Artificial-Intelligence Experts Are in High Demand | Hacker News Reader

Artificial-Intelligence Experts Are in High Demand

wsj.com

224 pointsmwytock111 comments

This is suspiciously close "data science" and "machine learning" experts.

Can't we just be honest and say that most of these are applied statistics jobs with a specialty in large volumes of data? Or is "statistics" just not fashionable enough nowadays?

I imagine the "AI experts" will be paid significantly more than data scientists and machine learning experts, just like "software engineers" are paid more than "software developers" and "programmers".

These subjects (plural) are all plagued by the same problem: definitions of terms. One mans intelligence is another mans dire stupidity, and so on and on it goes, chasing its tail.

The most value I got from this article was in the realization that, every few years or so, the academic globes align well enough (some paper de joure becomes well-read I suppose) that .. for a brief instant .. terms are defined well enough, and gain enough agreement, that progress is made .. which progress attracts more eyeballs, who tend to want to break off a chunk for themselves, and the terms begin to differ again and we have a whole new 'sub-sub-sub-' variety of the subject.

So its all about globes aligning, basically. I will now go off and implement an AI technique based entirely on the description of globes, alignment, and little chunks breaking off every now and then .. see you at the top of the AI heap in a year or ten.

As mentioned in the article, AI is the broader field that encompasses Machine Learning, and to a large extent also Data science (and Computer vision, NLP, Pattern recognition, etc.). And while data science might utilize a lot of statistical techniques, it is a huge stretch to consider the whole AI field to be 'statistics'.

In general, AI borrows many more techniques from mathematics than it does from statistics. However, the field of AI has been quite established since the 1960's, and many techniques have been developed within that field as AI techniques, it's more about being accurate than about being fashionable as AI simply isn't 'just' statistics.

I believe the OP's point is that the demand is for applied statisticians and not for AI experts (in the sense exactly as defined by you).

The article is about tech firms and universities stocking up on research centers of AI experts, with the claim in its title that there is a high demand for those AI experts.

There might also be a demand for applied statisticians, but that doesn't make AI experts statisticians. I understand the confusion, as the term AI is often misused, but when you see the names mentioned in the article it's clear they're talking about actual AI researchers.

I think there are two levels. On the one hand, many firms need big data experts who can reason statistically and apply machine learning techniques to their domain. This started out 15 years ago as predicting shopping cart basket items etc..

On the other side the big tech companies are investing heavily in Deep Learning for things like NLP, Speech, Vision, Siri, and wherever else these neural net approaches may work etc...

Perhaps more precisely, they're "statistical engineering" jobs. A machine learning PhD can derive an algorithm and provide you with a reassuring bound or guarantee regarding performance in terms of runtime, convergence, etc. They have to be able to understand not just the volume of the data but also how to trade off accuracy for speed, and myriad other constraints.

IMO, the "data science" label is too broad to properly differentiate statistical engineers. A fine definition for a data scientist is someone who runs experiments on user/company data and can assess the results. It's important work, but you don't need a PhD in stats or ML to do basic hypothesis testing.

You could simply call them "machine learning" experts, but that could be a bit too academic. People who are focused narrowly on theory or niche areas may be experts in ML, but they may also never do anything outside of running matlab simulations. It's unlikely that those people will make very good statistical engineers since they may never have had to think about the challenges involved in scaling algorithms.

Fully agree. I mostly lost interest in AI because modern AI goes toward statistics. Unfortunately, symbolic AI doesn't get much traction nowadays, and that is understandable.

Did you take notice of https://news.ycombinator.com/item?id=9432601 a while ago?

I often heard the Big Data guys hype that there's no sampling in Big Data, you have the whole data, so it's not exactly statistics.

from my distant memory if you sample size is the pollution its still statistics

you mean Pig Data?

Population I of course meant to say !

I've heard this too and it's a great way to demonstrate you don't really know what statistics is :)

Statistics is not (just) opinion polling, there's a lot more to it than estimating observable properties of a population.

If you're trying to make decisions, predictions or estimates which involve any uncertainty at all (and in my experience big data almost always is), then it's definitely within the purview of statistics even if you have data for the whole population.

Sources of uncertainty include trying to say anything at all about the future (do you have data on the future population? no didn't think so...), trying to make predictions which generalise to new data in general, trying to uncover underlying trends or patterns behind the data you see which aren't directly or fully observed.

Often people expect big data to be able to answer big numbers of questions, estimate big numbers of quantities, or fit big, powerful predictive models with lots of parameters. In these cases statistics can be particularly important to avoid reporting false positives and to make sure you can quantify how certain you are about your results and your predictions. (Amongst other reasons).

