> They think for a minute and reply: “by policy, you cannot purchase a digital camera through the procurement office. Therefore, your computers do not have digital cameras in them.”
They were being intentionally daft. The person who told him that almost certainly did so with a wink and a nod, and saved your protagonist a lot of pain.
> if you're working on gene editing, nuclear fusion, or some such radical and ambitious project, we want to hear. Tell us the path you took to get to this point, and your ideas for the future. Then select from everyone who writes in?
You are describing a scientific conference, or perhaps a grant proposal.
Unlike NSF/NIH/etc., they're not really asking for people who have dedicated their lives to a problem to write up a compelling and realistic plan of attack for a hard tech problem. That can't be done in a few minutes of a video, and certainly can't be done if in that video you also have to write your ideal obituary. What they're looking for, I think?, are charismatic and driven leaders.
When people get used to a comfortable life, a predictable job, and a reputation of succeeding at whatever they do, it gets very hard to leave that behind
I think that there's a sampling bias here; in fact, that bias is even pointed out in the paragraph: > At YC, we’ve often noticed a problem with founders.... I.e., this is an observation about FAANG employees who become YC founders.
But the fact is, very few of those drive/motivated Google/Facebook employees leave and become YC founders. So this observation is an observation about a very small subset of the people who choose to stay at GoogleFacebook.
I know a lot of people at the FAANGs who are there because they are primarily motivated by technical and scientific problems. The FAANGs -- and esp their research divisions -- are exceptionally good places to work for people who get meaning and fulfillment from technical excellence.
It's not pathological. Most universities have flirted with hiring non phds to teach. It can work, but there's a reason they prefer traditional candidates: thebfaikure rate is way lower.
the competition is often between people with phds. i am not 100% sure that a university teacher needs to have a phd, but it seems that even keeping that requirement, the present system is pathological. most people with phds do not make the cut for researcher, which leads to the university being short on teachers.
There's not a ton of overlap with boot camp grads. At least here, only the bottom barrel of CS grads even apply to positions that are good fits for boot camp grads.
> There's not a ton of overlap with boot camp grads. At least here, only the bottom barrel of CS grads even apply to positions that are good fits for boot camp grads.
That is not true at all. Maybe if you are in the upper-echelons at the Stanfords, MIT, and CMU's of the world doing advance research in AI/ML/Robotics/Security but for the most part most CS grads are going to school to get a job -- I was one of them.
> I'll be honest, I'd probably trust code that's gone through DO-178C more than I would trust code that's been formally verified.
Well... depends on the company/team :-)
But I think the instinct is fundamentally reaosnable because even the most abusive interpretations of DO-178C tend to demand a lot more attention to quality than even very well-run software projects in other industries.
> For the most part cars and medical devices and stuff don't use FM because there's pretty much no oversight.
This is starting to change, maybe. The automotive companies have been snapping up verification people (and letting them work on verification).
> But any FM stuff has to go through DO-330 and DO-333, which are just as rigorous and intense as DO-178C. The only stuff that has so far is, unsurprisingly, very expensive and very proprietary.
So, it's been a while, but I seem to remember DO-333 being written in such a way that if Tool X went through the process, and if you could produce a deep embedding of Tool Y's logic into that Tool X, then you could piggy back on tool X's certification by using the Tool X embedding. Which is still a lot of work, but hopefully substantially easier than getting a new tool approved from scratch. Also, I could just be mis-remembering or confusing a dream for reality... its been a while.
> The really simple answer to this question is: because at a high level of abstraction, it isn't useful. Most applications are being specified by people who don't understand in any detail what they want the application to do.
The world is changing.
In high-growth-potential areas, I think we're starting to see a shift away from the style of software where "what to build is hard but how to build is easier" and toward application domains/products where exactly the opposite is true.
IoT, next-gen factory automation, robotics, ADAS, automated trading, etc. all have two extremely important things in common:
1. It's really obvious what we want the system to do (minimize energy usage; produce widgets without killing Joe the machinist; go to the store and don't kill anyone; make money; etc.)
Unlike with social apps/online communities/productivity software/games, the "what did the customer want" problem is likely not THE key failure point.
2. It's extremely non-obvious how to map those high-level desires onto software.
These application domains, which I think will drive growth in the software industry during the next cycle, are kind of dual to our current ones (web and mobile apps).
So, at the highest level of abstraction we know what the customer wants. That's not the point where we need the most derisking. Also, the hardest problems -- the ones where derisking efforts will land -- are amenable to mathematical treatment. Because of this shift, mathematical modeling tools are going to play a big role in the industries that (I think) will drive growth in the coming decades. And that means the zeitgeist of software engineering will change.
