clickok

In imperfect information games (like SC2) the outcome prediction implicitly takes into account the unknown. Given what it has observed and what it has not observed, it is essentially comparing the present state to similar situations from the past.

You can see this in the replay-- at seemingly random times, its outcome prediction jumps up, even when it hasn't had any interaction with its opponent. But that's precisely why it's going up-- it notices that its opponent has not executed a faster rush or cheese that it's unprepared for, hasn't expanded early, and the scout has not been destroyed.

Similarly, after having won a fight, the worst thing that could happen is that a bigger force emerge from your opponents base, destroying your army and giving them a chance to rebuild. When that does not happen, you know that your opponent probably doesn't have such an army (because otherwise they are falling behind in resources due to being bottled up). Either way, after a certain point it's worthwhile to press on to cause more damage, because you're now far enough ahead in resources that you will win regardless if they manage to repel that particular attack.

It does, insofar as you can express anything meaningful in an inconsistent system. A formal system being inconsistent implies being able to prove some statement A, and also its converse ~A. If both A and ~A are true, then we can prove that every other statement in this system is also true. A quick proof:

1. A 2. A v B 3. ~A & A v B 4. B

We start with A. The union (logical-or) of true statement (here A) and any other statement (say B) is a true statement, thus A v B. Then we introduce another true statement, ~A, via logical-and to get ~A & (A v B), which simplifies (disjunctive syllogism) to just B.

So we have proved B, but B was arbitrary. It could be anything, including the statement that x = y for x and y two ostensible non-equal numbers.

Assuming that insurance is already priced as accurately as it can be according to the available data, then discovering that wealthy people tend to evacuate disaster areas while the poorer residents remain would suggest that you charge greater premiums for the impoverished and give a price break to the rich. As it stands, the poor are effectively getting subsidized by otherwise similar individuals with more wealth.

Insofar as the utility associated with money is roughly logarithmic, then the net utility from a pricing change will likely decrease. For example, a wealthy person will be pleased with a $500 reduction in premiums (that's a new pair of shoes or some decent concert tickets); a $500 increase might induce a poor person to forgo renewing their life insurance (with disastrous effects for their dependents should they die) or if they pay it, then other worthy expenses might get cut to make up the difference (home repairs, sending your kids to summer camp, retirement savings).

Even assuming that everyone values money equivalently despite their net worth, it still leaves a bad taste in my mouth. Here's one wealthy elite technocrat, already invading people's privacy on an Orwellian scale, and he's using that access to explicitly prey on poor people to make money for other people who are already rich.

In the article it mentions that the data comes from apps on the phone, so insofar as the user has clicked through the EULA, Privacy Policy, and permissions screens[0], it's voluntary in theory but obviously not in practice. The interesting thing is that the data source comes from "hundreds" of innocuous apps (weather, directions to ATMs, social media add-ons) that send the user's location data to a company that aggregates and then sells it.

This in turn is only possible because for useful but easily coded apps, you'll always have a race to the bottom on price, and once you hit free then the only way to make a return on your investment is by taking advantage of the user, by ads or by tracking. And it will be this way for the forseeable future, or at least until the harvested data indicates that there's uncaptured profit in making apps that aren't free, but promise to respect your privacy.

-----

0. It sure would be nice if someone could use natural language processing to parse the legalese and spit out a summary of what you're really agreeing to. Maybe combined with some sort of phone-farm (like they use for testing builds on different devices) to download various apps in order to acquire the corpus and maybe test the what servers the apps communicate with, and how often. I know that there are sites dedicated to interpreting EULAs from major companies/apps, but probably the greater threat is the myriad free utility apps that get used once and then forgotten.

Read the list of individuals to whom this "sophisticated" personality clustering technique was sent. You've got people like A.V. Aho, R.W. Hamming, J.B. Kruskal, J.R. Pierce, K. L. Thompson, J.W. Tukey (and probably more who I didn't recognize). That sort of sheer concentration of talent is a rare and precious thing.

So how long would you expect books bound this way to last? I took some pictures of the ones I've made for reference: https://imgur.com/a/60zP8vn . So far, none of them have come apart but I'm pretty careful with my books.

