make a shitload of noise and choke the crew. iirc freight trains are also diesel electric, so you can fuse the range of combustion engine vehicles without having to have overhead wires along the whole route, while also retaining the flexibility and ease of maintenance of electric transmissions alternatively, there probably isn't much of a market for an almost-EV that you have to throw fossil fuel into on a regular basis. if you're going to do a hybrid just do a hybrid, not a hybrid-hybrid, or a halfass electric car that still needs gas to go long distances
|
|
#
?
Jan 16, 2019 22:17
|
|
|
luxury handset posted:probably yeah, using a combustion engine to charge batteries to power an electric motor is like, well, you have a combustion engine on board, so why not just use that when it's more efficient to use a combustion engine rather than an electric one? this is how pre-nuclear submarines worked, you'd have a diesel engine that would power the boat on the surface and charge batteries to drive an electric motor when submerged, since you don't want to run a noisy, gassy diesel when you're underwater as you'll make a shitload of noise and choke the crew. iirc freight trains are also diesel electric, so you can fuse the range of combustion engine vehicles without having to have overhead wires along the whole route, while also retaining the flexibility and ease of maintenance of electric transmissions I've wondered why you don't see this sort of set up in hybrid sports/race cars to be honest. It seems like it would be rather useful in endurance racing.
|
|
#
?
Jan 16, 2019 22:57
|
|
|
luxury handset posted:probably yeah, using a combustion engine to charge batteries to power an electric motor is like, well, you have a combustion engine on board, so why not just use that when it's more efficient to use a combustion engine rather than an electric one? this is how pre-nuclear submarines worked, you'd have a diesel engine that would power the boat on the surface and charge batteries to drive an electric motor when submerged, since you don't want to run a noisy, gassy diesel when you're underwater as you'll make a shitload of noise and choke the crew. iirc freight trains are also diesel electric, so you can fuse the range of combustion engine vehicles without having to have overhead wires along the whole route, while also retaining the flexibility and ease of maintenance of electric transmissions To be clear, this type of hybrid already exists and is very popular, as it eliminates the transmission and allows a small internal combustion engine to run at a continuous level, which is more efficient than stop-and-go driving. I brought it up because I wondered why the same tech wouldn’t be beneficial to quads of a certain size.
|
|
#
?
Jan 17, 2019 01:25
|
|
|
Solkanar512 posted:I've wondered why you don't see this sort of set up in hybrid sports/race cars to be honest. It seems like it would be rather useful in endurance racing. Top tier endurance prototypes in the WEC are hybrids. Electric motors powered by batteries for instant torque and greater acceleration, recharged by braking energy.
|
|
#
?
Jan 17, 2019 02:19
|
|
|
Solkanar512 posted:I've wondered why you don't see this sort of set up in hybrid sports/race cars to be honest. It seems like it would be rather useful in endurance racing. Maybe... it doesn't work very well? Running a heat engine at optimal efficiency and using it to charge batteries is all well and good if it's in a stationary or intermediate performance application (read: generator or Prius). However, if you're trying to squeeze maximum speed or endurance out of a vehicle on a short race track or a long constant-ish speed run, just saving the weight or having a bigger gas tank is probably more effective.
|
|
#
?
Jan 17, 2019 02:20
|
|
|
Solkanar512 posted:I've wondered why you don't see this sort of set up in hybrid sports/race cars to be honest. It seems like it would be rather useful in endurance racing. I believe the efficiency of turbines is compromised at smaller sizes. Also, I'm only familiar with them in aviation contexts, but they are expensive as all gently caress compared to piston engines.
|
|
#
?
Jan 17, 2019 02:36
|
|
|
suck my woke dick posted:Maybe... it doesn't work very well? Running a heat engine at optimal efficiency and using it to charge batteries is all well and good if it's in a stationary or intermediate performance application (read: generator or Prius). However, if you're trying to squeeze maximum speed or endurance out of a vehicle on a short race track or a long constant-ish speed run, just saving the weight or having a bigger gas tank is probably more effective. The current top level WEC cars, as PCM notes, are all hybrids. In fact, they’re only allowed to operate under battery power when leaving the pits. Though there’s been a massive shake up in teams over the past few seasons, Porsche, Toyota and Audi have had very successful programs. To get back on topic, it will be very interesting to see what happens with the Roborace series this spring.
|
|
#
?