Not to mention: having all the data, and comprehending all the rows on an individual level, are two very different things. Doubly so if the data is irregular (I'm currently doing fuzzy matching on really mangled street address data. ICK).

Once you hit millions of rows, it's not humanly possible to survey the data. All you can do is make assertions about the data's structure / buckets it will fall into. You then try to disprove that assertion, or establish an error bounds on it. You will never see all the data, only the results of assumptions you've made about it.

The refined pieces of information that people can look at to make decisions are called "statistics".

Presumably you want to draw an inference of some sort from the data. Otherwise what's the point of even looking at it?

There is quite a big difference between statistics and machine learning. A lot of the most successful machine learning algorithms do not have a statistical grounding or did not when they were invented. E.g. neural networks, SVM, low rank matrix approximation, k-means, decision trees/forests. Statistics is one of the tools in the machine learning toolbox.

But, to look at one example: neural networks. Neural networks may have been inspired by attempts to recreate the structure of the biological nervous system, but the way in which they are used commonly, e.g. "learning" via back-propagation, is really just a statistical regression for a gigantic equation with many free variables.

My preferred term is "predictive analytics," which I feel kind of straddles statistics and machine learning, and also serves as a nod to a common difference -- "statistical" methods often yield understanding, while "machine learning" methods are often opaque to human insight but yield predictions.

I feel annoyed with opaqueness of ML algorithms like neural networks. I hope ML doesn't unwittingly define itself as a field where machines learn, but humans may not learn. I'm referring to predicaments like the story about 42 from hitchhikers guide to the galaxy.

That's definitively an interesting problem. Just note in many cases we're not even interested in learning the tasks. For example, you don't need any person to actually know that consumers aged 25-29 years old prefer a certain product 10% more than consumers aged 21-25, and so on.

But humans are still the ones responsible for important high level decisions, so it still makes sense to maximize information transparency to enable good decisions in those contexts.

A neural network that given a prediction 'X is most likely' and could answer the question "Why?" with 'Because Y' would be amazing.

SVMs were invented by a couple statisticians/mathematicians in the 60s. k-means also harkens back to the 60s, by mathematicians and control theorists. Decision Trees and Random forests were invented by a famous statistician, with the latter related to bootstrapping, a statstical technique. PCA and factor analysis, forms of or closely related to low rank matrix approximation, were pioneered in the early 1900s, by some of the most famous statisticians ever.

Something that was invented by a statistician is not necessarily statistics, and that certainly applies even more to something invented by a mathematician. I guess with a broad enough notion of statistics some of these would fall in the field of statistics, but if something does not use at least one probability distribution it's probably far fetched to classify it as statistics.

It would be a lot more fair to classify machine learning as a subfield of convex optimization. Yet even that classification does not quite fit, so it makes most sense to just accept that it's a separate field which uses techniques from statistics, convex optimization, computer science, and more.

What's become very clear in the past ~5 years is that we're seeing the emergence of a new field, very distinct from statistics. The closest equivalent of the new machine learning field is electrical engineering, which has now heavily shrunk. Indeed, many former EEs have made a natural transition into this new field.

The new machine learning is about building layers of components on top of each other, very much like circuits seen in EE. The "circuit" components being used are no longer well defined mathematical pieces built from the bottom up using ideal assumptions, but less well understood, somewhat black-box newer components that were built from the top down. Far more like a type of engineering than a type of statistics.

If you haven't been seeing all the latest Arvix papers, you're really missing out. It's evolved to look sharply different than statistics now.

Just because you're using statistics it doesn't mean that you are a statistician. For example, most of particle physics is based on statistics, but this is not enough to motivate them to rename the field. Whole fields of engineering use statistics every day, machine learning is just another.

Your comment might mistake "data science" as being more than a neologism for describing something that hasn't yet taken enough concrete shapes to be clearly defined.

I think many would agree that "machine learning" and/or "deep learning" are at least cornerstones for "artificial intelligence". After all, nobody singularly defines intelligence.

> Can't we just be honest and say that most of these are applied statistics jobs with a specialty in large volumes of data?

But isn't this the approach Nature herself is taking?

The knowledge engine you carry in your head spends years just "learning" the world - which means, it absorbs huge amounts of input, sorting the good stuff from the bad. It "knows" what works simply because that stuff happens more often; it "knows" what doesn't work because that stuff doesn't happen very often.

And sure there are higher layers of integration there, but the whole process is strongly supported by a statistical approach.