I took some pictures to demonstrate the final product[0]. I am not an expert bookbinder, but it works well enough and is cheap and easy to do by hand. This is the result of tinkering; presumably someone who knows what they're doing could achieve better results.

-----

0. https://imgur.com/a/60zP8vn

The problem with most recent books is that they use a glue-based binding, but the glue seeps into the spine far enough that it makes laying them flat difficult.

However, if you bind it yourself, you can get the desired behavior pretty easily. I have a summary of my process in the footnote of an older comment[0], but I've been able to get it working with hardcover books, and experimented with using a little bit of thread to reinforce the binding. So far, none of my books have come apart and they all lay flat. The problem with this is that you either have to print your own books or remove the binding from existing ones and then rebind them, which is not especially time consuming but not exactly fast either.

I wonder why this style of binding isn't more common? Is it just too expensive or are there durability concerns?

-----

0. https://news.ycombinator.com/item?id=15877525

I am not sure that he's the first to note that the brain is a predictive machine, or to notice that models of state or the environment can be formulated entirely in terms of predictions[0].

Personally, I find that idea very convincing, because we already know how to learn to predict (reinforcement learning), and we can use the same framework to optimize an agent's behavior to maximize against these predictions. There is some evidence that the brain uses something like temporal difference methods to learn.

The hard parts are: how to do this in a stable and efficient matter, and how to build up our representation to go from relatively simple goals (e.g., survive and collect resources) to more abstract goals (e.g., translate a poem in such a way that the original's charm is preserved[1]). That is, you want to have a system that is capable of climbing up Maslow's hierarchy (or something like it) and actually proceeds to do so based on emergent properties of the system itself.

Proposing and investigating new ways of constructing learning systems so that there is a continuum between primal drives and complex abstract goals is hard, and even if grid cells are an epiphenomena, then it is worth studying them to see how they might develop from simpler rules. If Hawkins' team can make something useful, then that can still be beneficial by allowing for more powerful and flexible AI, even if it's not a theory of everything.

---

0. See for example: http://web.eecs.umich.edu/~baveja/Papers/psr.pdf (Littman, Sutton, and Singh), which describes and analyzes the equivalence between belief-state based representations and state representation based on predictions.

Or just a private firm[0], since models need training data and Facebook has a truly staggering amount, all pre-tagged. There is also a small amount of irony in the fact that Facebook's initial incarnations were pre-seeded with scraped data from Harvard student websites.

---

0. For example: https://www.businessinsider.com/ntechlab-google-facial-recog...

Originally I found the Magic Leap extremely exciting, because in contrast to phones or tablets, whose touchscreens limit what sort of applications are feasible, AR could allow for arbitrary interfaces. These interfaces could be present alongside the real world, making them both more convenient and maybe less of an attention-vortex.

This is (among other things) what Magic Leap pitched, and comparing the actual product to their press releases is to note a vast discrepancy. Still, hope springs eternal, so while I wasn't impressed, I know how hard AR is and I wasn't inclined to judge too harshly-- maybe it will get better as technology improves.

After having read the article, I find myself rooting for them to fail. This is a company that has taken (and arguably squandered) billions of dollars to release a slightly cheaper HoloLens. Instead of feeling abashed, they continue blithely on, even having the temerity to invite developers to their platform because the rest of the world (and in particular, other platforms) has so much "baggage" and "negativity".

As the article notes, this is a company that is propped up by oligarchs, settles sex discrimination lawsuits, and is pursuing military contracts with repressive regimes. Although to be fair, improving soldiers' lethality is the only use case they've articulated that justifies the price. Now that the missed expectations and general shadiness are starting to catch up with them, they're moving to attract developers. Somehow, despite all the money, they haven't been able to pay for an idea to justify the hype-- what else could you call it but a con?

Yes, you could have different possible production rules for a given symbol and sample from them according to some distribution. You could also just randomly perturb/fuzz a given iteration of the system.

