You've missed the point. He isn't advocating for everyone to suddenly use JPEG where they wouldn't before. He's advocating that, since there are so many JPEGs already in use, the architecture for displaying them should avoid unnecessarily converting them into RGB 32bit, and instead display them without conversion in a YUV 420 format, to align with the internal encoding of JPEGs which is YUV 420.
Essentially, rendering JPEGs directly as a GPU texture (since GPUs can already natively use 420 encodings for textures). His point is that this would improve performance, very slightly improve quality, and reduce power usage as there's no colour space conversion.
>global capitalism is the major force for extreme poverty in this century.
The complete opposite. Where do you think globalised capitalism gets cheap labour from? Hint: it's not developed countries. Then to get them to work for them, they have to pay them something, and people aren't going to accept the jobs if it's not enough to live on.
The empirical data also bears this out, as the other reply notes.
Now, whether capitalism is a force for increasing wealth gap and relative poverty in developed countries is a different question; but the "poverty" in that case is certainly much less "extreme."
Why the fuck are you worrying about "metaphysical frameworks?" The things EA mostly focuses on are people dying of easily preventable starvation and disease. There is no need to bring up highly abstract ideas about "metaphysics" (whatever you mean by that - morality isn't generally considered to be in the field of metaphysics) when the suffering is so blatantly concrete.
Wow. If Amazon is being held up as a good example, things are truly bad. My only experience is with Audible, and it is way too difficult and obscure to figure out how to cancel your subscription. Although at least once you do they don't try to pull anything except give you a discount, I suppose.
Although that shouldn't be held up as a positive either. Where I am, it's literally illegal for utility companies to offer discounts or any other type of benefit to customers who wish to cancel or move to another provider. They aren't even allowed to contact you unsolicited after you have requested to cancel the service, as it could be interpreted as attempting to offer enticement to continue. Ending a subscription to a service should not be a difficult thing.
At the very least, if you are showing the instalment price, you have to show the number of instalments, and the fact that it is an instalment rather than a pay-as-you-go subscription.
You could certainly argue that describing the plan as $x paid monthly is outright false, when it should be "12 monthly instalments of $x".
So you're in favour of people phone-scamming the elderly? They just need to take personal responsibility and do their due diligence, right? Or do you want to take away the ability to buy things of people you deem incompetent?
It is a basic necessity for beneficial trade to occur that the terms are made clear, not just technically available, to the people potentially wanting to undertake that trade, before anything legally binding occurs. The issue is not that there are cancellation fees, it's that they are not made clear.
The exaggerated, but still representative, example might be the Earth demolition plans in Hitchhiker's Guide, which were "available for 50 years in the local planning office in Alpha Centauri." Totally fair, right? Anyone could have just gone there and read them.
Demanding "personal responsibility" as a solution to avoid bad deals is the same as accepting that it's OK for some people to be scammed. No one is perfectly on guard all the time. If it is permissible for companies to engage in scammy behaviour, and scammy behaviour increases profits, then in order to be competitive, companies will engage in scammy behaviour. So it will be everywhere. Since, as noted, individual humans aren't perfect all the time, eventually someone will get scammed.
The obvious, morally superior, and more economically beneficial solution is simply to not allow scamming to be profitable. That, among other things, means ruling in favour of the consumer when companies perform shady practices.
Postscript: People with this kind of "libertarian" stance really annoy me. To give the benefit of the doubt, I'll assume that "personal responsibility" isn't just thinly veiled personal exceptionalist elitism. If that's the case, the only reason you can even have these beliefs is because you take so much about the modern world for granted. There is a lot going on in order to create an environment where mutually beneficial trade can happen.
In the absence of strict, neutral enforcement of fair trading rules, people scam the shit out of each other. It's way easier, and more profitable, than actually creating something of value that people want to buy. The total effect of this, if it were allowed to occur, is that it creates an environment of distrust and uncertainty, which means mutually beneficial trades that would otherwise happen won't happen, making everyone poorer. Advocating for "personal responsibility" as a solution to avoiding scummy business practices is directly advocating for this situation.