This trend is in most companies business-driven, in others it is technical-driven. Few companies have technical leadership that can manage true AI resources. If you remember the ML courses from Uni and experts in that field, you can imagine why. In many universities AI departments are assigned to schools of psychology and philosophy. Only companies with a deep engineering culture as those mentioned here can build up true AI departments.

The other driver is business-driven. And this is where management demands 'AI experts', when what they really want is data-miners. And in many cases management prides themselves on 'AI algorithms', but we know that this is a term for anything that gets the results that management wants and may be far from intelligent and in most corporate cases a bunch of SQL scripts.

What's the potential path forward (say projecting 10 years ahead) from the current growth in demand for data mining centric people?

I mean people go and study in response to demand. They learn data mining and AI at Universities. I think it's often people with backgrounds or aptitude in maths. What will the 22 year old with an aptitude for maths that is learning R, SQL, AI-for-business and such be doing in 10 years?

I don't know if the starting point matters much. "Results Driven," even if its optimising inventory or making ad purchasing decisions or data mining old DBs is not a bad place to "search" for advancements. Not everything needs to be fundamental research.

I doubt the 22 year old you are referring too will run out of problems to solve. I also feel at some point it will be like a lot of software engineering is today. Working for companies implementing solutions similar to what already exists but tailored to the context of that companies specific needs. As of now I feel that this field is so young that a lot of the solutions are almost completely custom built to the problem at hand and that a lot of work is needed to abstract away those solutions into higher level reusable pieces.

> The other driver is business-driven. And this is where management demands 'AI experts', when what they really want is data-miners. And in many cases management prides themselves on 'AI algorithms', but we know that this is a term for anything that gets the results that management wants and may be far from intelligent and in most corporate cases a bunch of SQL scripts.

I wonder if this is a case where managers end up believing their own bullshit. "AI driven" is basically the marketing-speak for "a bunch of SQL scripts".

Here in Germany hardly any big company have any clues about the insights which Machine Learning and Data Mining can provide.

Most upper management sees IT, only in a business suport role not as a business driver!

The big thing that prevents me from getting into AI is the lack of practical projects that I can build.

It is a very interestimg field, but as a self-taught programmer I'm used to learning by building things, and it's hard for me to come up with some project that would be practically useful and yet doable.

Does anyone have any ideas?

Game AIs, Stock market bots, character recognition, adaptations to game of life, computer art wherein you get an AI to paint or to draw and can seed it values. There's lots of cool stuff you can hack away at. Even training a neural net to recognise a '3' is quite interesting.

These things suck. I want a robot with sensors, the ability to move and an arm that can be programmed with a language that is appropriate for AI. That doesn't sound like something technically difficult.

So, what's holding you up exactly? RC cars are cheap, people have done cool stuff with old android phones for sensor packages. If you need more horsepower, stream the data back to a PC, you've got wifi on the phone. New industrial arms are expensive, but you can scrounge one, or get a hobby one. sparkfun had a uArm that would probably work for you.

Because I am a software developer and intelligence hobbyist (from the biology/ethology camp.) I don't know a damn thing about RC cars or android phones nor do I have the time or desire to learn. Sadly, the hardware tinkerers don't know a damn thing about programming intelligent behavior. Until there is some sort of API to connect hardware to a software environment programmable by specialists in that domain, hobbyist robotics will remain in the realm of Battlebots.

Hmm. There are a few options. You could get a roomba, and put a laptop on top of it, use the laptop camera to sense, and the serial port to command the motors. I kinda think you could use python to glue everything together.

Mindstorms are really easy to get started with if you want more control over what the robot will look like, but you'll be facing more mechanical engineering problems.

diydrones has a lot of good information about getting things connected. Most people look to RC cars because they're so inexpensive.

So, the big thing with hardware is it kind of sucks. It always feels like using a butterknife to tighten screws. Things are challenging in ways you don't expect. With software, if you make a mistake, you fix it and recompile. With hardware, you hope nothing breaks.

I too wish it were easy. It's not, it's complicated in ways that suck. If you have money to throw at the problem, i bet you can find a nice industrial platform that's built like a tank that has a nice API. That kind of stuff is thousands of dollars, and i'm not familiar with it.

Maybe I'm misunderstanding you, but robotics is now, and probably always will be, inherently about combining skills from electronics, mechanics, and software. I'm not sure what you're looking for, a complete robot that you can simply program? You could look into the NAO robot[1], probably too simple for you, but I think they're pretty cool. Wish I could afford one ;)

1 -- https://www.aldebaran.com/en

After competing in an AI contest a few years back, I had fun and wanted to do more, but ran into the problem you describe.