A deterministic system is nice, however, because you can make changes to the production rules and rendering without having to worry about as many possible outcomes (introducing stochastic production rules increases the number of possible outputs combinatorically). This is not too much of a problem if you're just making cool fractals for fun, but if you're doing procedural generation with L-systems then you have to keep an eye out for degenerate cases. For example, if you're generating foliage this way, you might get plants that grow into the ground or in weird non-biological patterns.

That might defeat NLP-based analysis, but the use of obfuscation and the methods employed still tells you something about potential authors.

If someone in an organization writes an essay disparaging that organization, but does so in an unnatural style that seems intended to defeat author identification, then you know that they're aware of stylometric analysis, and have created or downloaded a tool designed to interfere with assigning attribution. Depending on the organization, that might shrink the candidate pool more drastically than clustering based on sentence length, favorite words, or other linguistic quirks.

There are a couple of places where RL/Evolutionary Methods/Something Else are jockeying for the top of the leaderboards, but usually if one technique succeeds over the others it just means that someone's put a lot of work into that one particular approach.

Anecdotally, I've heard about some optimization problems that should be a good fit for dynamic programming or reinforcement learning where it turns out they actually don't seem to work as desired.

For example, optimizing elevator policy: an agent controlling an elevator wants to achieve high throughput of passengers while also minimizing the wait-time of each individual. We want people to get to their floor quickly but also want to avoid having any individual wait for an excessively long period of time. The agent can only observe information about the buttons pressed on each floor, although it gets a reward like that reflects our desiderata[1]. As it turns out, most elevators are not running some sort of cool machine optimized policy, but rather something hand-coded.

This is not for lack of opportunity, either-- apparently, some researchers pitched Otis on this in the 90s, and it didn't work as well as what they already had, despite this looking like a case where theory should match reality. Why this is, I don't know, but it might come down to the fact that there's really a lot more to a "pleasant elevator experience" than might be assumed (or modeled by the reward function), or perhaps the hand coded policy incorporates background knowledge about human vertical locomotion unavailable to a machine.

-----

1. Maybe something `R_{t} = Transporting(t) - ∑ Waiting(i, t)`, meaning something like "number of people currently being transported minus the people who are waiting times how long they've been waiting for".

Pretty cool, this is actually a great reference for a lot of things. Even if you're familiar with RL, you might be reminded of something or learn something new.

Sutton and Barto's book is also good if you want to do more than dip your toes into RL: http://incompleteideas.net/book/the-book-2nd.html That page has the PDF with links to code, problem solutions, and course material.

If you just want an idea of what an RL algorithm looks like, I've got a (heavily) documented version of TD(λ) for linear function approximation here: https://github.com/rldotai/rl-algorithms/blob/master/py3/td....

I skimmed through this and have already found a bunch of interesting sections, but there's also a ton of background information on topics related to bandit algorithms.

The authors say that this is the first draft of the book submitted to the publisher, so I suppose it's nearly complete? More details available at the site they put up, http://banditalgs.com/

It doesn't have to be literally false to misinform, and informing the public is (in my opinion) the main point of journalism. How erroneous/slanted does it have to be before we can assert fakeness? I think it's fair to say that if I have a worse understanding of the world as a result of consuming news media then that constitutes "fake news".

I am not so sure; there's rarely a strong incentive to misrepresent a paper on e.g. number theory, but that is not the case when it comes to issues in politics or business. We already see this to some extent with Wikipedia-- the pages on math and hard science are generally accurate and well-written. Biographies are much more a mixed-bag, even aside from the fact that the subject themselves can edit their page, I know there are companies that offer to create or manage wiki pages for PR purposes.

It might be better to take an approach similar to "open science". Have the reporters publish their raw footage and interview transcripts to address misrepresentation/cherry-picking, and get them to commit to answering follow-up questions from readers to allow the readers to get additional information. Having more of a back-and-forth could improve the quality of the reporting, allow for spurious objections to be debunked, or ensure that the author gets taken to task if their bias taints the story.

This seems time/capital intensive and therefore impractical in the age of listicles, but on further consideration I realize that many "name brand" bloggers popular on HN essentially follow this model-- for example, Matt Levine and Brian Krebs. However it's not clear whether their model can be applied to all areas of journalistic interest instead of specific niches, or if it requires a rare kind of individual to make it work.