Jan 17, 2019 04:29
|
|
|
I was taking a look at the GAN thread and it reminded me of a number of the reservations I have about the fully automated future that tech companies want to bring us to, and since we just finished a riveting discussion on zoning for personal drones, it seems like as good a time as any to bring it up. The gist of it is this: I feel like machine learning has the potential to be hugely destructive to personal privacy and how we make decisions, and will ultimately consume tens of billions of dollars of wealth while being a net negative on society. Now, obviously, the field of statistical modelling is nothing new. People have been making decisions based on data spit out by computers for as long as computers have existed. However, the packaging of ML models in modules such as Sci-Kit Learn combined with easy access to massive amounts of computing power have lowered the barrier of entry considerably. There is no reason that anyone with the ability to install modules with pip is incapable of creating a functional implementation of a ML model within a few days. And if your model is too complex for your computer to handle, Google, Amazon, and others can provide essentially unlimited computing power at a price and level of convenience that would have been unbelievable even 10 years ago. This ease of access has led to a significant problem: you can learn the skills needed to run a csv file through a model without developing an understanding of how the data you’re putting in affects the data you’ll be getting out. All ML models do is recognize patterns in your training data and extend that to new inputs, meaning that any bias in your training data will lead to a bias in your model. Amazon faced this problem relatively recently when it came out that their HR hiring model was biased against hiring women). The model was trained on their past hiring information and was used to some extent in their hiring process such that “[Amazon] did not dispute that recruiters looked at the recommendations generated by the recruiting engine.” The article tries to frame this as a problem with the model, but it’s actually a problem with the training data, which clearly demonstrates that Amazon employees are biased against hiring women. Only now, it’s getting passed off as a computer problem rather than a human one. Biased training data is probably going to be the single most destructive element of applied machine learning. Since ML models can only recognize patterns, you are bound to see those patterns repeated. If a MAGA chud was doing your hiring for 5 years, then that data is going to continue to influence your model unless it is purged from the training set. Even worse are situations where models aren’t trained on enough features (each feature essentially being a column in a spreadsheet) to adequately understand the domain, and it creates patterns where none exist. A mortgage lender training a model based on historical default rates that include the time period of the Great Recession might inaccurately characterize certain groups as risky, unless the model was trained with features that charted the health of the economy at the time. The opposite problem is true as well, where adding too many features can also lead to patterns being created where none exist. This may not sound like a problem until you’re denied a loan because a ML model inaccurately decided that folks who play Candy Crush are a higher default risk. This will become an even bigger problem as machine learning gravitates to smaller and smaller domains. Your local credit union does not have the massive databases of loan history that giant banks do, and smaller datasets make biases more prevalent and result in less accurate, “noisier” models. And when some of the largest tech companies in the world are falling victim to this, how do you think the one guy at Bumblefuck Bank and Trust who “knows computers” is going to fare? Features, which as I said are essentially the columns of a spreadsheet, are the currency of the industry. Modern ML packages and computing power have given us the ability to automate the analysis of thousands of features and select the ones that are (or seem, see above wrt biases) relevant. As a result, there is an incentive to pack in as much information about each entity (be it a customer, client, whatever) as possible in order to gain the tiniest edge over their competitors. The industry-wide desire to train on more features is going to lead to an unprecedented land rush for personal information. Everything would be on the table, because anything could be relevant. Huge amounts of money will be exchanged to obtain everything from your social media likes, to past addresses, to any grocery purchases tied to a loyalty card. This has all been available for a while, but as these techniques pass even further into the mainstream you’ll see the demand (and, as a result, the value) skyrocket. Going back to the low barrier of entry for a second, a more generous person might say that is a huge benefit for the field, allowing single-person operations to compete with multinational behemoths. But as those behemoths absorb the majority of the (almost certainly insufficient) regulatory oversight, individuals will fly under the radar due to the lack of infrastructure needed to get up and running. Ultimately, this tendency towards applying machine learning techniques to everything in pursuit of efficiency and higher profits is bringing it into domains where the consequences are less abstract. People have been passed over for jobs, they have been denied loans, they have been denied housing, and overwhelmingly this has happened to protected classes. The whole ML discipline is codifying discrimination while claiming to take human fallibility out of the equation.
|
|
#
?