The state of hobbyist robotics is very hardware-centric. I not yet discovered something appropriate for use by AI programmers.

So I decided to step back and work on theory. I did this for several years and recently finished. Hobbyist robots still aren't ready to implement my ideas so now what do I do? I started working on a software architecture based on my high-level research. But how to test it? So now I am working on software to simulate an environment for my AI to interact with because I don't have the time or skill to build a real robot.

I honestly don't think this is a bad thing because I would imagine that when we start to work on "real" AI (not this nonsense that passes for it today) testing behaviors in a simulated environment before deployment to hardware would save a huge amount of time (and physical damage.)

Now I am hoping that once I have taken my virtual habitat as far as it can go, hardware will be available for me to apply what I have been working on. My hopes are low.

I don't know much about AI, though I know a bit about ML, statistics and neural nets. And one good way I've found to learn those subjects has been to find interesting papers and then figure out how to replicate/reimplement them.

This might sound a little uninspiring, but it's a solid start, and often you can make improvements over the authors' original stuff -- especially in areas where you may have more experience, e.g. computational efficiency.

Try to build a Siri clone, or try any of the Kaggle projects using neural networks.

- Politics: Predict which candidates/bill will do well. Or predict which bills you might care about. Identify the strongest features. Publish a writeup with nice maps/charts. Inputs: FEC donor data, vote history.

- Sports: Predict game outcomes, player performance. Make money playing fantasy sports?

- Real estate: Build a better tool for house assessments that identifies "comps" and predicts what the house should cost.

- Astronomy: Something about exoplanets?

- Amazon prices: Find opportunities to buy/sell.

- Stock prices: Inputs are news/Twitter, outputs are predictions of closing price.

I'd start with your hobbies and go from there!

AI is very interesting but not very accessible because it's so specialized. I have a fairly strong programming background but feel like I'd need to study theory for a significant amount of time to even get my feet wet with AI.

If you have (condensed, especially) AI resources that you think would help bridge that gap, please share! Toy-scale project ideas would also be appreciated.

There was a tutorial link trending on HN a few months back, I can't seem to find it. But these links are helpful (though not condensed) as well:

https://github.com/ChristosChristofidis/awesome-deep-learnin...

https://github.com/owainlewis/awesome-artificial-intelligenc...

These links can be helpful too:

https://news.ycombinator.com/item?id=7783550

https://news.ycombinator.com/item?id=9432952

https://news.ycombinator.com/item?id=2645671

https://news.ycombinator.com/item?id=5508261

https://news.ycombinator.com/item?id=9004689

Buy and work through "Artificial Intelligence: A Modern Approach". It's a huge book and the de facto standard for pretty much every AI 101+ course. Some of the stuff may not interest you some might but it covers a broad range (from logic based agents to Bayesian networks). It's systematic and has excellent references and further reading notes for each chapter. The focus is not on the currently sexy "data science" aspects though (however you will find plenty of material that is relevant).

The edX class from Berkeley is pretty fun and hands on. It uses Pacman as a running example and essentially teaches the agents stuff from AIAMA:

https://www.edx.org/course/artificial-intelligence-uc-berkel...

The Stanford class by Thrun and Norvig himself (one of the authors of AIAMA) is also good but I prefer the edX one:

https://www.udacity.com/course/intro-to-artificial-intellige...

Edit: changed to direct links for the courses

The AIMA book is sort of a Good Old-Fashioned AI (GOFAI) book that focuses a lot on agents and planning. The jobs this article is talking about are really machine learning ones-- taking large volumes of data and extracting knowledge, so as to build recommender systems and such. For that, Kevin Murphy's book, "Machine Learning: A Probabilistic Approach" is without a doubt the best book out there, both in terms of explaining things from the ground up and being the most comprehensive/up-to-date source.

There's still quite a bit of material on Bayesian networks (with the dreadfull dentist example :D), neural networks and support vector machines but overall you're right the focus is on agents. The relevant chapters are great staring points though and as always filled with great reference material for further reading.

+ I'm pretty sure if you apply for an AI job somewhere and it's labaled AI and not "data science" they'll expect that you know the material in AIAMA.

Murphy's book is actually subtitled "A Probabilistic Perspective" -- "Machine Learning: A Probabilistic Approach" is a different book by a different author.