Historically, before commercial law was explicitly codified, the means of ensuring that merchants were trustworthy was essentially reputation. The way this scaled was with merchant groups, where the reputation of one member going bad affected them all, so they all had incentive to police each other (or alternatively, collude). Plus the fact that people in general were much less dependent on markets for necessary goods, the labour force being mainly agricultural. However, this made opportunity for trade highly exclusive, and largely only available to the privileged ("merchant class").
I guess my question for you is, do you want to have to read every single EULA, thoroughly scan the terms of every transaction every time you buy something, and otherwise do a lot of work and research any time you have to make some vaguely commercial agreement to avoid getting into an agreement you, in retrospect, didn't want? If not, I suggest you carefully re-examine your beliefs and what they imply.
A final remark. The economics of scamming are highly asymmetrical. If individuals have to do a lot of due diligence even for routine transactions, that is a massive cost to that individual. In contrast, organisations willing to partake in scummy behaviour can spend almost endless resources optimising their ability to ensnare customers, because once the process is developed, the marginal cost is very low if not zero. So individuals cannot be expected to "keep up" in this environment.
I mean, the only reason things like this aren't commonplace is cost (of the skillset, tools). Basically anything's possible these days with CGI, everything is purely a matter of the amount of effort you want to put in.
And for artistic purposes, why does it matter how the final result was arrived at? If we have powerful and easy techniques for realising an artistic vision, that doesn't seem like a bad thing?
Yep. To be useful for exploring potential "true" values, a system would probably need some way of showing you the distribution of its guesses, so you can get an idea of whether there is any significant information there.
That aside, you'd still probably need a ML PhD to have a chance of correctly interpreting the results, given the myriad potential issues with current systems.
I will admit I don't fully understand the implications of this.
Doesn't this mean it's essentially game over for running untrusted JS by-default? Doesn't default-deny functionality like NoScript have to become mandatory in browsers for security? If not, why not?
There is also this which seems to have been produced in parallel, closely related to the first but not exactly the same: https://arxiv.org/abs/1804.11271
Thanks for your [2] link paper, it's an interesting one.
If you think about it, the high-level idea makes some intuitive sense. Since NNs are known to be "equivalent" to kernel methods - which are, to oversimplify, essentially nearest-neighbour + a similarity function - then the ability of NNs to memorise specific training examples can be analogised to adding another "neighbour" to interpolate from. So maybe it's not too surprising that NNs which can do this can have better generalisation performance. (Although it is still surprising, since I certainly wouldn't predict it a priori!)
Really, what is the difference between a "feature" and "memorising a data point"? It seems only a matter of scale - how much of the input space is "relevant" to the learned representation.
Why exactly is it a problem that powerful models can memorise training data with random labels, if they generalise well on non-memorisation problems?
I also think that a lot of progress is being made seemingly in parallel on this "understanding NN [generalisation]" topic, and the different threads don't necessarily seem to be aware of each other.
This article seems to have failed to mention most of this recent progress. Possibly this is an artifact of the time it takes to produce and publish a completed article, but it presents a picture that we are more clueless than we actually are.
The most promising, IMO, is outlined in these blog posts:
which summarise 3 research papers. The basic idea is that because NN outputs are usually thresholded to produce a prediction/classification, there is not always a change in the output if a parameter changes. This gives space for there to be a non-uniform distribution over output functions, if weights are initialised randomly. The authors investigate this empirically (with some theoretical justification) and find that this bias is toward "simple" functions, which often generalise well, due to "simple" properties mostly being what we care about in real data. Then generalisation can be explained by SGD simply being more likely to find parameters which cause the NN to express a function that is simpler, out of all functions that perfectly match the training data.
