I wonder how many IQ test misclassify intelligence? Feynman had a score of 125 and many people here have a high score than that. I bet no one here with a higher score of 125 are actually smarter than Feynman though. I think this is just one of many possible such examples where IQ test miss the purported people they were suppose to capture. Feynman’s score wouldn’t even qualify him for MENSA.
I wonder how many IQ exams misclassify the intelligence of smart people with disabilities of some sort or high functioning autism. Not saying any of this applies to Feynman, he is just an example of where the test didn’t accurately capture his intellect.
I'm around 150, and I don't consider myself smarter than Feynman. And I'm no savant, am socially well adjusted, and not on the autism scale. This, among many other data points, has led me to question the entire concept of IQ.
The measurement of IQ borders on pseudo-science. It measures something, but what exactly? Everyone has a different answer. Ok, sure, intelligence. Now try to define "intelligence" empirically. What isn't a particular IQ test measuring that is relevant to intelligence? Different IQ tests often yield different results. And I'm not talking about online tests, but proctored tests administered by professionals.
How do you account for that amorphous quality known as "intuition"? You can't measure it or quantify it. You can barely define it. It is a finicky, inconsistent beast. Yet it has a profound effect on one's problem solving ability.
If the factors that determine IQ are individually difficult to define, differentiate, and measure, what does that say about the score itself?
In the broadest terms, IQ matters, but not in a way that makes one's score relevant for any practical purpose. You don't need an IQ test to measure competency in a specific task or occupation.
FWIW, IQ is not pseudo-science. The research around IQ is some of the most robust psychological research that exists. In fact, it is one of the few areas of psychology that has basically been unscathed by the replication crisis.
You have to engage in some really convoluted thinking to have something like IQ, which significantly predicts a whole bunch of important life outcomes, and somehow convince yourself that it doesn't have sweeping implications at the individual and societal level. It would be like convincing yourself that height had nothing to do with success in basketball. Surely it's not the only important factor, but it makes a huge difference.
Why are you so sure that IQ isn't just a useful fiction that masks a large number of underlying factors that we don't fully understand? Why are you so sure that it is the best predictor of success in these fields. Perhaps other traits are equal or better predictors of success in a particular field, like tenaciousness, or conscientiousness, and perhaps IQ doesn't matter as much if one of those aren't present.
Perhaps you have some citations to share?
The comparison to height is quite weak, because height is a far simpler and more objectively verifiable measurement, and the causal mechanism that connects height and success in basketball is plain as day simple: proximity to the hoop.
> Why are you so sure that IQ isn't just a useful fiction that masks a large number of underlying factors that we don't fully understand? Why are you so sure that it is the best predictor of success in these fields. Perhaps other traits are equal or better predictors of success in a particular field, like tenaciousness, or conscientiousness, and perhaps IQ doesn't matter as much if one of those aren't present.
Because the study of human intelligence is a well established scientific discipline with a lot of well-regarded research that replicates and provides explanations of real world phenomena. And it all points to there being this (not yet fully understood) thing called g that is the strongest known predictor of tons of things related to life outcomes.
Here's a blog post by Razib Kahn about an episode of his podcast, The Insight, with guest Stuart Ritchie, a researcher in human intelligence. The podcast episode is a good start.
"However, critics of g have contended that an emphasis on g is misplaced and entails a devaluation of other important abilities, as well as supporting an unrealistic reified view of human intelligence. "
You don't need to "define" it in some philosophical sense to use it.
Compare "mass". What is mass? Can you really define it? You can say what it does, but what "is" it? Well, nobody can define it, except in terms of the things that it does. Because those are all we see of it. We can measure it, we see those measurements let us predict other outcomes. So we can use our knowledge of mass in practical, predictive ways to build things and accomplish things. Physicists still don't know what it "is" and they don't have to.
Now intelligence. You can measure it. The measurements stay stable on a given person over time and between tests. The measurements predict a wide variety of outcomes - both of a large life scale (income, longevity, criminality, etc) and on an individual task scale (this person can learn X in Y time). It's the most solid, repeatable, predictive result in all of psychometrics - more than personality traits or anything else we can measure about the mind. but what "is" it? It's the thing that predicts all those outcomes. No practical use for it requires that it be anything else.