+1 for the Stanford course. Great intro to AI and super easy to follow - I've done it after my uni class elsewhere, and it helped to internalise what I've learned there.

reading the book still requires a significant amount of time

'AI' is not programming, it's mathematics (well the current 'flavor' of statistics-based AI, that is - the 1980's style AI people I used to work with were philosophers, legal scholars and the like). Anyway, there is no 'bridging the gap' - you need to start from a good foundation of statistics (and the 'prerequisites' - algebra, calculus, linear algebra) and in the end, the technicalities of the software and the theory come together naturally.

(source: have tried to 'bridge the gap' for 2 year, including taking MSc courses, before admitting to myself that it's a lost battle. Am now starting to build a solid math foundation before revisiting ML applications.)

This is exactly my experience as well. I'm an alright programmer, but it is insufficient, because machine learning and AI are a table resting on four legs: linear algebra, calculus, statistics, and programming. I've also found myself going back to build up those foundations.

Here are two tutorials I've written on natural language understanding:

https://honnibal.wordpress.com/2013/09/11/a-good-part-of-spe...

https://honnibal.wordpress.com/2013/12/18/a-simple-fast-algo...

When true strong A.I hits, I feel the confusion will quickly lift. You'll know because all of the people with weak A.I :

> Used mainly to strip information value from people without compensation

> Who are dumping money into foundations to prevent the coming of it's more true form

will be screaming 'It's the end of the world'. Until then, enjoy the algorithms. It's the nature of business to over-sell. Don't be too upset by it.

I'm starting an SF-based robotics/AI/ML workshop/meetup/club next week. Hit me up at dan@techendo.com if you want an invite -- or join this group: https://www.facebook.com/groups/762335743881364/?ref=br_rs

We went through a round of this in the 1980s. The first commercial graphics workstations happened to be LISP machines. So management confused non-numeric code with A.I. There was demand for workstation experts. Not to loang after this UNIX graphics workstations like Sun, Apollo and MicroVAX came out and the market switch to UNIX/Linux.

Second was the expert systems boom in the mid-1980s. This was fanned by Stanford professor Fegeinbaum who wrote the infamous book The 5th Generation about expert system computers being the future and Japan was building the best ones. These would either be LISP machines or an interesting French niche language called prologic. Prologic basically traversed a databse "if-then" rules (modus pons). These machines went nowhere and Japan economy tanked in the early 90s. Lot of Silicon Valley VCs lost big on this.

Prof Feigenbaum may still be correct, but 40 years early. However the new A.I. is driven by massive database matching possible in modern peta-level computers and not so much logical computing.

The first four generations were defined by hardware: (1) vacuum tubes, (2) transisitors, (3) integrated circuit boards, (4) microprocessor full CPUs on a chip. I would define (5) clusters and (6) mobile. Candidates for next generates include huge data engine clouds, wearables and internet of things.

What you're referring to is the advent of AI bubbles, and thus AI Winters - http://en.wikipedia.org/wiki/AI_winter

Stating that a bubble cycle has emerged in the means only accentuates the importance of the end, to note that the desire for AI is so strong that futility hasn't kept people from trying.

Virtual reality is another such example.

That would be prolog. You'll have much better luck googling for prolog. I played with "turbo prolog" in the 80s and accomplished nothing (from the same place as turbo C or turbo pascal or turbo basic (am I forgetting any?)). A modern variant (of a logic oriented language) can be seen here:

https://github.com/clojure/core.logic/wiki/A-Core.logic-Prim...

It tends to suffer from management by scalable procedure disease. Its possible to successfully replace a human assembly line worker with a robot arm and a very small shell script, which inevitably leads overactive imaginations to think of replacing engineers or doctors with an immense set of unfortunately undefinable unscalable procedures and rulesets, so it always collapses with complexity at implementation time. Its like moths to a flame, you should be able to replace an engineer with a very long list of if/then statements, but it turns out to be impossible in practice. Meanwhile the more advanced techniques butts up against the rapidly scaling "DBA" "IT" type of traditional solutions or non-traditional big-data techniques.

Its hard to find something to logic program that isn't less verbose in a non-logic language or unwritable in any language including logic programming. Its like the Perl regex thing where you got a problem, so you write a regex, and now you got two problems. Its a very narrow although interesting niche. Finding something that fits would be pretty cool, although probably very difficult to maintain.

No, that was the wrong approach. Back then we believed that solving AI means being strong on logic and rules, with data being a secondary aspect. Nowadays we do the exact opposite: data is king, and the rules are expected to somehow emerge from it.

when the wsj writes about it means that the trend is over

I wish there were any position like this available in Germany ...