I'm not sure if those mathematical modeling tools will look anything like today's theorem provers and model checkers. Probably only in the same way that Python looks like FORTRAN.
As a counterpoint to that, another change happening in the industry now is the increasing tendency to delegate more and more to ML models, which are even less amenable to formal verification than mainstream PLs.
You can prove the ANN works, but how do you check it conforms to any specification, other than by testing and accepting high probability of real world failure?
Eventually, you have a high level fuzzy requirement which is hard to even specify right. Just look at, say, latest YouTube guidelines about content they try to ML around to make tractable...
And that's a classifier, now try to do the same with a machine translation system.
>> It's really obvious what we want the system to do
>Nope.
Hedge funds and trading firms are some of the top donors to/supporters of the formal methods community. They sponsor conferences, hire tons of formal methods experts, buy products and services from formal methods startups, and develop/use formal methods internally.
So...
Yup.
(I do understand there's a substantive point being made, and I think my reply to the sibling comment by temporal addresses that point. So maybe that's the right place to continue discussion.)
I'm not buying it; I believe the problem is more on the human side.
"Most applications are being specified by people who don't understand in any detail what they want the application to do", because most humans absolutely suck at clear communication and clear thinking. They're not aware what they're not communicating. They're not aware what scenarios they're not considering. Examples abound - remember that time where your spouse/parent asked you, "go to shop and buy milk", and you knew you now had to drill them for things like "when", "which shop", "which milk (%/brand)", and "what other %/brand is OK if the one you're looking for is sold out", because they'd be unhappy if you did the wrong thing? Most people who don't work with actual STEM, programming or law seem to be like that, almost completely unable to articulate in a clear and complete way, and completely unaware of how much they're not communicating.
> 1. It's really obvious what we want the system to do (minimize energy usage; produce widgets without killing Joe the machinist; go to the store and don't kill anyone; make money; etc.)
This is not "obvious" and is a good example of an unclear specification. The "not kill Joe" may be an implicit assumption shared by all moral people, but at some point it'll have to be quantified, e.g. for risk analysis, and the question "how low probability is low enough" will pop up. "Minimize energy usage" is limited by goals and tradeoffs that need to be stated; I can make you a sensor run a decade off a coin battery, but there's a good chance that's not what you want. Etc. Etd.
Circling back to the GP's statement, does the customer who wants the energy-efficient, safe, money-making contraption understand enough about either requirement to do a formal specification on it? I believe they do not, and it's the same situation as with cat sharing apps and all the other software of today.
I've worked on software for internal use in manufacturing, and number #1 problem I had was that the customer had a bunch of unclear ideas (ideas that I, a lowly developer with zero experience in the domain, had to fix for them on the whiteboard during meetings many times, because thinking clearly is not a common skill, and they did not have it). A lot of consumer-grade software try to get around this by giving up on clear thinking altogether, and just do gradient descent on customer demands (as perceived through A/B tests) in order to maximize money made, regardless of whether the product ends up being useful or not. I don't feel industrial software lends itself to such method; hell, I hope it doesn't - I'd prefer the firmware running in factories to be effective, not addictive.
You and twic make similar points, which I think are reasonable but miss a very fundamental difference between e.g., ecommerce webapps and something like a robot.
The how and what questions in robotics/manufacturing ground out in formal methods style mathematical modeling in a way that the sort of questions that dominated the last couple cycles of tech did not (because those tended to be behavioral specifications about humans -- click more ads, book more taxis, click the buy button, etc).
Formal methods didn't capture "engagement" or "click-through rate", but they do capture "safety" and "liveness".
Consider any of the examples you give. At the top level of the design process, we need to derisk things that demand mathematical modeling (battery life, car behavior). How do you go about determining the probability of a car colliding, or tease out the various things that might inform that probability? Not (only) by white boarding with business stakeholders or asking consumers, but (also) by building models of how the system behaves and analyzing those models.
Obviously, formal methods aren't a silver bullet, and yes, software engineers will always spend a lot of time at whiteboards and in meetings.
But when it comes to the future of formal methods, the question is whether the systems we expect to take center stage in the software world look like to sorts of systems where properties like "safety" and "liveness" are a) hard to specify, b) lend themselves to mathematical analysis, and c) are critical to business success.
In particular: ...[he's] welcome to keep shouting that tools won’t help, better languages won’t help, better process won’t help, the only thing that’ll help is lots and lots of unit tests. Meanwhile, we’ll continue to struggle with our bugs, curse our model checkers, and make software engineering just a little bit better. We don’t believe in silver bullets.
Again, your second paragraph already kind of agrees with this point, but the brilliant rhetoric in the essay is actually true: "no silver bullets" is a reason for, not against, the use of formal methods (and any other quality assurance mechanism we can get!)
