darkmighty

The real world is expensive. Everyone has a smartphone in their pockets (and most a moderately powerful PC of some kind depending on region), where hundreds of hours of challenge and entertainment can be had for free (or a few dollars). A piece of hardware for entertainment costing hundreds of dollars would need to be absolutely unforgetable (or provide a very lengthy/rich progression) to compete realistically.

I think either it needs to be an amazing teaching tool and/or have other useful capabilities (like merging a roomba/educational robot, or merging educational/house monitoring); or the cost needs to come way down for mass adoption. I believe both have future (my dream is widespread robotics within <$10).

Is the relevant model here simply a sum of 2nd order terms?

It's widely known to practitioners vanilla Neural Nets don't do very well on small datasets because of overfitting issues -- which can probably be corrected with very careful regularization and parameter optimization, but where generally it's best to just use simpler methods.

Polynomial regression is one of those, but I'm not sure how it compares to other methods such as random forests and KNN variations.

Those methods don't scale very well, in particular PR (if I'm interpreting correctly) doesn't scale, because depth is essential to solving complex problems with computational efficiency.

The researchers conveniently refer to those more complex tasks as a 'work in progress'... when they are what NNs are mostly used for (i.e. it starts getting interesting with CIFAR-10). That said, I would be curious to see performance of deeper polynomial regression networks. They become almost identical to NNs then, but maybe some methods of PR could be extended to give an edge in some cases (e.g. perhaps better regularization techniques than dropout/etc).

> why shadows have fuzzy edges

? Shadows have fuzzy edges in geometric optics as well (unless you're mentioning usually micrometer-scale diffraction effects). It's because illuminators are not point-like, so when viewed near an edge your light source progresses in size until it disappears (penumbra to umbra)

https://en.wikipedia.org/wiki/Umbra,_penumbra_and_antumbra

As soon as your photon interacts with a realistic scene (thermal), I believe it loses coherence with any apparatus you might have prepared in the detector. In short, no, I don't think this is possible, unless you consider idealized environments.

Intuition tells me variable extrusion generally should yield stronger parts. Adhesion between stair-stepped and non-planar layers in this method is probably subpar (with the emergence of new gaps). With variable extrusion, as long as the variability isn't too large layer adhesion should be kept fairly unchanged, maybe even stronger if there's a certain waviness (the waves increasing surface area).

Actually I wonder if internal waviness could be a useful method to add strength in general (area should be increased as ~1/cos(angle), or ~(1+angle/29) for very small angles, i.e. about 3% per degree).

> I don't actually understand how Markforged and BASF can print with a material that obviously becomes runny at about 250°C, and yet retains its shape while in the sintering furnace at 1300°C. It would be good to learn more about how this works.

A blind guess: the original filament has a resin to make the metal powder flow, that becomes viscous at certain temperatures. Perhaps this resin could be volatile and evaporate at temperatures below its viscous temperature.

Or maybe it could be quickly burned off by supplying extra oxygen (so that burning temperature isn't high and there's not enough time for the piece to deform).

In any case you could probably introduce some kind of automatic software correction for the deformation to the piece. As noted in the article this has some limitations/drawbacks, like limiting some geometries you can achieve (and requiring careful knowledge/control of your process).

The use of hyperbolic space to encode hierarchies is absolutely genius (in not small part because it should be obvious), given the properties of hyperbolic space to "expand" around fixed points exponentially, exactly like trees do* . It seems like a continuous generalization of trees essentially. Beautiful.

*: https://en.wikipedia.org/wiki/Hyperbolic_space

"Another distinctive property is the amount of space covered by the n-ball in hyperbolic n-space: it increases exponentially with respect to the radius of the ball for large radii, rather than polynomially. (...)"

The first versions used what you might call 'camerahack', it just sees the entire map at once but as the player himself would see (i.e. fog of war included). The latest version has camera control, only seeing one screen at a time like humans.