Interesting. Do you know if there are any positions in the UK or any other part of Europe?

I think the Amazon ML Group have something going on in the UK.

I will check them out. Thanks!

I wonder what all the AI is going to be used for. Is everyone working on their own Siri and recommendation engine now?

(Is anyone building an AI that can come up with its own agenda?)

I think (not sure, as I don't closely follow the work) AGI people and MIRI work along the lines. (i.e: how to make sure a super AI doesn't go wrong, but comes with Objective functions beneficial to humans. https://intelligence.org/)

Considering being an employee such as in the OP, I have two reactions:

(1) Take statistics, machine learning, neural nets, artificial intelligence (AI), big data, Python, R, SPSS, SAS, SQL Server, Hadoop, etc., set them aside, and ask the organization looking to hire: "What is the real world problem or collection of problems you want solved or progress on?"

Or, look at the desired ends, not just the means.

(2) Does the hiring organization really know what they want done that is at all doable with current technical tools or only modest extensions of them?

Or, since artificial intelligence is such a broad field, really, so far of mostly unanswered research questions, and the list of topics I mentioned is still more broad, I question if many organizations know in useful terms just what those topics would do for their organization.

So, for anyone with a lot of technical knowledge in, say, the AI, etc., topics, it is important for them to be able to evaluate the career opportunity. I.e., is there a real career opportunity there, say, one good to put on a resume and worth moving across country, buying a house, supporting a family, getting kids through college, meeting unusual expenses, e.g., special schooling for an ADHD child, providing for retirement, making technical and financial progress in the career, etc.?

So, some concerns:

(A) If an organization is to pay the big bucks for very long, e.g., for longer than some fashion fad, then they will likely need some valuable results on their real problems for their real bottom line. So, to evaluate the opportunity, should hear about the real problems and not just a list of technical topics.

(B) For the opportunity for the big bucks to be realistic, really should know where the money is coming from and why. That is, to evaluate the opportunity, need to know more about the money aspects than a $10/hour fast food guy.

(C) As just an employee, can get replaced, laid off, fired, etc. So, to evaluate the opportunity, need to evaluate how stable the job will be, and for that need to know about the real business and not just a list of technical topics.

(D) For success in projects, problem selection and description and tool selection are part of what is crucial. Is the hiring organization really able to do such work for AI, etc. topics?

Or, mostly organizations are still stuck in the model of a factory 100+ years ago where the supervisor knew more and the subordinate was there to add muscle to the work of the supervisor. But in the case of AI, etc., what supervisors really know more or much of anything; what hiring managers know enough to do good problem and tool selection?

Or, if the supervisors don't know much about the technical topics, then usually the subordinate is in a very bad career position. This is an old problem: One of the more effective solutions is some high, well respected professionalism. E.g., generally a working lawyer is supposed to report only to a lawyer, not a generalist manager. Or there might be professional licensing, peer review, legal liability, etc. Or, being just an AI technical expert working for a generalist business manager promises in a year or so to smell like week old dead fish.

(E) If some of the AI, etc., topics do have a lot of business value, then maybe someone with such expertise really should be a founder of a company, harvest most of the value, and not be an employee. So, what are the real problems to be solved. That is, is there a startup opportunity there?

Really, my take is that the OP is, net, talking about a short term fad in some topics long surrounded with a lot of hype. Not good, not a good direction for a career.

AI and hype? Just why might someone see a connection there?

However, with the abysmal standards of hiring, I'm sure a lot of companies would pass on very good candidates because they won't write FizzBuzz on the board for you, or companies would pass on Peter Norvig because his code is not Pep8 compliant

Not sure if you're being serious but are you sure they would even care whether these kinds of candidates could program? I am skeptical they hired them to code; I imagine they mostly spend their time doing research and then have the SDEs implement things.

It happens all the time, and it comes from a mix of things.

One way it happens is that you get a PhD in astrophysics with years of data analysis experience in for a data science job. Have a software engineer interview her and he might find that she doesn't know a number of basic computer sciences concepts [traversing a linked list, tail recursion, implement breadth-first-search]. His knowledge background says these basic ideas are fundamental, there are therefore serious questions about the technical ability of the interviewee.

This is a good example. At the same time this CS interviewer may not know what's the second central moment of a probability density function.

Uhhh... the variance?

I have no idea how hiring in this field works but I'd expect a data scientist to have a pretty good (algorithmic) programming background in this day and age. You pretty much have to "play with the data" and get a good intuition for it when it comes to gigantic data sets and programming is how you accomplish that.