> adding a layer of additional effort like Formal Methods necessarily either (a) increases development time significantly or (b) exponentially increases headcount.
It really depends. Especially in the case of formal methods for specification (as opposed to implementation formal methods), formal methods can save substantial amounts of time and manpower. And there are already a lot of success stories, and not just in safety-critical industries.
Even a sedan will do if you just want the bed. Not ideal, but definitely possible. Used to haul drywall, gravel, lumber, etc. in a little trailer hooked up to my freaking Corolla.
Trucks are great because the extra degree of freedom is a PITA, but unless you're towing serious loads they are mostly just the country version of a lexus or BMW.
> The current state of the art of ML assumes nonlinear relationships between all parameters. It can't assume simpler & reasonable models, and therefore it can't extrapolate easily with reduced data.
I'm not really sure what "low-dimensional intuition" means, but I pretty regularly build models that do not "assume nonlinear relationships between all parameters".
The common theme of these problems is that when the dimensionality increases, the volume of the space increases so fast that the available data become sparse. This sparsity is problematic for any method that requires statistical significance. In order to obtain a statistically sound and reliable result, the amount of data needed to support the result often grows exponentially with the dimensionality. Also, organizing and searching data often relies on detecting areas where objects form groups with similar properties; in high dimensional data, however, all objects appear to be sparse and dissimilar in many ways, which prevents common data organization strategies from being efficient.
As the dimension d goes to infinity, if you sample integers from d-multivariate normal distribution, your samples will collect near the unit d-sphere; whereas in low dimensions they'd collect near the origin. This is because as d goes to infinity, your samples become more and more far apart. This is a topological intuition of why higher dimensions don't work like lower dimensions.
I reserve the possibility that I am completely mistaken, in which case I apologize in advance, on the condition that you refer me to an actual calculation or derivation (not just a quote in the same spirit) in a paper or text or textbook.
I have seen such insinuation multiple times, that samples "collect" near the unit d-sphere when samples are drawn from the unit d-sphere in high d dimensions.
From a physics perspective this is very familiar to me, but not in the sense of fact, but in the sense of misinterpretation.
I do believe the observation is very useful in the educational sense as long as it is pointed out as being a paradoxical illusion. In this sense I can appreciate (even encourage) a professor or a TA showing this phenomenon, on the condition they finish up by explaining why this seems to be the case, but is nevertheless a misinterpretation, and they should make sure the connection with Jacobian determinants etc are made clear.
Consider a normal distribution of any dimension (as high as you want), but I will showcase the phenomenon even with low dimensions (here merely d=3) to illustrate this has nothing to do with high dimensions.
Clearly the probability density is maximal in the center of the distribution.
In computer processing of data points, we typically loop over points, calculate some hopefully interesting function on each sample, and then plot the samples say by binning with equal bin sizes. A programmer typically disregards transformation properties like the Jacobian determinant. Suppose the value we calculate for each data point is the absolute length or distance from the center. The further we go from the center the smaller the probability density of the normal distribution becomes... but the larger the volume of a shell of radius r!
Since we are binning with equal bin sizes (equal length intervals per bin), for small lengths, then even though the actual probability density is highest near the center, we will get relatively few samples because the volume under consideration is small compared to the volume under consideration for a shell of a larger radius of equal thickness (area of a sphere grows quadratically with radius). However for even larger distances, the exponential decay of the normal distribution dominates and the number of samples in highest radius bins will decrease again. So in between there will be a peak.
This explains the fact that [ the probability density of [ the absolute distance from the center over [ the sample points ] ] has a peak at some non-zero length.
But it is a conceptual mistake to interpret this as if those sample points in the original d-dimensional space form a dense shell on some "unit sphere" ... This is a pure illusion which illustrates the interpreter is not familiar with jacobian determinants etcetera.
Consider volume or triple integrals over some volume element dx dy dz and for symmetry you prefer integrating in a spherically symmetric coordinate system, theta,phi,r then you can not simply replace dx dy dz with dtheta dphi dr, you need to use dV = dx dy dz = r^2 sin( phi ) dtheta dphi dr.
It is this necessary factor that is ignored when processing sample by sample and causing this illusion in the AI community. I did not follow conventional machine learning courses, but given the learned language used whenever I see statements to the effect of samples lying near the unit sphere in high dimensions, I can only conclude it has its origins in 1) direct observation or experience of plotting in bins of the length of the vector without guidance in interpretation; or 2) guidance during education, where the phenomenon is shown, and the origin of the paradoxical illusion or confusion adequately explained and then the illusory nature subsequently forgotten or 3) a teachers assistant having gone through 2) and showing the phenomenon to students without emphasizing the illusory nature of the misinterpretation.