Negative comments are usually much less useful also. They are usually one of those:

1) Animosity (the person dislikes you)

2) A broad description of a mistake

3) Different taste

Only (2) can be somewhat useful, although (3) is permissible sometimes I guess. Often the person already knows he might be doing something wrong or not good enough; simply pointing it out leads nowhere, knowing what not to do still leaves unancessibly vast possibilities.

It seems much better to directly (and probably privately) suggest alternatives, offer solutions, etc.

I like this opinion piece about this:

http://fanaro.com.br/negative-language-why-you-should-avoid-...

Diverging should be discussed under some circumstances though, specially if you clarify your rationale (or when discussing unconsequential things like foods and such).

Overall pointing out solutions and affirming good values, qualities, outcomes if much more useful (and has the bonus of being charitable, even affectionate).

Sleep is a very technical thing, it's a 'Brain maintenance time'. It's not about saving energy or anything that could/should be bypassed (otherwise we'd be tired awake at night, useful for defending against predators in the wilderness -- nature accepts a great cost for this shutdown). I find my memory improved significantly once I started being more careful with my sleep, for instance. I wish I had come into contact with this instead of the bragging 'I barely need any sleep!' culture.

Emissivity is mostly what matters in this case.

Whether the material is reflective or transmissive just means you'll be "seeing" IR sources reflected off the target or from behind the target, respectively. You do have to keep that in mind if there's an object not in ambient temperature in those locations (e.g. a clear sky is of very low temperature, a person is around 35C).

Indeed, but note it usually works well enough, because most surfaces have high IR emissivity (in particular with organics like plastics, papers, woods, fabrics, etc.; soils and liquids as well).

The one exception you have to be careful is with unpainted metal (high reflectivity). If polished it can be so low that even calibration doesn't help -- because you'd need extremely precise calibration that isn't practical and because reflections and noise/other heat sources will pollute you readings. In that case a simple sticker (or thermocouple reading) would be more adequate.

Yea some state of the art results are just inaccessible without large budgets, simply because models can scale (and because some orgs have a lot of money to train those scaled models).

You can always just use smaller models and/or lower resolutions though; of course the results won't be on par but it may reach a qualitative result (for research and experimentation purposes) or good enough result (for personal application purposes). E.g. hobbyists don't need AlphaGo-level go playing AI (which I'm sure had aggregate costs in 5 figures or more to train), reduced versions play all far above our levels -- although in this case there's the interesting effort of pooling hobbyist resources to indeed reach SOTA, see LeelaZero[1] and LCZero.

Some kinds of research will be effective only at large orgs, that's always been true. There was indeed a brief period when people realized GPUs could unleash deep learning/CNNs that you could do anything with a good GPU, but that was very much an exception. To borrow from another field, you cannot do a level of car engine research without all infrastructure to fabricate and test engine prototypes (though you can do some/other kinds of theoretical analysis).

[1] http://zero.sjeng.org/home

This is almost all of the software we use, I guess :) (that isn't strictly for entertainment)

The most notable productivity boosts/cognition boosts found so far have got to be Google [search engines] and Stack Overflow. Yes you could use books but the speed is incomparable. I can accomplish something extremely complicated that would require lots of thinking with just a few queries, reading solutions and glueing them together. I can get expert advice on anything instantaneously. Interestingly this augmentation is mostly 'oracular', just giving you the right answers on demand. Though sometimes the rationale behind an answer is well explained and you can somewhat incorporate the general solution model.

Should we value more augmentation devices that are less oracular? E.g. maybe a high level programming language is a non-oracular augmentation device; you still have to correctly encode all logic, only now you can glance over some details (in sacrifice of some performance).

Do we risk of atrophy from this practice of getting everything almost ready? The productivity increase would indicate you can just use your time to accomplish more; but we all know sometimes "wasting" time is needed to learn or maintain useful skills (e.g. you could always use a calculator, but we chose to learn arithmetic; you could use wolfram to solve integrals instead of learning calculus; we spend time exercising).

It does sometimes feel like I'm almost a god at the computer, and a naked mortal away from it.

Do you have any purchase advice/recommendations?