Or in other words...I'd be skeptical if a candidate hadn't learned programming on their own even if it wasn't required because it's pretty much impossible to get any practical experience otherwise.

I think often then have experience in programming, but in languages that don't map well to the "real world" of programming -- R, octave, matlab, etc. Those languages are also usually loaded with very helpful libraries to avoid having to do any nitty-gritty programming.

But there would still need to be a fizzbuzz style filter for AI jobs to keep out the people who overstate their abilities. Is there a good test for basic AI job capability?

You just interview people.

Ask them to explain what a SVM is. Ask them to explain how training a linear perceptron works. This kind of stuff.

That tests book knowledge, not actual understanding of how to think about AI, and leaves itself open to interviewer bias, the very things fizzbuzz avoids.

The idea that FizzBuzz has no interviewer bias is laughable at best.

I can guarantee some will pick and reject people because they did things in a way they didn't like

No it does not. Without practical experience one just cannot discuss these things freely, and pure-book knowledge is clearly visible instantly.

Sketchy, but in the other direction. In high school I did a project that involved a support vector machine, and over the course of this I learnt how they worked to a reasonable level of detail.

Two years later, I interviewed at a large SV company, and they asked what projects I'd been working on. I gave a description of this project (2 years prior) and my explanation was phenomenal; in order to understand SVMs two years prior (without the 'necessary' mathematical background) I needed to develop all of the intuition (up to 'is kinda a high pass filter' etc) (which you might not ordinarily do at a time-pressured university course).

The interviewer was correspondingly impressed, and the SV company gave me an internship almost directly off the back of this interview.

The kicker is that at this point, I had a rudimentary knowledge of linear algebra, and absolutely no knowledge of any other machine learning; I had no business interning in their data science team.

My point being that even as a first pass, the bookwork questions don't work fantastically. FizzBuzz is no better, but a data science alternative would have weeded me out pretty quickly.

You're pretty harsh with yourself. You showed that you were interested in the topic and that you could learn it by yourself. I don't think anyone expect an intern to already master the subject he'll be working on, picking a smart motivated person is usually what you're looking for.

Ah, but if you combine it with the fact that when I did study those things I didn't really enjoy them, you can probably see why I look back on the experience with healthy scepticism :P

The question I'm more interested in: when you worked there as an intern, did you suck?

(If you didn't, maybe the interview worked...)

Apologies in advance for the bluntness, and don't feel like you need to answer.

I didn't work there; I went somewhere else for a number of reasons. I wouldn't have sucked, but it wouldn't have played to my strengths - my work would probably have been very average.

The Chinese room?

I'd hire a fast-enough Chinese room.

How can there be experts on a subject that doesn't yet exist?

Actually there can be anything possible in the articles that are published by so called Tech Journalists who have no idea of the fundamentals of the tech.

What do you mean it doesn't exist? I work in that field, it exists.

What do you mean it exists? How do you definite AI? Can your project reason with you? Or is it simply an Input-Output type of program? Just because you use natural language with it instead of punch cards doesn't mean it is intelligent.

People tried to do AI in the 60s to 90s era. It is dubbed symbolic AI. It didn't work out. A good chance that it never will. Today machine learning algos and a bunch of automated statistics is called "Artificial Intelligence". It's not intelligence at all. Intelligence implies something more than I/O computation.

I tend to agree that AI today is a long way from what the founders of the subject imagined – it's become something more like "Applied Computer Science". But what's now called "Artificial General Intelligence" isn't dead, and people are still working on it.

Also, it's more tricky than you'd think to narrow down what counts as intelligence. There aren't really any hard lines between an I/O program and an intelligent agent, even though they seem pretty far apart.

Just because you don't see the hard lines doesn't mean they aren't there. We are deluding ourselves by avoiding a hard definition of intelligence so we can keep believing that we are creating AI when its really nothing of the sort.

Just because you can't see unicorns doesn't mean they aren't there, but at some point you have to give up the search. It's fine to talk about how, broadly speaking, rats are more intelligent than ants, plants or microbes (which are basically I/O rules with a body), chimpanzees more so than rats, humans more so than than chimps etc. But in general there's a ton of overlap and the qualities we associate with intelligence – memory, planning, self-awareness, tool use, whatever – are only loosely correlated continua.

There are a few more binary measures in intelligence research, such as the mirror test, but at best they're only a small piece of the puzzle. There's no sudden point where everything clicks into place.

Of course, if you have such a good definition of intelligence, feel free to enlighten me.

You are repeating the same mistake as before. The difference is significant between admitting we don't know what those lines are and claiming that they don't exist.