Jan 17, 2019 22:02
|
|
|
I think machine learning is still in early stages and it's important to understand what kind of results you get from the technology. As you said it's really easy to get started with ML tools being as user friendly as they are, but that doesn't mean the results are meaningful. There are certain types of data sets that can produce good results for very specific type of targets, but when you start boiling complex data points into generalized targets the results start to become less meaningful. At the end of the day all machine learning is doing is taking in data and filtering it into buckets that can be represented as binary or continuous variables. If your problem doesn't have an answer that can be represented that way it's difficult to use machine learning to provide anything useful. Your example scenario about hiring models is a good example of a complex problem that isn't easily solved with machine learning. To have a ML model provide a result that you could quantify as "good" or "bad" would require you to know what a "good" or "bad" hire actually is outside of the context of machine learning. I have yet to see a person or algorithm that can assign a numerical value of worth to an individual contributor. I think it is definitely a human problem and not a computer one, you can't blame the computer for creating a measuring stick that doesn't include women.
|
|
#
?
Jan 17, 2019 22:36
|
|
|
ElCondemn posted:I think machine learning is still in early stages and it's important to understand what kind of results you get from the technology. As you said it's really easy to get started with ML tools being as user friendly as they are, but that doesn't mean the results are meaningful. But it’s something we still need to be wary of. You’re absolutely correct that such results won’t be meaningful in a literal sense but that won’t stop managerial assholes from believing that they’re meaningful and implementing incredibly harmful policies and practices because of it.
|
|
#
?
Jan 17, 2019 22:42
|
|
|
Baronash posted:
 this is probably one of the better opinions on machine learning people have right now.ElCondemn posted:I think machine learning is still in early stages and it's important to understand what kind of results you get from the technology. As you said it's really easy to get started with ML tools being as user friendly as they are, but that doesn't mean the results are meaningful. The thing is, most people don't realise that garbage in = garbage out. Machine learning algorithms outside of a select few companies suck my woke dick fucked around with this message at 00:02 on Jan 18, 2019 |
|
#
?
Jan 17, 2019 23:59
|
|
about artificial stupidity
|
suck my woke dick posted:
Not to mention their proud, MBA holding manager.
|
|
#
?
Jan 18, 2019 00:20
|
|
|
Solkanar512 posted:But it’s something we still need to be wary of. You’re absolutely correct that such results won’t be meaningful in a literal sense but that won’t stop managerial assholes from believing that they’re meaningful and implementing incredibly harmful policies and practices because of it. Ultimately, technology can't solve people problems. The problem isn't even one of new policies and practices. It's more that technology is used to shield the existing problematic practices, because people tend to have too much faith in technological solutions (especially when buzzwords like "AI" and "machine learning" are thrown around), and the marketing and public relations divisions of every company are well aware of that. In non-expert fields, people just assume that a computer solution will be better at decisions and less biased than a non-computer solution, even though the computer solution is typically running on the same bad data and flawed assumptions that were used in the non-computer solution. Part of it, too, is that we've spent half a century generating bad data that backs up existing biases. Remember, the term "meritocracy" was originally coined in a satire piece. Our growing reliance on quantifying people in every way possible was often just a roundabout way of justifying existing class distinctions.
|
|
#
?
Jan 18, 2019 06:36
|
|
|
(Sorry if this just sounds like gobbledygook to anyone who's never played Starcraft) DeepMind just did an event where they showed their new AI agent, AlphaStar, playing Starcraft 2 vs pros. Prior to this, AIs in Starcraft were sometimes very good at executing one strategy, but only if you didn't know it was coming, because they were very exploitable if you picked out the build's weakness. More general purpose AIs were mostly quite bad, even a moderately competitive regular gamer who ladders frequently would trounce them quite easily. DeepMind, of course, is the Alphabet subsidiary that made AlphaGo, the world-dominating Go-playing AI that shocked the Go world when it beat the best players in the world. Starcraft is quite a jump up in complexity for an AI, for two main reasons: imperfect information, and the state space. Go and chess are perfect information games, you can see the entire game state at all times. Starcraft has a fog of war, so you had to infer/guess at what your opponent is doing a lot of the time (although of course you scout occasionally to see what's actually happening). Getting an AI to guess is hard. The decision space in Go, while large, pales in comparison to a game like Starcraft 2. An RTS like SC2 just has an astronomically larger decision space to play around compared to a discrete, turn-based game with relatively straightforward rules. At any given time, in a game like Go, your possible actions number a few hundred, tops. With an RTS...just between having dozens of buildings and units, times each of those usually capable of