The whole "what is intelligence anyway" thing is a giant red herring; it's a specific case of a general-purpose counterargument that can be used to attack literally any statement about anything by demanding endlessly more rigorous definitions of the terms involved.
I never claimed that intelligence is nothing. I said that the IQ measurement is itself not worth much except as a general guide in fairly broad ranges.
Intelligence is certainly something. It is relevant. But it is not precise, and is useful, as I said, only in the broadest strokes. And like all psychometrics, the manner in which it is measured is itself not stable.
But comparing intelligence to mass or any other physical measurement is a non sequitur. Mass is precisely measurable with perfectly repeatable results, and has perfectly repeatable interactions.
I think it's worth separating out different tasks in terms of how g-loaded they are (that is to say how much they require intelligence).
General life success is g-loaded but not to an extreme degree. As many have noted, it's certainly possible to have success without great intelligence by working around one's limitations and finding other strengths.
But consider other tasks like "invent a new theorem in particle physics and get it published in a top journal" or "improve a mature database/load-balancing system to save a million dollars a year for a large computing company". These are extremely g-loaded tasks. Intelligence, as measured by IQ, is an absolute requirement to be able to do these things at all in my opinion. My sense is I don't think anyone could ever do such things without scoring 120+ IQ at absolute minimum and probably much more, though I'd be happy to hear counterexamples.
That's an example I'd say where intelligence as a concept and measure is useful in narrow strokes: When you need such a task to be done and done well, you can use intelligence measures to filter who does it (the same way you'd use stature to filter who you put on your basketball team).
In any case, however useful intelligence is, it's the most useful of all psychometric measures. Everything else is worse. That makes it not a great tool necessarily, but the most generally important among the tools we have.
> I said that the IQ measurement is itself not worth much except as a general guide in fairly broad ranges.
IQ predicts about 30% of a whole host of life outcomes (mostly related to job performance, educational attainment, earnings, etc.). That might not seem like a whole lot, but it's the best single predictor of life outcomes that exists, aside from looking at a person's parents and siblings.
But IQ itself is predicted by parental earnings and parental educational attainment (and, IIRC, mostly by the sheer volume of reading for pleasure during childhood.)
There's a chicken/egg thing here that I think is a cart/horse thing.
As someone who does research in this area, broadly defined, I think you're on to something, but I also think there are some misleading things about this article (which I nevertheless think is interesting) and caveats to what you're saying.
Lots of thoughts:
1. Intelligence is a broad construct. It is by definition, and it is not the only cognitive construct. It does have a lot of utility for certain purposes though, such as in identifying pervasive neurological disease.
As others are noting, this is relevant to the article in that we tend to focus on extremes when making these kinds of comparisons, when the full spectrum is really what's important sometimes. We tend to fixate on whether someone went to some prestigious university or less prestigious university, or whether our incomes are in the upper middle class or upper class, but in the sense of outcomes, compared to all outcomes, these can be relatively minor distinctions and hard to predict.
2. There are other variables that are relevant, like conscientiousness, ruthlessness, and so forth. This is certainly true.
3. There are still other variables that have nothing to do with the individuals involved though. The elephant in the room are societal and other random factors that prevent any individual attribute from mattering as much as they could. The article starts out by dismissing prediction among females out of hand because of societal limitations, which is reasonable. But there are lots of other variables involved, random and nonrandom societal and environmental forces at play. The hidden story is that there are limits to predicting outcomes at all from the individual at hand, meaning that other variables in the environment are working.
4. Measurement of intelligence is fuzzy and imperfect as you're alluding to. It's stochastically imprecise, in the sense that giving the same test twice, or two different tests, will give you somewhat different answers. But it's also imperfect in that the thing it's measuring isn't really what we probably want to measure in an ideal case. Even if the tests were giving the same answer all the time, it wouldn't really be intelligence in the way we want to talk about intelligence.
5. I'm not sure that we really want cognitive functioning measures to be perfectly stable, because I don't think cognitive functioning is actually perfectly stable. It probably varies across the day, for example.