There are several kinds of procrastination, so it's impossible to generalize.

Some projects require very large time commitment to get right, reworking several times, etc. Some others reach a diminishing return stage very quickly and continued effort is wasted. You definitely shouldn't just take whatever you have in front of you and just rush it until it's barely workable.

The critical skill is to be able to judge without anxiety and with some precision how much additional time is going to improve the outcome, and what is the rough global value (to you) of this improvement. Some projects have long ramps of effort and refinement, some are just not worth the time and can arrive at a good enough stage very quickly.

I think the best advice I could come up with is that if you're having this sort of problem, recognize it, try to deal with the anxiety and get a perspective on the risks vs benefits, and improve. If possible find an environment/hobby /side occupation where you have to make a ton of such decisions to get a lot of practice. College/personal projects tend to be pretty sparse and give a lot of room for failure and delaying I guess.

For me it's been a very slow, progressive realization that I'll never get things I wanted done by just leaving them to an arbitrary later date, or obsessing too much over details. In nothing vs something, something wins.

I don't see any figures comparing the error of the formula with the previous error from numerical methods. While mathematically the discovery seems significant, it isn't clear to me this will actually improve lenses.

The coldest parts usually have double doors at the entrances right? So temperature doesn't really swing so much from people transit. Beside, office buildings are usually extremely large compared to volume of exchanged air at the entrance.

A fun dream I've had (context: I've always been confined to apartments) is some sort of cat/pet utopia in a block, so they wouldn't be stuck inside alone. But that wouldn't be feasible since they have to navigate the fire-proof doors between floors...

Unless your cat door goes to market. May luck be on your side :)

(really, don't underestimate niches, you can probably find plenty of places to sell this sort of product...)

There's such an elegant solution here: cups could have individually tailored patterns (or chosen from a large library of a few 100s) on the outside s.t. you wouldn't confuse yours for another. We're (at least I am) pretty good at memorizing combinations of odd shapes and colors, so there's no need to name cups (i.e. they can be pre-named).

Europe, and to a certain extent the, US seem to be doing fine, thank you very much -- without need for artificial scarcity or dictatorships.

Europe in particular, isn't, for the most part a "Failure of capitalism" -- people are living well enough, much better than in China for instance.

The problem here is that even though by most humane standards the average european or american could be better off than the average chinese, perhaps exactly because of this China will achieve more competitively. Maybe optimum standards of living doesn't coincide with optimum productivity. What then?

That's a problem for the US in particular because its core justification, so to speak, [for the American social-political-economic system] was this entire premise: 'More productive, more competitive, means ultimately means better in every way'. 'The US was more productive than the USSR, therefore Socialism is bad'; 'We win in competition therefore we have the greater future'. But if China, as a an authoritarian surveillance state is more productive than the US, how to sustain this thesis?

And how to deal with China if they may be bound to outgrow their peers in the future? What if they start spreading this cultural modus operandi justified precisely by economic productivity?

For me the answer is obvious: this thesis must be abandoned. There are two ways to deal with an enemy that is willing to sacrifice humanity in the name of slightly better performance.

(1) Through international law: make agreements and regulations demanding standards of living, democratic minima, etc. that apply globally.

(2) Though conflict/direct competition. If an authoritarian hegemony is expanding and threatening our core values, and there is no other way to stop, then you should do the necessary to stop them, degrading temporarily the standards of living to outpace them and override their doctrine in the expectation to restore to a more humane program afterwards. I don't condone the practices of the USSR (it was a very authoritarian and problematic regime on its own), but this is mostly their rationale through the Cold War; this aspect of their ideal at least seems very reasonable to me.

Let's hope that either

(a) China's approach ultimately doesn't actually show to be as effective as it seems;

(b) They fail to expand their regime/doctrine and collapse from internal democratic pressure (unlikely as control/surveillance technology expands);

(c) International treaties control the situation and enforce better human rights

Before the West needs to enter a conflict.