It is, and has been for decades. Usually, when I read articles on a topic which I have researched, I find the reporting tends to be "wrong" in some sense. Charitably this can be blamed on time or space constraints: the articles need to be ready for publication while the story is relevant, so the author has insufficient time to achieve expertise; or even if the author is an expert, they are incentivized to publish only the most interesting (and perhaps most superficial) details on the subject. Less charitably, you might suspect that the author (or their editor) has a bias depending on what they elide or emphasize. Herman & Chomsky wrote about this topic[0] examining the various filters that determine what gets reported and from what angle.

However, it's recently become politically controversial to acknowledge this fact, I think because Trump uses "fake news" to disqualify criticism of himself and his administration[1]. But let's not pretend that journalism is without faults or bias just because (on average) the virtues of the fourth estate outweigh their flaws.

-----

0. https://en.wikipedia.org/wiki/Manufacturing_Consent

1. He's hardly the first person to have issues with his press coverage, but it seems like he's getting a bit more traction with his complaints. It helps that he has an observably hostile relationship with the press. But the term "fake news" was originally applied to right-wing blogspam/memes that showed up during the 2016 election, which were highly slanted, based on cherry-picked information if not outright lies.

It sounds like their sudden viral popularity created too many obligations to fulfill, but not enough to scale up sustainably. Truly victims of their own success.

I know that stretch goals are popular but there's something to be said for not overcommitting. I have in mind a few other crowdfunded campaigns that were much more successful than initially envisioned, with the bonus rewards/additional content that they promised apparently causing delays or cost overruns for the creators.

I would not dive into GANs right off the bat. Ng's course is a good start, maybe also buy or borrow a copy of Russel and Norvig's "Artificial Intelligence: A Modern Approach"[0]. If you decide to look into deep learning, I've heard good thing about Bengio's book (see also Hugo Larochelle's lectures on YouTube) and if you want to learn more about reinforcement learning, "Reinforcement Learning: An Introduction" by Sutton and Barto is frankly amazing (and there's a great course on YouTube by David Silver).

From reading your code, you're probably familiar with most of the stuff you need to learn ML, so I thought I would take a stab at explaining how you get from "no machine learning experience at all" to "surely GANs will solve all our problems". Broadly speaking, a lot of machine learning is just deciding what your goal is (carve an image, play a game, classify objects, group similar data points, etc.) definining an objective function (e.g., number of sharp edge removed, player score, a classification loss function[1], distance over points in a cluster) and then maximizing or minimizing that objective function.

In classification, you want a function `f(.)` such that given a data point `x`, `f(x) ≈ y`, where `y` is the correct label for that data point. Say you have a bunch of these data points in a matrix `X`, and a vector of correct labels `y`–– how do you come up with a function that correctly classifies the data in `X`, and how can you do it so that data you might collect in the future would also be correctly classified?

You could do this with linear least squares (come up with weights `w` to minimize `|f(x) - y|`, with f(x) = x^T w) but unless your data is cleanly separable, that probably won't work so well. There are a lot of possible techniques people have tried, but lately neural networks have become popular because they are powerful approximators, relatively easy to deploy, and pretty versatile. But the training process[2] still boils down to tweaking weights until `|f(x) - y|` is as low as you can make it.

General adversarial networks are kinda a tweak on this idea-- you start with a bunch of images, and you train a network to generate images that are "similar" to them; call the initial images the "real" images and the generated ones "fakes". You want your network to generate fakes that are similar to the real images. How do you define similar? Well, you have a second network whose objective is to distinguish between real and fake images. You train both at the same time, using feedback from the second network to improve the first network, and using the images generated by the first network as training examples for the second.

Initially, the first network will generate splotchy garbage that the discriminator can easily tell apart from the real images. Over time, the first network improves, creating increasingly realistic fakes, while the second network gets better at determining whether an image is real or generated. After much training, you have two networks: one that is good at generating images that are similar to the seed images, and another network that is good at identifying if an image is real or fake.

––––––––

0. It even has a website and example code http://aima.cs.berkeley.edu/

1. https://en.wikipedia.org/wiki/Loss_functions_for_classificat...