6. Physical measurements are certainly more precise. But the objects systemically are much less complex. It's easier to talk about measuring the mass of a cubic meter of oxygen than it is to talk about measuring climatological variables; something analogous is in play with things like intelligence.
Also, even physical measurements at a certain level become fuzzy and highly interdependent. Measuring mass "precisely" depends on your scale and other variables.
You can quantify mass identically for humans, for other animals, and for inanimate objects. Intelligence-as-measured-by-IQ-tests apparently only works for humans. I would like to see intelligence measurement procedures that work alike for corvids, humans, and digital computing systems. I don't know if they exist. It appears that ordinary IQ tests cannot be applied that way.
My main grievance with IQ as "general" intelligence is its human parochialism. It does not generalize much at all. "The most solid, repeatable, predictive result in all of psychometrics" sounds like very faint praise to someone with a background in the natural sciences.
> I would like to see intelligence measurement procedures that work alike for corvids, humans, and digital computing systems. I don't know if they exist.
Of course these objective psychometrics exist, but it's usually only referenced in fields like information sciences, xenology, or speculative futures studies. The obvious reason is because IQ or g factor is only relevant as a barrier to entry in human societies for eliminating potential revolutionary competition against the higher incumbent echelons, pseudoscientific justification of persecution and political subsidy gerrymandering by fracturing demographics into enclaves of special interests to play off each other, or simply as a cultural shibboleth to identify peers for collusion and enemies for swindling.
To start with, there's metrics like the encephalization quotient which are still empirically based (read: gimmicks to support a pre-defined conclusion) curve fitting of anthropocentric expectations of how a genre of organisms should be judged relative to how humans perceive themselves in ability. Yes it's an improvement from IQ or g because it simplifies what was a constantly changing, completely opaque, and wholly arbitrary metric into something that at least measures one physically real property, namely brain volume. But I guess you could say the same about things like phrenology or any other quackery.
There's further refinements to metrics like the sentience quotient based on the density of computational matter or surface area of the I/O boundary of an organism against its environment. Even metrics like this still have incredible assumptions on the nature of intelligence as if we should prioritize the bandwidth or latency of interconnect, I/O, or memory among many other considerations such as algorithmic efficiency -- a whole other can of worms because that means you now need to define the relevant sources and sinks of information and that mapping essentially implies the "purpose of life" which is still a tricky thing to classify. Maybe it's the ability to minimize the time required to maximize the diffusion of an energy gradient? It's unclear.
And that's pretty much the crux of all this that has been stated before on this site[0]: "Metrics, even if not quantified, are always goal-oriented in providing an explanation or use."
Sometimes that use could be nepotism, mental masturbation to relieve some angst brought on by realizing a life wasted in pursuing nonsense like psychiatry, or maybe it's just to survive when the only source of funding demands you to tailor a patch of woven bullshit to mend the emperor's increasingly tattered clothes.
Would you please stop creating accounts for every comment or two you post? This is in the site guidelines, and we ban accounts that do it, for reasons explained at length here:
You're not measuring; "intelligence". You're measuring someone's ability to perform well in a very contrived test. A test that no one could even take without a very specific education experience.
So in that context, I'm sure the results are robust. But I don't think that means much
> You're not measuring; "intelligence". You're measuring someone's ability to perform well in a very contrived test. A test that no one could even take without a very specific education experience.
You're not measuring; "mass". You're measuring an object's ability to attract other objects in a very contrived test. A test that no object would even undergo outside a very specific physics experiment.
The difference being that an object's "mass" has very broad implications for all the theories that incorporate the notion of an object's "mass", all of which have been tested to high precision. (In tests that range from being contrived to not at all.) You can't say this for "intelligence" in anywhere close to the same way.
But it seems more HNers are dead set on believing that there is a genetic factor that can make certain people more "superior" than others, to justify existing social hierarchy.
Does it also bother you that tall people are statistically more successful basketball players? Or that wiry people are statistically better marathon runners? If no, then why does it bother you that intelligent people are statistically more successful in positions that require complex mental work?
The intelligence trait is real, not imaginary. Sure, it's a lot more complex than a simple, well-defined concept like mass. But then again, that applies to everything related to the human mind.