Note how this goes against Nature: we'd be refusing to let the most 'fit' nation out-compete and dominate others through some global coordination, and try to enter a suboptimal (in the competitive sense) quasi-steady-state. It isn't clear we can achieve this, but this would a most beautiful achievement of humanity. Finally taking destiny into our own hands free from the current make of A Blind Watchmaker. Being blind, those Watchmaker very often fail to prevent disasters stemming from their work, another significant sign of its troubles: with our inability, but also necessity, to achieve global coordination in sacrificing maximum naive development speed (but helping humanity's environment in the long run).

https://news.ycombinator.com/item?id=17572185

Just wanted to say thanks. This is a little piece of magic. That's what technology is for, I guess.

Well, it should be possible to hide components inside the PCB, see:

(1) https://hackaday.com/2017/09/11/the-components-are-inside-th...

(2) https://hackaday.com/2019/01/18/oreo-construction-hiding-you...

Analysis should be quite difficult; if you hide them between copper layers, or worse behind components and behind copper layers, I believe it'd require vert inspection with x-rays to detect.

The main question is that of utility. Data exfiltration is highly unlikely. It's possible to trigger malfunctions, but it's not clear this would be useful (when triggered at a certain date, randomly, etc) -- it would be easy to detect and cause the companies to switch to suppliers. I think the main possibility/atractivity would be some kind of radio activation of malfunction. But then you need a significantly sized antenna [1] and a very strong signal to penetrate the circuit shielding (which is probably emp-resistant).

But the main impediment I think is the total erosion of trust in Chinese manufacturing this could bring. Most of the world manufactures in China and they don't want to change this. They wouldn't sacrifice this trust except at a very good opportunity. The same goes from a strategic spying perspective: don't use bugs frivolously or you'll risk raising the awareness level compromising important opportunities.

[1] Which cannot be hidden between layers, although it might be possible to use some of the traces themselves as antennas. If the pcb has an obvious external antenna I/O then of course this would make things easy.

All that said, it does make sense to keep an eye on this possibility, but probably not in this particular case.

> Gradient optimisation gets stuck in local minima, and it will always get stuck in local minima, and therefore always be dumb as bricks. Which is evident in the way deep neural nets overfit and are incapable of extrapolating outside their training datasets.

This seems extremely harsh, given the vast successes of deep neural nets calling them 'dumb as bricks' (there are too many examples of extrapolation outside datasets to count[1]). They have proven generalization capability when care is taken against over fitting (almost any model has this ability). Also there are results showing getting stuck in local minima is unlikely in highly-dimensional parameter spaces (the curvature in a large number of directions needs to coincide), and in any case rarely you're seeking the global minimum because of said over-fitting issues.

[1] It should be self-evident because they would hardly be of any use if unable to extrapolate at all. See GANs (https://www.thispersondoesnotexist.com/), AlphaGo (https://en.wikipedia.org/wiki/AlphaGo), etc.

> Note also that you don't need something to be Turing-complete for it to be a program. For instance, a regular automaton is a program, but it's not Turing complete. Turing completeness makes something, well, a Universal Turing Machine, which can compute any program. But a program is just a set of instructions.

A regular automaton is a program (and a function is a program), but all programs cannot be expressed as finite (regular) automatons (FAs). So FAs are a subset of programs. If we equate programs with mathematical algorithms, in general they really need the unlimited memory aspect (or at least the capability of abstracting memory, which in practice is always finite of course).

To make things clear, an example: a fixed function (or FA) cannot sort a list of numbers of variable size. You can train your function to sort any K numbers, which are the input variables. But because it has no memory, you cannot input variables sequentially (in the FA case you cannot input arbitrarily many variables). You've essentially created a sorting network[1].

Not only that, but for each input growth you need to re-train your architecture. The nice thing about program inference is that it hopefully captures the fundamental algorithm behind a program, which generalizes to arbitrary sizes. It's not that arbitrary sizes are necessary in practice (again computers are always bounded), it's that this abstraction/generalization is extremely useful, saving data and allowing large (arbitrary) variation in input.