2. See here for a nice explanation of the backprop algorithm, http://neuralnetworksanddeeplearning.com/chap2.html or if you want the OG (original Geoff) paper on the topic, it's in the proceedings for Parallel Distributed Processing 1986 -- "Learning representations by back-propagating errors", Rumelhart, Hinton, and Williams.

It's interesting that so much neural net weirdness emerges from exploiting errors in physics simulators or floating point math. I am now expecting the next generation of perpetual motion machines to include AI to try to take advantage of physics bugs in our own universe.

On a related note, does anyone know how you might go about fixing a simulator that allows collisions to generate more energy/momentum than was initially supplied? Or otherwise violates known invariants?

I'm glad to hear that. The main point I try to get across to people regarding Bellman equations is that they are very special-- these sorts of recursive equations allow us to express the value of an observation without knowing the past, and to improve our estimates of a state's value without having to wait for the future to unfold.

In most other situations you're forced to "wait and see" when you want to learn how a given strategy will turn out. This is not the case if you're dealing with an MDP. If the current state is `s`, the next state is `s'`, and the reward you got for transitioning between the two is `r`, then for a given value function V(.) you can express the temporal-difference error (which is sort of a gradient for the value function) as: δ = r + γ v(s') - v(s) ≈ ∂v(s)

Other formulations of rewards/objectives don't tend to permit such elegant constructions, which is why MDPs are so special (and reinforcement learning so successful).

However I feel like it's a struggle getting that point across, so I'm interested in reading your next post to see how you convey things.

Yeah, so here we can "solve" the problem without using a POMDP by really expanding what constitutes the state space.

Initially, your state is your hand plus the initial rules (who goes first, which suit is trump, etc depending on what kind of game you're playing). After each player's turn you transition into a new state, which is <<starting hand>> + <<history of cards played>>, so at each point in time the state has all the available information. Obviously this leads to a combinatorially huge state space, because there's O(52!) different possible sequences for a standard deck of cards.

Typically you have to use some sort of approximation architecture because evaluating all different possible responses to a given sequence of cards would both take too long and even if you could do it, the memory requirements would be enormous. See for example DeepStack[0], which uses a neural net to approximate the quality of positions + possible responses (along with other techniques), rather than having a lookup table for what to do in each possible game state.

So you can formulate such games as MDPs, but the question then becomes: is this a useful way to think about it? Sometimes it is, because your approach to the problem scales and so you can throw enough resources (training data, compute power) at it to learn a useful strategy; other times there is a better way of formulating the problem[1].

I have some implementations of reinforcement learning algorithms and code for working with MDPs if you're interested in that sort of thing[2], and of course I highly recommend Rich's book.

---

0. https://www.deepstack.ai/ -- note that it's not really an RL system, but it does illustrate using neural nets to learn complicated functions in a large state space.

1. For example, I am not sure that such an MDP for hearts is well-posed: the convergence results that I'm familiar with rely on the MDP being ergodic and aperiodic, but the suggested construction is obviously periodic.

2. The algorithms: https://github.com/rldotai/rl-algorithms An MDP library: https://github.com/rldotai/mdpy (check out notebooks/example_notebook.ipynb for a basic demo).

Consider reading "Reinforcement Learning, an Introduction" by Rich Sutton[0]. It is very accessible (and probably the most used textbook in the field), and goes into detail about the connection between RL and MDPs.

If you want a high-level answer right now about how the two are related: reinforcement learning focuses on predicting or maximizing the discounted sum of rewards, G_t = R_{t+1} + γ R_{t+2} + γ^2 R_{t+3} + ... which is called the return Crucially, the process generating the rewards (the environment) is assumed to be memoryless[1]. So we can reformulate the return as a recursive equation: G_{t} = R_{t+1} + γ G_{t+1}, since the return from "t+1" doesn't depend on the reward you got from the previous time step ("R_{t+1}").

This in turn allows you to define Bellman equations (which define optimal solutions to the problem of maximizing return, which is what you want to do). More generally assuming the problem is Markovian makes things substantially easier to reason about, because you don't have to worry about complex histories leading up to the agent's arrival in a state.