> [...] to justify existing social hierarchy.
What a weird attitude... Would you prefer that someone's position in the social hierarchy is even more strongly determined by their parents position or money? As was the case in medieval Europe?
It is relatively easy for a good IQ year to deliver a “false negative” — that is, a significantly lower score than a theoretical true score. This could be as simple as something like not sleeping well the night before or just simple ambivalence.
It is relatively difficult for a good IQ test to deliver a false positive — that is, a significantly higher score than a theoretical true score.
I personally don’t think that this issue is taken into consideration enough when discussing IQ.
Right. My boys all achieved tests in grade school that varied from 100% to 3%. My wife worried that something was 'wrong' with them. I assured her by noting, its easy to not care about a test and get 3%. Its essentially impossible to accidentally get the 100% score
Your comment doesn’t address my question of how many people are misclassified by IQ test. I suspect smart people with high functioning autism and the like are not accurately measured. Not saying any of this applies to Feynman by the way. Just used him as an example of where an IQ exam underestimated someone.
It´s fairly well known that people from impoverished backgrounds get lower grades for several reasons - not ever having practiced similar questions and vocabulary issues: there have been some quite notorious questions, especially over in England that were dependent on different social classes having different vocabularies, etc.
Feynman is a notable exception to the trend. Using a single exceptional result as evidence for decrying a trend is foolish.
IQ tests obviously do not account for __every__ variable which we cumulatively call "intelligence", but IQ is heavily correlated with the ability to succeed at tasks society classifies as "smart people stuff". The difference in intelligence between two individuals is much less pronounced if both individuals are already two standard deviations above the norm. But if you were to compare two individuals, one who is one standard deviation below, and the other one above, you would observe a clear difference in 'intelligence' between the two. Even in this comparison, there will be outliers, but the outliers are not statistically significant.
>Feynman is a notable exception to the trend. Using a single exceptional result as evidence for decrying a trend is foolish.
You missed my question. I am concerned with how many smart people with disabilities IQ exams misclassifies. I am not using Feynman to say he had a disability or anything, I was using him as an example of a case where his IQ score doesn't match his actual intelligence, and I am in no way claiming why that may be. I am merely stating IQ exams aren't perfect. If you are interested in replying, then reply to the question I posed. I think IQ test potentially misclassifies people with high functioning autism and other highly intelligent people with mental disabilities, showing they have lower intelligence than they actually have.
>Or, more precisely, IQ is heavily correlated with the ability to succeed at taking IQ tests.
...and succeeding at taking IQ tests is also predictive of an individual's potential to succeed at, using the Feynman example, physics.
Lets say you were to gather two pools of individuals, one pool with IQ test results of 2 standard deviations below, the other with results of 2 standard deviations above, and provide them with the necessary materials and instruction to understand physics 101. Incentivize the participants by offering a $2000 cash reward for demonstrating the knowledge they were tasked with learning, lets say by answering X questions presented to them in whichever format would be optimal for their success.
How many from group -2 SDs do you think would receive the $2000, compared to group +2 SDs? -1 SDs compared to +1 SDs?
The choice of physics was arbitrary as well. You could assign almost any field of study, and you would still get almost identical results in performance.
my guess is that a well designed IQ test actually tests what it tests for with good accuracy/precision. i am skeptical that the IQ criteria are very close to what we consider "intelligence".
Furthermore, most IQ tests are not meant to measure intelligence past a certain point. A person with an IQ of 160 is not inherently smarter than someone who has an IQ of 150 in the same way someone with an IQ of 100 is inherently smarter than someone with an IQ of 90.
There can be measurement errors, like with any metric, but statistically speaking, a person with an IQ score of 160 will be more intelligent than somebody with a score of 150. And significantly so. 10 IQ points is a big deal (2/3rds of a standard deviation). Increasing the average IQ of any population by 10 points would be a total game changer.
Are you really arguing that most IQ tests have replicable sensitivity above 130-140?
Because that is very different than what I’ve heard. My impression is that most (perhaps not all) IQ tests are stable and highly replicable for results up to that range (assuming consistently motivated test takers, a very big assumption, particularly for bored smart people), but test scores above that range are mostly noise.