[1] https://en.wikipedia.org/wiki/Sorting_network

> a Universal Turing Machine, which can compute any program

I feel there's a bit of confusion here. Turing Machines themselves can compute any program (i.e. you can represent any program as a TM) -- that follows essentially from the definition of algorithm (a set of well defined steps given unlimited paper); again languages accepted by FAs are a restricted subset of TMs (the former accepts a regular language while the latter recursively enumerable languages). There are special Turing Machines that simulate arbitrary Turing Machines -- those are called Universal Turing machines. It's a desirable characteristic of any computer, of course (the ability to input arbitrary programs).

We refer to program description languages that can describe any Turing Machine as Turing-complete. You could make a Universal Turing machine that accepts this kind of language directly or you could in principle just build a Turing machine that executes the given algorithm (for example using a gate array/FPGA to encode the state automaton, plus some large memory for the tape).

> In fact, I'd go as far as to say that this is a new kind of Inductive Programming.

I wouldn't. Traditional NNs are widely called funciton approximators, or function induction. That's because they're not Turing-complete, and in this case don't even have memory. This makes them decidedly not programs, just functions.

Of course, they could be modified with recurrence (a suggestion of future work), or even more complete memory systems (e.g. equipping a neural controller with a tape), making it Turing-complete.

Keep in mind the functional approximation power isn't increased by architecture search alone (there's no restriction from doing weight training). What seems novel here is how significant architecture alone is (I wouldn't guess it being this powerful), and this possibly has implications on the connection with the human brain -- in which my vague impression is that connections themselves are much more tuned than weights; also possibly giving greater emphasis on architecture training.

> Wealth = freedom to do what you want

Once you're wealthy, kind of :) (there are some complications and the fact that often the main limitations of doing what we want are usually not wealth, but not knowing what we'd want the most in the first place, or lacking courage/skill to pursue it)

But there's the process of getting wealthy, which can include self-restriction of freedoms...

> - Earn as much as you can from your own work. Take a 2nd job, change to a higher paying job, ask for more responsibility and a raise at your current job, go back to school for a more lucrative degree, ...

I think that's a much better advice than any I found from a cursory read of the article; at least with my limited experience with friends. Most of my acquaitances doesn't seem sufficiently agressive with salaries -- they're mostly worried about having a job, a "good enough" pay, and that this job doesn't suck too much. That's ripe for exploitation -- if you don't actively demand and seek the best deal, you're going to be paid the minimum you can to just scrape by and definitely won't get wealthy. You don't have to own a business to be relatively wealthy (I'm taking 'wealthy' to be relative because the threshold for wealth is indeed much higher in developed countries).

Also I think being wealthy is completely beside the point; it doesn't make too much sense as an objective. The article tries to make the point for wealth being ethical; but in any case ethical are possible consequences not the accumulation of wealth. Health, general satisfaction (including from making ethical choices) are what really matter. That does necessitate some wealth, but in general any wealth threshold is not necessary nor sufficient for those prime goals.

Those two points may seem contradictory, but it's not because wealth is a non-necessary correlated variable that you should just leave it at the table. Get all you can from your employer without compromising on what matters (and hence obtaining the conveniences and securities of wealth while retaining life satisfaction).

I'm not going to venture into defining life satisfaction, but I do think it's good to ask yourself from time to time. For me it is having decent amounts of free time, being able to help people, having fun socially, producing beautiful/impactful work, etc. -- but it could include whatever you deeply enjoy.

I recommend this (well fundamented as usual) post from SSC:

'Basic Income, Not Basic Jobs: Against Hijacking Utopia.'

https://slatestarcodex.com/2018/05/16/basic-income-not-basic...

I think the main argument should be fundamented by charitable assumption, that most people actually have decent plans for employability and could use the available time much better than being stuck on a Job Guarantee (that could have very low marginal value to the employer anyway).

Moreover, if power becomes significant you can add regenerative breaking to regain a large fraction of energy (if they're not doing so already; maybe not because of the age of the trains).