While it's honestly a pretty strong assumption, it's one that we make across many fields and particularly in RL. It's actually not too inaccurate; for example most games are Markovian (or close), and in physical problems (e.g. robot locomotion) if you've got position+velocity then controlling the robot can be viewed as a (continuous-time) MDP.

The weaknesses are obvious: not every environment is Markovian, sometimes there is long-term dependence on the past. For example, translating prose or poetry in such a way that preserves the point across is very tough[2], not really something you can formulate as an MDP. Other times, the weaknesses are easier to overcome (think of a poker game, where you can combine the present state information with the betting patterns from the past; this augmented state should be able to tell you everything you need to know).

---

0. http://incompleteideas.net/book/bookdraft2017nov5.pdf

1. Another way of phrasing this is that each state provides all the useful information, and knowing about previous states does not tell you anything new about the environment. Think of perfect information games, like Chess. It doesn't really matter how you got into a particular position, just the moves you make from there.

2. Compare the poetry translations of Jerry Lettvin with what you'd get using existing translation software (https://sites.google.com/site/lettvingroup/Home/Projects-His...).

Yes.

In particular, it's possible to learn the variance of the return using TD-methods with the same computational complexity as learning the expected value (the value function). See [0] for how to do it via the squared TD-error, or [1] for how to estimate it via the second moment of the return, and my own notes (soon to be published and expanded for my thesis) here [2].

It turns out that identifying states with high variance is a great way of locating model error-- most of the real-world environments are fairly deterministic, so states with high variance tend to be "aliased" combinations of different states with wildly different outcomes. You can use this to improve your agent via either allocating more representation power to those states to differentiate between very similar ones, or have your agent account for variance when choosing its policy. For example, Tesla could identify when variance spikes in its model and trigger an alert to the user that they may need to take over.

Additionally, there's work by Bellemare [3] for estimating the distribution of the return, which allows for all sorts of statistical techniques for quantifying confidence, risk, or uncertainty.

---

0. https://arxiv.org/abs/1801.08287

1. https://arxiv.org/abs/1607.00446

2. http://rl.ai/posts/fun-with-the-td-error-part-ii.html

3. https://arxiv.org/abs/1707.06887

I can't comment on #4 but with regards to 1-2 we can make an argument that high dimensional space is generally friendly towards stochastic gradient descent.

Consider a neural net that produces a single output given `N` inputs– it's basically a function `f(x) = f(x_1, x_2, …, x_N)` A local minimum x* has the gradient `∇f(x) = (0, 0, …, 0)` and an NxN Hessian of the form `[H•f(x)]_{ij} = (∂^2 f)/(∂x_i ∂x_j)`.

The critical point is a local maximum if the Hessian is positive definite at x and a local minimum if it's negative definite; it's a saddle point otherwise. This corresponds to the Hessian having a mixture of positive and negative eigenvalues.

Heuristically, we have N eigenvalues, and absent prior information we can expect that the probability of each eigenvalue being positive is 1/2, and 1/2 for the eigenvalue being negative instead. So the probability of an arbitrary critical point being a local minimum is `(1/2)^N`, which becomes extremely small as N grows large. Large values of N are kinda the bread-and-butter of deep learning, so we expect that most critical points we'll encounter are saddle points. Additionally, the inherent noise in SGD means that we're unlikely to stay trapped at a saddle point, because once you're nudged away from the saddle, the rate at which you slide off it increases rapidly.

So if your net appears to have converged, it's probably at a local optimum with a reasonably deep basin of attraction, assuming you're using the bag of tricks we've accumulated in the last decade (stuff like adding a bit of noise and randomizing the order in which the training data is presented).

As for 3 & 5, we kinda cheat because if your model appears to have converged but is not performing adequately, we step outside of the learning algorithm and modify the structure of the neural net or tune the hyperparameters. I don't know if we'll ever have a general theory that explains why neural nets seem to work so well, because you'd have to characterize both the possible tasks and the possible models, but perhaps we'll gradually chip away at the problem as we gather more empirical data and formalize heuristics based on those findings.