In another part of this thread[0], Bartweiss says that Raven's Advanced Matrices is valid up until about 150 IQ. So I was close. And my bigger point is that, though the measurement lacks precision at that level, if you lined up a bunch of 150 IQ people and a bunch of 160 IQ people, there are probably some aggregate group differences. Anecdotally, I've spent a lot of time around really smart people. There are gradations even at the top (just as there are in all other areas of talent; there's only one Lebron James, after all)
Sounds like you went to a good high school. Mine was a terrible high school that offered 0 computer science classes.
>One really intelligent decision the department made was to not just mimic a college curriculum.
It would have been GREAT if my high school even mimicked a college curriculum. We barely had any AP courses & I think it would haven't been a surprised if you didn't go onto college from my high school. You re definitely speaking from a place of immense privilege and it's amazing your school offered that.
> 1) Do FAANG companies hire non-PhDs for machine learning positions? Most seem to require a MS or PhD
That depends on what you mean by "machine learning positions" and what your bar for normalcy is. For research roles - these generally have distinct titles like "Research Scientist" or "Quantitative Researcher" - it is extremely difficult to get an interview without a PhD, let alone an offer. It happens, but rarely, because there are many capable people with PhDs and other relevant experience applying for the same roles.
If instead you relax the bar to also include software engineers who work with research scientists on model implementation and optimization then yes, people with "only" an MSc are routinely hired for these positions. These engineers are still credited on papers published as a result of their collaboration, and they still need to have a firm understanding of how the models work. The difference is that they don't tend to have leadership roles and don't develop novel theory - they are responsible for supporting the core research team and helping the research output become production ready software.
Both of these types of roles require strong coding skills, but the "hard" research roles require significantly stronger mastery of linear algebra and probability theory. As an example, compare the roles for Research Scientist[1] and Research Engineer[2] at Facebook. You'll find a similar bifurcation at other industry labs like Google, Microsoft and IBM.
If you have the opportunity to do either and you're optimizing your career for wealth maximization or research impact, it's better to obtain a role as a research scientist. That being said most PhDs do not end up at Google Brain or FAIR, so it's not a cut and dry decision. You can't just choose to trade n years of your life and an easier-to-obtain role on the periphery of research for the ability to do theoretical research at the best tech companies in the world later on.
Thanks. Can you list what courses one should focus on as a baseline to do research in ML? You mentioned Probability theory & Linear algebra, but what specific topics in those areas should one learn? Any other fields of math?
Probability, statistics and linear algebra are recommended because all of machine learning uses these as foundations. Any graduate level course in these areas will cover what you could possibly need. For people working in the more theoretical side of machine learning, maybe some analysis (functional specifically) might be useful.
I work at NVidia as a GPU architect doing ML applied research. I'm not a hiring manager, but I have interviewed a bunch of people.
1) Having a PhD will make it easier to get an interview, but it is not necessary. Relevant experience counts just as much. I only have an MSc.
2) That will vary a ton from one team to another and from one position to another. The interviewers are not going to change their questions depending on your education.
3) The same requirement as for PhDs. It will depend on the role, but in general I expect people value hands-on experience more than theory.
Except for some particular hiring managers with strong opinions, a PhD is not going to be a hard requirement for nearly any job. However, a PhD in a relevant area is going to be useful, just like any other relevant experience you may have.
5) No, but I am obviously biased due to not having a PhD. I doubt you are going to get an unbiased answer from anybody; everybody wants to believe they made the right choice.
> I doubt you are going to get an unbiased answer from anybody; everybody wants to believe they made the right choice.
I'm not sure where choices come in here. I specifically mentioned a mixed team; the choices have already been made, so the performance of the existing team is what I'm asking about.
As far as not getting an unbiased answer, that's why I'm asking HN at large -- hopefully there are enough people in enough environments to give an interesting and informative combination of answers. :)
I think it's a pretty good starting point. I also like Cracking the Coding Interview and I think there's definitely a place for timed coding challenge sites like leetcode - especially if you've been in a role where you're mostly working on larger-scale problems rather than on producing smart, working code quickly on the fly.