several different actions, times those often being targetable across at least tens of thousands of effectively unique positions on the map... you're looking at any one action being chosen out of theoretically millions of options. And then multiply that with each 'turn' (game tick) being kind of inconsequential because it's real time, you have to choose potentially several of those actions every second (human pro players can spike to 8-10 actions per second during an intense fight). It gets pretty crazy, from a computational perspective. It gets pretty crazy, from a computational perspective. The branching factor (average possible legal moves at any given time) for chess is 35, according to the googling I just did. For Go, it’s 250, which is a big reason why Go was so much harder than chess for AIs to get good at, and why a lot of people thought pro-level Go AIs were still a decade away when DeepMind unveiled AlphaGo. At a branching factor of 35, looking 5 turns ahead exhaustively means exploring about 50 million possible states. At a branching factor of 250, looking 5 turns ahead exhaustively means exploring just under a trillion states, so being able to look ahead in Go meant having to prune the state space/decision tree much more aggressively. Even if you are unreasonably conservative with Starcraft and assume only a hundred thousand possible actions, with, say, effectively five turns a second (in real games pro players often do more actions than that per second, but 300 APM is a reasonable average for top tier players), that means you have 10,000,000,000,000,000,000,000,000 possible states after one second. After five seconds, you’d have 1E125 possible states, 1 followed by 125 zeroes, or astronomically more possible states than there are atoms in the universe. So you’d need a way to very aggressively bucketize values in those states for ones that are effectively the same, without losing nuance that might be important in odd cases. I mean really you’d need a completely different way to explore different possible strategies, I’d be shocked if what they did looked very similar to AlphaGo, other than the obvious things of still using a neural net, reinforcement learning, etc. To elaborate on this a bit more, at most, during Go, an AI has to see into the future a few hundred turns, because the board fills up over time. At a tick rate of ~22/second, a few hundred 'turns' in Starcraft amounts to maybe a dozen seconds, and you obviously have to be able to plan further ahead than that. MUCH further ahead. But, without getting so bogged down in long term planning that you lose the moment to moment fights. The state of the game is much more complicated, too. Each ‘unit’ in Go is identical, and each board space of which there are (19*19) can essentially only have three possible states: empty, black stone, or white stone. In contrast, for every unit in SC2 you’re looking at at least hundreds, maybe thousands of unique positions across X and Y for positioning, plus there are dozens of different unit types, each with their own behavior. And each unit, in addition to positioning, may also have variables tracking its current health, energy (mana), ability cooldowns, buff/debuff status (of which there may be multiple), animation/direction state, and upgrade status. You could probably fit an entire Go board, data-wise, in the amount of bits it takes to hold a handful of Starcraft units. Anyway, first they showed a...let's say medium-high strength professional player, TLO (TheLittleOne) playing his offrace (he usually plays zerg but the agent only does protoss vs protoss). If you wanna see the games for yourself, go check the videos here, spoilers after these links: First pro: https://www.youtube.com/watch?v=cUTMhmVh1qs&t=2660s Second pro: https://www.youtube.com/watch?v=cUTMhmVh1qs&t=5558s Live exhibition: https://www.youtube.com/watch?v=cUTMhmVh1qs&t=9067s AlphaStar beat him 5-0, but it looked fairly beatable still. There were little mistakes and weaknesses that were apparent that looked exploitable by a stronger player. Oh, I guess I should mention that actually TLO was playing a different agent each time. In some sense this is more fair; the agent isn't learning from just playing TLO, but TLO can't really learn either since each agent has different preferences, which was very clear. However, in another sense it's kind of 'unrealistic' in terms of how humans usually play each other in tournament: multiple games vs the same person in a series, so if you're a one trick pony, your opponent can easily exploit that. Oh, it was also the same map every time, I guess that's another restriction the agent has right now. After that, they had Mana come in, who's a pretty top tier protoss player. Not the best in the world (that's nearly all Koreans), but quite close, definitely one of the best foreign protosses (in the Starcraft world, "foreign" refers to anyone who is not Korean). He played new agents who had been trained in the week in between the showmatches, and he also lost 5-0, and the agents looked much, MUCH stronger this time. It was much harder to see any clear errors in its play. In particular its micro handling just looked superb, like what you see the best in the world do when they're really in the zone, except consistently throughout basically every game, and even when its forces were split up and fighting on multiple fronts. In Game 4, there was a situation where Mana had clearly the stronger overall army composition, but AlphaStar was managing three separate bands of stalkers on all sides of Mana's army perfectly cleanly, dancing back and forth to pick away at it with no mistakes. Apparently AlphaStar is limited in its APM (actions per minute) down to reasonable