Might it have been Lyrebird[0]? They're (as far as I can tell) they have the best text-to-speech that you can actually use. It's kinda annoying when Google makes these announcements about their advances in TTS only to reveal that it's not actually something you can make use of, and no, the dataset they used to train their model is not available.

0. https://lyrebird.ai/demo/

According to: https://www.rmv.se/aktuellt/det-visar-tre-manader-av-medicin...

I don't speak Swedish, so I had to translate it (the site I found the link from[0] after Googling seems like it has an axe to grind). They indicate that, of ~2400 refugees where the National Medical Board rendered an opinion, some ~2000 were estimated to be eighteen years of age or older, with ~400 estimated to be under eighteen years of age (so ~5/6).

This may be different from the patterns in general because these are cases where the board was asked for and subsequently rendered an opinion. It might be that (contra the headlines) this sort of misrepresentation is relatively rare, and the rate for the subset of the population where no questions were raised is substantially lower.

-----

0. https://nationaleconomicseditorial.com/2017/07/15/sweden-chi...

I am disappointed in the level of fear-mongering pearl clutching; it could[0] get so much worse!

First, the person creating the fakes is using an off-the-shelf algorithm[1], adapted accordingly. A system designed purely for the purpose of swapping faces (or morphing a particular face into others, one-to-many style) would obviously yield improvements.

Secondly, it's the work of one person without commercial backing[2]. A studio that invested in lidar or used multiple cameras would be able to incorporate volumetric information.

Third, it's limited to individuals where you have a performer that's an approximate body double for the target in question. Why involve real people at all? That limits the kinds of scenes you'll be able to shoot, and it's clearly a production bottleneck besides. Since it's feasible to create realistic animations of locomotion via deep reinforcement learning[3], applying the same technique would remove the need for human actors.

With that in mind, we can invent scenarios substantially more terrifying, for which this is but a teaser trailer.

As computation becomes cheaper and AI knowledge more widespread, the delay between when something becomes possible and when it becomes ubiquitous grows ever shorter. Once custom pornography is just another something-as-a-service, the question becomes how do we optimize it? IoT sensors and wearables can provide the data needed to customize an experience tailored to your particular tastes on a moment-by-moment basis. Facebook is sinister, but it's sinister in the sense of a malevolent external force with inscrutable goals; you can defeat it by not using Facebook. The idea of a machine that can produce something more compelling than most stimuli available in the external world is scarier, because all it would be doing is giving you exactly what you want.

Okay, so the riff on wireheading in the style of Infinite Jest might not be adequately terrifying. How about something more realistic, like character assassination campaigns? We've seen how intimations of sexual impropriety can damage a politician's electoral chances or outright cause them to resign. Sometimes, the thing that allows them to hold on to office is the fact that there's no convincing proof beyond circumstantial evidence. However, if you've already primed the public with rumors, a fake sex tape might be the just the thing to push past the tipping point. Imagine: a future where an individual could effect a bloodless coup using nothing more than a Twitter botnet to spread slander and a few GPUs to generate scandalous footage.

No? That's perhaps more in the vein of William Gibson. The last thought that came to mind was along the lines of "the future is weird and terrible, yet somehow boring and familiar" a la Neal Stephenson. The redeeming value of porn is that there's at least some intentionality behind it. Real people had to be involved, from the actors to the crew to the guy who wrote the script in traffic on the way to the shoot. The future will be endless porn spam, half-coherent plots generated by LSTM-char, just glitchy enough to be discernible as fake, yet seemingly unavoidable due to the sheer quantity made possible by automating the production pipeline.

------

0. And when it comes to AI, "could" should be read as "will" but with an indeterminate time frame.

1. Apparently, the face-swapping is accomplished following "Unsupervised Image-to-Image Translation Networks" (paper: https://arxiv.org/abs/1703.00848 video: https://www.youtube.com/watch?v=Bwq7BmQ1Vbc), with the face detection provided by FasterRCNN.

2. The reddit account of the artiste in question: https://www.reddit.com/user/deepfakes (not exactly safe for work, but not particularly unsafe, either).

3. http://www.cs.ubc.ca/~van/papers/2017-TOG-deepLoco/