But perhaps I am wrong, and the normal distribution in high dimensions actually has a higher probability density near its "unit sphere" if the dimension is beyond some critical dimension d_c... but again, I'd like to see a derivation showing it :)
EDIT: For more precise language: they do "collect" (reach a peak) at a certain non-zero length or distance, but they do not collect to a unit sphere of such radius in the original space!
It seems you're mostly arguing with the description of "collection", because to you it implies high density in the full n-dimensional space. But the point of the curse of dimensionality is very much that most points do not lie in regions of high density, because regions of low but still non-negligible density are so much larger. If you prefer, you can say that they are "close" to the surface of a sphere, without implying high density.
There are also descriptions of the curse that do not involve spatial analogies at all. Assume that the data is independently identically distributed along each dimension and is an outlier if it's sufficiently far along one dimension, which happens with probability p in the one-dimensional case. Then in n dimensions, the proportion of outliers is 1 - (1 - p)^n -> 1 for n to infinity. Most points are outliers along at least one dimension.
I am not arguing with the word "collect", as it can be interpreted in best faith.
I DO argue against the "collect near the unit d-sphere", if it does not come with an explicit pointer or reference to the explanation of this illusion, I don't care if one points to Maxwell Boltzmann speeds vs velocity vector distribution, or Jacobian determinant, but one should point to something else communication makes no sense. We communicate to teacch and learn. Only when explaining why there appears to be a sphere in the higher dimension, and how this illusion arises is communicating about this pseudosphere justified. The unit d-sphere makes no sense on the 1-dimensional axis of absolute length on which we project the samples. A reference to a "unit d-sphere" only makes sense as residing in the original d-dimensional sample space. But in that space there is absolutely no packing of samples near the peak radius as it appears in the distribution of lengths.
I was not responding to the "curse of dimensionality". I show that this effect already exists at low dimensions, and physicists are very acquainted with it because within their first years of university study they get drilled in 1) jacobians for non-linear coordinate transformations 2) the velocity distribution of molecules in the kinetic theory of gases, where there is a similar plot for the absolute speed (Maxwell-Boltzmann) distribution [0] showing a peak at a non-zero speed, even though the velocity vector distribution is a normal distribution... Every physicist worth his / her salt, immediately recognizes the phenomenon as relating to the Jacobian determinant, that this has nothing to do with velocities aggregating to some sphere of non-zero radius, and that this is a misinterpretation of magnitude distribution plot...
Clearly in this physics example d=3. Would you say d=3 already shows the curse of dimensionality? I call bullocks, and suspect a misinterpretation of the magnitude plot...
Again the peak in the magnitude plot is very real, any reference to a sphere with radius of the peak in the magnitude plot and residing in the original d-dimensional space is purely imaginary!
Of course the person with simultaneously avera height, avg weight, avg income, avg capital, avg age, avg ... is very rare, ... but less rare than a similarily accurately specified person with non-average weight, but all other variables still average...
I think most of the time when the "if you sample points [at random] from an n-dimensional sphere" line is written/told, "using uniform distribution, where each unit volume has equal chance of being picked" is substituted in the place of "at random", and not normal distribution. I think with normal distribution you are right, and GP is mistaken, but with uniform distribution it isn't just an illusion.
That's just a d-dimensional cube of uniform density and zero probability outside of this d-dimensional cube, there is no meaningful d-dimensional sphere though...
you seem to be referencing the fact that repeatedly convolving uniform distributions approach a gaussian distribution
I agree it's beautiful and concise, but it is not correctly explained.
I can give you a beautiful and concise explanation of clouds too: as we all know objects fall back to earth, or more properly speaking they follow orbits, close to the surface they seem to follow parabolas, but in fact they follow ellipses. the moon orbits the earth in a nearly circular ellipse, as do geosynchroneous satellites. When water evaporates from the ocean they are in fact launched into an ellipcitcal orbit, however each elliptical orbit that intersects the surface of the earth will intersect it again, this explains rain, the individual water drops fly for a while until they fall back down...
beautiful, and relatively concise in comparison with real cloud physics, but totally wrong of course...
My understanding is: It's a fancy way of saying "lots of variables", because a lot of these problems are converted into linear algebra representations (which can have geometric analogies, hence: dimensions).
Meaning that you can run PCA and the first few components will explain a vast majority of the variance. Some things may have so many moving pieces that it's just not possible to grok how all the variables interact to get the results we see.
They were being intentionally daft. The person who told him that almost certainly did so with a wink and a nod, and saved your protagonist a lot of pain.