Can you tell me what companies find attractive about: "...producing smart, working code quickly[;] on the fly," (emphasis mine)?
I understand there are a lot of substandard programmers in the marketplace. However, why is it that this specific criteria is the one the industry is so attracted to? Could it be that kids are proficient at these kind of games? Because I can tell you that in the 1980s, 1990s, and early 2000s, I wasn't being asked to write a regular expression parser under time pressure.
Others in this thread say they've done Triplebyte take-home tests, only to end up in a "go fast-fast-fast!" interview in the end.
Why is speed so important? Every popular [aA]gile methodology today is implicitly -- if not explicitly -- against such "machismo" programming. If you're pair programming, how is this ever relevant?
Whenever I see someone say leetcode, hackerrank, and Cracking the Coding Interview is the "answer" it translates to "only the young need apply" in my head.
Is this why these companies have such intensive leet coding interview processes? In other professions a degree + exams alone is enough.
I can't imagine an interview where a doctor has to perform surgery "as a take home interview" without pay or be able to solve obscure problems from chemistry, biology, etc.
I've interviewed a lot of people at this point and I just can't rely on a CS degree as a guarantee of competence. I've interviewed (and worked with) people with degrees from excellent CS programs who were bad engineers. Some of them must have cheated in their CS courses because they were outright incompetent -- but they were able to get degrees.
It's a serious problem that we don't have a good way to tell if someone is good at software engineering. Unfortunately degrees don't guarantee anything. So we use whiteboarding, which isn't great either.
I'm not really sure we should want degrees to be a guarantee of competence anyway. If the industry relied on them it could lead to a culture of credentialism that would make it more difficult for good engineers without CS degrees to get jobs. Some of the best engineers I've worked with have degrees in other subjects.
One can say this about any other profession, can't one?
The notary I met while buying my last apartment was a sham, he probably cheated while studying. Guess what? My acquaintance rubber-stamps papers like no one else, one of the best paper rubber-stamping guys I know.
> Some of them must have cheated in their CS courses because they were outright incompetent
I suspect they just didn't want to be software peeps. I suck at anything I don't wanna do, and I find a way to get better at anything I'm really interested in.
Agreed. Considering how important this problem is, you'd think there would be a generally accepted solution that was both effective at finding the right candidate and efficient with resources (both the candidate's and the company's). As far as I know, there isn't one.
To be fair, there's a board certification process for doctors that they are typically expected to pass before being considered for employment. You could argue this is their version of a 'take-home' interview.
>Thats why some companies stop hiring ppl after 30s
This is a huge generalization. Some people's careers start in their 30s. Would you think some 30s something ML researcher that just got a PhD is over the hill? Not everyone is married and has kids in their 30s either.
all the CS postdocs I know see a) pretty happy b) have a really easy back up of a high paying industry gig that they can pivot to because of currently insanely inflated demand for academics to make deep learning on the blockchain on the cloud on embedded devices doing quantum backprop
I am primarily interested in the experience of ML post-docs that forgo going into industry. Can you elaborate a bit more about their experience? Salaries look a bit lower for post docs vs industry work. Some I've seen start in the mid-40k to 70k range. Could be different based on geography.
are you already a grad student or are you considering it?
in my department [we are an outlier probably] post doc wages are actually pretty comparable to industry i.e. >= 100k. 40k to 70k seems low to me.
I would expect the range to be more like 60-100k.
I haven't been on the real job market yet-but the key to getting a TT or postdoc position seems to be a) collaborators who want to hire you b) people who have heard you give a talk in person and are familiar with your research.
I am considering going into a CS PhD focusing in ML. The mid 40k-70k range was from quick google search I did for CS post docs in ML in lower cost of living areas where the cost of living is much lower than on the West Coast. I am trying to look at career prospects and weigh whether it makes sense to stay in academia or jump to industry (after I complete a PhD). If wages are closer to 60k-100k for post docs, then I may consider staying in academia for some time after completing a PhD depending on whether my career interest shift.
Well I would be happy to provide some context.
I just finished my first year of CS Phd in ML (more on the theory side) and I really like it. I think most of the places you would want to do a post doc in CS are probably going to be moderately high CoL. My phd is in a place with pretty low CoL (but a still a top 10-top 20 school (depending on who you ask) ) so the graduate stipend goes reasonably far.