human levels (by pro standards), but it was obvious there that it was incurring no context switching hit in switching between controlling each separate force, whereas when humans are in that position there's a tiny delay each time, even with top tier pros. As DeepMind explained, while AlphaStar is limited in how fast it can dole out actions, it does see the entire map zoomed out at once, which is cheating a bit since humans can't do that even if they want to. After showing those replays -- live commentated by Artosis and Rotterdam, who hadn't seen them before and didn't know the outcome so that their reactions would be genuine -- they then did a single live showmatch between AlphaStar and Mana, this time with another new agent who didn't have the zoom out trick, but who their internal metrics said was equally strong. At first it seemed to be gaining the edge of mana through superior micro, particularly it had some really strong oracle harass that Mana struggled to fend off efficiently, and between that and expanding faster than Mana it looked like it was in a very strong position. But, then Mana came in with a warp prism with two immortals to harass AlphaStar's main base, and AlphaStar brought in its entire army, painfully slowly (because it was big and had walled off at its main choke so the gap was very small) to defend, at which point Mana retreated after doing a little damage. Mana had an observer sitting in a clutch position at AlphaStar's natural (first expansion), and then he just sort of went back and forth with his warp prism harass, pulling back each time AlphaStar came to defend, then going back in once AlphaStar pulled out its defenders, etc. Did this like four times total I think before AlphaStar finally killed the observer spying on it, and then had oracles follow the warp prism (oracles are flying but can't shoot other air units, so they could only follow). At this point, AlphaStar was still probably in the stronger position, maybe army was weaker but not too much weaker, and much stronger base/economy setup. But then AlphaStar's main force just camped out near where the warp prism was, just out of reach, so Mana came in with his army and flanked AlphaStar, doing some damage, but more importantly getting his army in a position between AlphaStar's army and its base. Then he waltzed over to AlphaStar's natural, demolished it with AlphaStar putting up little resistance (probably calculated that it couldn't win the fight), then walked over to the third, AlphaStar finally coming in but with a terrible army composition and kind of awkwardly positioned, Mana took him out and at that point it was basically GG, so Mana finished off its main base and then everything else. Overall, AlphaStar looked very strong, but much more so on attack than defense. It looked fairly...trickable when responding to harass in that live match. Would be interested to see it against various cheese strats, and also when it's more flexible (more race matchups, more map choices, able to have a single agent throughout a series). Cicero fucked around with this message at 23:55 on Jan 24, 2019 |
|
#
?
Jan 24, 2019 21:53
|
|
|
Cicero posted:DeepMind, of course, is the Alphabet subsidiary that made AlphaGo, the world-dominating Go-playing AI that shocked the Go world when it beat the best players in the world. Starcraft is quite a jump up in complexity for an AI, for two main reasons: imperfect information, and the state space. Go and chess are perfect information games, you can see the entire game state at all times. Starcraft has a fog of war, so you had to infer/guess at what your opponent is doing a lot of the time (although of course you scout occasionally to see what's actually happening). Getting an AI to guess is hard. The interesting thing there is that "incomplete information" and "state space which can't be practically expressed as a finite list of choices, and which plays out in real time" also describes most potential applications of robotics outside a controlled environment. Such as, perhaps, driving a car. So this could potentially lead to some cool stuff in the near future.
|
|
#
?
Jan 28, 2019 03:42
|
|
|
Cockmaster posted:The interesting thing there is that "incomplete information" and "state space which can't be practically expressed as a finite list of choices, and which plays out in real time" also describes most potential applications of robotics outside a controlled environment. Such as, perhaps, driving a car. So this could potentially lead to some cool stuff in the near future. Or it simply demonstrates how fundamentally inadequate the AI is and always will be for the task. Actually watching the video makes it clear just how constrained the parameters have to be for the AI to have a chance at all in a real-world scenario without cheater powers, and even then an exploit is ultimately found that trivializes all its so-called advancements.
|
|
#
?
Jan 28, 2019 10:32
|
|
|
|
| # ? May 8, 2024 12:17 |
|
|
Kerning Chameleon posted:Or it simply demonstrates how fundamentally inadequate the AI is and always will be for the task. Actually watching the video makes it clear just how constrained the parameters have to be for the AI to have a chance at all in a real-world scenario without cheater powers, and even then an exploit is ultimately found that trivializes all its so-called advancements. "Have a chance at all" is an interesting interpretation of a 10 - 1 score against top level players. The matches demonstrate that it would 100% beat most human players in those scenarios, so I don't know how you're getting the "fundamentally inadequate" reading from this. AFancyQuestionMark fucked around with this message at 10:42 on Jan 28, 2019 |
|
#
?
Jan 28, 2019 10:40
|
|