It's not a mistake – I see the difference and I am claiming that those lines don't exist. There's plenty of evidence for that, including the continuous range of intelligent behaviour from plants to humans. It's just an empirical fact that there's no hard line.

Of course, that could be wrong – new evidence may come to light, after all. But even so, it doesn't make any sense to say that trying to understand and replicate intelligence is deluded, just because we don't know where that line is – because figuring out what intelligence is is exactly the problem that people are trying to solve. AI is one part of that, along with more empirical research in fields like cognitive science and biology.

Are people researching quantum gravity deluding themselves because they don't yet have a hard definition of quantum gravity? Figuring that out is exactly the point!

If you are going in the wrong direction, you won't get to your destination by going faster. You should stop and rethink and you won't do this unless you can admit that you might be wrong.

How much research had you done before you assertively proclaimed that those lines don't exist? Because it looks nothing like a smooth transition to me.

Well I would say that "intelligence" is learning and inference with causal models rather than just predictive or correlative models. You can then cash it all out into a few different branches of cognition, like perception (distinguishing/classifying which available causal models best match the feature data under observation), learning (taking observed feature data and using it to refine causal models for greater accuracy), inference (using causal models to make predictions under counterfactual conditions, which can include planning as a special case), and the occasional act of conceptual refinement/reduction (in which a model is found of how one model can predict the free parameters of another).

It's an interesting perspective, but the thing about this kind of definition is that it's very much focused on the mechanism of intelligence rather than the behaviour that the mechanism produces – which flies in the face of our intuition about what intelligence is, I think.

If we found out that one species of chimp learns sign languages through a causal model while another learns it through an associative one (for example) we wouldn't label one more or less intelligent, because it's the end result that matters – don't you think?

Likewise, arguably the ultimate goals of AI are behavioural (machines that can think/solve problems/communicate/create etc.), even if it's been relatively focused on mechanisms lately. Any particular kind of modelling is just a means to that end. Precisely what that end is is still a bit hard to pin down, though.

>Brains are complicated, and there is a huge amount of heterogeneity in how people process information and think about mathematics (and indeed all topics, but it is clearer in mathematics perhaps). Some are very visual, some are big on calculation.

What do you mean by "associative model"? That doesn't map to anything I've heard of in cognitive science, statistics, machine learning, or Good Old-Fashioned AI.

But actually, I would expect different behaviors from an animal that learns language via a purely correlational and discriminative model (like most neural networks) versus a causal model. Causal models compress the empirical data better precisely because they're modelling sparse bones of reality rather than the abundant "meat" of correlated features. You should be able to generalize better and faster with a causal model than with a discriminative, correlative one.

I think I meant correlational, but it was really just a placeholder for "a different model". You could replace the chimp with some kind of alien whose thinking model is completely, well, alien to us – but still proves its intelligence by virtue of having a spaceship.

I'm not necessarily saying that different models lead to exactly the same behaviour. Clearly, chimps' models don't generalise as well as ours do and they don't have a language model that matches ours in capability, for example, which leads to different behaviour. But given that their behaviour is generally thought of as less intelligent as opposed to not intelligent at all, it seems like the mechanism itself is not the important thing.

Experts on a subject that doesn't exist? Like professors of theology?

Theology exists.

AI experts don't know how to create an artificial intelligence. AI researchers study how to solve various problems in CS traditionally performed by humans that humans don't solve by carrying out an algorithm by hand like Natural Language Processing, machine learning, automated reasoning, search (e.g. chess).

There are at least two flavors of A.I. "Strong A.I." seeks to build something human-like and could pass Turings immitation game test. The new movie Ex Machina explores this test. "Weak A.I." replicates just a single cognitive skill like game playing, pattern recognition, natural language or something more trivial. Most recent A.i. worked on the latter, or its theoretical background.

Face recognition, speech recognition, machine translation, text classification and more don't exist?

That's the plot line of a hyper critical, yet insightful, joke that came out of one of the many AI downturns in past decades, that as soon as an algo or implementation works it isn't AI anymore, its (fill in the blank specialization). So AI is just the present set of algos that don't (yet?) work.

So that fuzzy logic, thats not AI anymore, thats a footnote in the EE control systems theory class, isn't it? And the face recognition is a parallel processing assignment in FPGA class, speech recognition is an advanced section in DSP theory class, etc.

Right, it's not yet at that point for face and speech recognition, but that certainly happened to game AI (Chess), constraint solving (Sudoku), and planning.

Reply on news.ycombinator.com