The other thing to note in ML is that it seems like a few people go to industry research labs for a few years i.e MSR/FAiR/google brain and then come back to the academy since there are industry roles that involve research and publication. for instance moritz hardt.
my personal plan for the first 3 years of grad school is to work really hard and try to keep both academia and industry open and after year 3 evaluate the number of publications I have and my current skill set to see if I can make it in the academy or shift more towards industry.
I think the biggest factor I would comment on is look very closely about what jobs the graduated students from the department you matriculate at AND more importantly the professor you want to work with go on to do post Phd. There are a lot of naysayers in this thread about the risks of an academic career and I share those concerns but I felt a lot more comfortable taking the plunge after I looked at the career record of the graduated students of my advisor. They were all either tenure track or had good industry positions.
edit: if your advisor has collaborators in industry groups I think it is pretty straight forward to get an industry gig.
When evaluating the warnings from naysayers you have to keep in mind that CS is quite an outlier as far as backup career prospects go. I made it all the way to CS postdoc and every step of the way I had to keep swatting away industry recruiters waving wads of cash at me. I finally made the leap for other reasons but it was effortless. I think this is absolutely not how it works in other disciplines.
One exception I can think of is what I call "closet programmers," which are folks that work in various areas which rely on software such as experimental physics, astronomy, molecular biology. and end up mostly doing programming because they love it. We have a bunch of engineers like that and they are all excellent :-)
100% correct. I think we are both very lucky in that we are able to do fun science and chase our intellectual interests with a realistic and still fun safety net.
Also those closet programmers are always really fun to talk to since their problems and culture are breath of fresh air.
Thanks for the context. That sounds like a good plan to me regarding post-doc locations. I am also interested in theory side of ML. What areas of mathematics should one learn really well that apply to the theory side? What blogs, papers, books would you point one to to learn the theory side more? To your knowledge are their applications of abstract algebra to ML? If so, what areas of algebra apply & what problems do they solve?
I could rant about this for hours. I actually just went to a defense for a deep learning paper that had a ton of abstract algebra. I am honestly not really a fan of deep learning and algebra because all the papers to me like- seem to stop at describing some really basic feedforward network as some really specific mathematical structure but these theories a) provide very little explanation of empirical phenomena b) provide no new directions of research in terms of like useful network architectures.
I haven't really come across algebra in machine learning other than people applying it to deep learning.
This book is nice since it really balances theory with a more practical understanding.
-An Introduction to Computational Learning Theory by kearns is a classic [low priority]. this is fun since the proofs are simple and deep but is very very far away from practical algorithms.
-convex optimization by boyd
Course Notes:
[I think a good alternative to blogs is stalking course notes for other schools-they are very often
public.]
>In my experience, bachelor's in CS counts more than master's in CS
This discounts people that transition to CS from another field like mathematics, chemistry, physics, etc. Just because you didn't major in CS as an undergraduate doesn't mean you can't program. That is a bit of gatekeeping. Plus, I'd argue someone from a field like mathematics and physics might be more suited to do certain types of programming jobs due to their background skillsets.
>But they don't have as much experience programming?
Then by that metric you don't even need a degree and your initial point of a BS > Masters > Bootcamp is nullified by the self-taught programmer that has programmed longer than anyone else in that group that happens to apply for your position. But that doesn't necessarily mean they will be a better programmer in the long term.
Do Data Engineering jobs require you to have knowledge of statistics, machine learning, etc? Or can you get data engineering jobs where you are focused on connecting pipelines and data flow? I am interested in a data engineering job, but it appears a lot of data engineering jobs require machine learning expertise in job listings. I am interested in just a data engineering focused role.
Look for functional programming data engineering jobs e.g. Haskell, they might be what you're looking for, since most people don't run their algorithms in Haskell (some do!) and use it as a duct tape/pipe.
I wonder how many IQ exams misclassify the intelligence of smart people with disabilities of some sort or high functioning autism. Not saying any of this applies to Feynman, he is just an example of where the test didn’t accurately capture his intellect.