There is also some interesting investigations into infinite-width limits - the most well known being the result that NNs are equivalent to Gaussian Processes at initialisation, then Neural Tangent Kernels [1] showing a way to view them through the lens of kernel methods throughout training as well.
But then there is Feature Learning in Infinite-Width Neural Networks [2] which seems to be at odds with the kernel approach (due to kernels not learning features) - this paper provides a different method for parametrisation to admit feature learning.
So there is progress both in understanding generalisation and understanding when feature learning does & does not occur. (And this is hardly a comprehensive sampling.)
Absolutely. The fact that China almost completely disregards IP and shows no signs of "innovation" slowing down disproves the notion that you need government backed monopolies on inventions in order to encourage innovation.
These arguments get trotted out every time this comes up, and they are bad.
>1. The inventor would pay the R&D costs with no good way to recoup them, because he'd be competing against knockoffs who didn't have to make that investment, and can price accordingly. You'd be putting the actual inventor at a disadvantage.
It's now clear that there are plenty of ways to fund R&D without relying on IP law to recoup the costs. Another reply goes into this better, so I want to also point out another benefit: the incentive to perform R&D, without IP law, is less likely to be "in order to make a profit," and more likely to be "because whoever is funding it wants it to be done," whether for itself or because they want the end product to exist.
Remember that an open market capitalist economy is _a method for allocating resources within a society towards ends that people actually want_. Doing R&D for profit is an indirect way of incentivising companies to produce things that people want. Without monopoly licensing, the incentive to the company is directly to research & develop something that people want - the funders, at a minimum.
Some lessons from e.g. machine learning research show that any slight misalignment in objective when you have a powerful optimising system, like a corporation, will be ruthlessly exploited, until what you end up with is nothing like what you wanted. In particular, instead of solely making things that people want, companies spend massive amounts of effort on making people want the things they make.
>2. There'd be no incentive for inventors to actually publish designs for their inventions. Most people who argue against patents as a concept forget is that to get one you have to publish plans for your idea openly, which makes it much easier to copy. The limited monopoly is an incentive to part with that information. You'd end up leaving the public domain poorer.
This argument is nearly completely redundant these days. Nowadays, to reproduce something it's almost always enough just to know it's possible, let alone knowing precisely how or even having a working example. If that wasn't the case, why are copyrights necessary on manufactured physical products like drugs? The producers wouldn't be concerned about their products being copied if it wasn't possible to do so without explicit instructions.
Another poster brought up that patents are now written to be as obscure as possible anyway - also true, nearly removing all public benefit that might have been obtained from such a system in the first place.
>3. Invention and production are different skillsets. For an author to get paid for writing, they shouldn't need to own a printing press and distribution network. It's a good thing that good, successful authors can make their livings by writing and not always be forced to sustain themselves with a day job.
This just makes you seem out of touch with the modern reality of the writing industry. You may have heard that most musicians, apart from a few mega popular ones, make their money from live performance & merchandise. Writing is going a similar way. Self publishing is a bigger and bigger thing, lots of authors have communities or are supported directly by fans & merchandise. There are multiple companies that can print any arbitrary work into a physical book for you without requiring ownership of the IP.
Of course under the current system there will always be a few big winners, but for the aspiring new writer it's pretty unlikely you'll get very far unless you're exceptionally talented or you know people high up in the industry.
Essentially, because the average creator doesn't have a realistic shot at making money from licensing anyway, they are already adopting models that work in the absence of licensing revenue. If I were to boil it down, they're mostly selling one of three things: community, experiences, and anticipation of future creations that wouldn't otherwise exist (in other words, investment, except the investors don't own the creator's work or otherwise receive any formal stake.)
The original ideas behind IP were conceived in a different time, and many of the assumptions simply don't apply any more. It's possible that there are still good reasons why some form of IP needs to exist, but these tired justifications aren't it.
Par for the course with these captcha-like tests though.