Recently I came across a well-meaning, disillusioned Engineering PhD student from a top engineering school, not unlike me who has come to the conclusion that most academic research is useless and should we even fund them? I would probably go further and ask whether we should penalize the researchers who are wasting tax payer’s money, but I digress.
His point was simple, it seems like a large area of research is becoming a joke that will never have any practical relevance and in fact, there doesn't seem to even be any pathway towards practical relevance. He gives examples in Power Engineering with the hottest paper mill, machine learning. The responses were honestly quite predictable. One group called him an idiot for claiming that he can judge practical relevance, surely the scientists know better and hey number theory eventually became useful in cryptography (am I rite). I wonder if they will use this one example to completely negate an argument logic on climate change? Seems pretty cold outside, doesn’t it. Another response hilariously was that research that does not end up practically useful, itself is useful as it tells others “What doesn’t work”. I’d be all for cancel culture if it cancels people like him. I can see the racism and whiteness in his words, can you not? I can do research and tell you that Machine Learning cannot in fact land a rocket on the moon, or I can ask a control engineer and accept his “You’re insane” response and be on my way. I’m not against moon-shots, sometimes they do work out and humanity is better off for it but since it has such a humongous failure rate, I think it is should be common decency that you pursue moonshots with your own funding. The search space of all possible solutions is infinitely large, limited only by human imagination and saying that research into any and every area in this search space is useful is like saying a navigator that gives out random directions is useful as it shows you which directions don’t work. Alas, unfortunately, it seems like the above post doesn’t seem to try to understand why research is in such a dismal state, or even convincingly argue that it is in such a dismal state and so I will put the nail in the coffin as if it matters in more words than Elon Musks heuristic: “95% of all academic papers are useless” which is honestly way too generous for my standards. Should be more like 99.5%.
I honestly have little to no problem with natural sciences research (Math, Physics, Chemistry etc). Looking at the log distribution of research, it does seem that a fraction of scientists contribute to more than 80% of the knowledge, so I would maybe downsize the amount of research done. But they seem to be suffering a lot anyway (3+ postdocs lol), so let’s cut them some slack. Engineering Research though is an entirely different matter. There are two types of engineers in the society, let us lovingly call them ‘Cosmopolitans’ and ‘Townies’. Or if we look to the past, India had a nice way of categorizing them as the “Brahmins” and the “Vaishyas”. The Vaishyas are, of course, no disrespect to the Brahmins, the productive members of the society, the actual engineers who make stuff. Recently, I have found possibly the best Vaishya in the world. So many projects, so much precise and useful engineering, find me any professor or so-called Engineering expert with even half the portfolio of useful inventions as him. In the current climate, 100 research papers would be equivalent to his one research project. I exaggerate but only a little. It turns out that he has joined Boston Dynamics, the perfect example of a Brahmin-Vaishya synergy. Things as they should be. Unfortunately in the current world, Brahmins are out to get the Vaishyas and terrorize them to remind them who rules, not unlike Ancient India. The Brahmins, of course, are the academics, the grad students and the research community as a whole. The townie-cosmopolitan analogy is an apt comparison to the engineer-researcher dichotomy. The Brahmin is the creative thinker, the inventor of new technologies. He is after eminence, not lowly townie pursuits like money (Just look at how professors sneer at those who leave to go to industry). Starting a startup is fine for a Brahmin but only for reasons “to make the earth a better place”, “To improve Humanity” and increasingly “to fight climate change”. He will never start a startup to get rich, that is only reserved for the lowly townies. The Brahmins regularly may live insane and completely dysfunctional lives filled with drugs, polyamory or are insane over-achievers as long as it signals how different, cool and creative they are. The townies, on the other hand, go to the actual task of engineering. They don’t invent new technologies, instead, adapt them for the needs of society (in no ways a simple problem and in many ways a much harder problem as any startup that experienced the scale problem will let you know.) Townies tend to be more well adjusted, more hard-working, more precise engineers. They are rarely insane over-achievers or members of significant counter-cultures, they live lives similar to the status quo of their times. Every society in this world at all times requires both classes of people though of course brahmins are always in smaller demand than townies. Brahmins, however, ultimately are far more responsible for the success or failure of the society.
Good Brahmin research almost always creates a large number of skills for townies to learn and subsequently a large number of new jobs for townies. Research that makes it to industry, even if not as a product then at least as a relevant skill for the engineers in the industry is the only useful type of research. Lots of what we know like transistors, networks, VLSI design fall into this paradigm. Let me stick to my expertise in robotics. The most useful thing to possibly come out of Robot Controls is the Proportional Integral Derivative (PID) Controller, invented by Nicolas Minorsky a Russian-American Researcher in 1920, who was a professor at the University of Pennsylvania for a while. He did of course invent this while working for the military which has an incredible track record of creating useful technologies (something we will investigate later). Almost the entire industrial Robotic arms use PID Controllers and Computed Torque Control (A more modern version). Developing PID Controllers is unfortunately quite hard. It requires a lot of trial and error, experience and intuition to do it correctly. Basically a Townie kind of task. It is a skill, no creative thinking required but instead a skill you develop while making robots and tuning controllers. Modern Academics of course hate this. They don’t proclaim to hate townies, as they always love it when a townie uses their creations and they sneer at anyone who leaves the brahmins to become a townie, even then I don’t think they sincerely hate engineers who work in the industry. They do however hate the townie sort of work, they hate doing townie work and regularly see if they can completely get rid of the townie work if possible. Welcome to auto-tuning PID Controllers, a useless research field that provides immense complexity while simultaneously eliminating the need for a townie. The academic doesn’t need to do actual engineering, he increasingly doesn’t need to even hire townies to help him, how great, how useful. Assuming this does actually not be hopelessly complex, even then it often replaces one townie for another townie. Barring that, the industry rarely if ever asks for this. Spending any time in the industry disabuses you of the notion that the industry wants to get rid of its engineers, especially in robotics. Most often, they don’t mind hiring as many engineers as required as long as the robots work. A good engineer is a very valuable resource, and anything that makes his entire skillset useless is more of a headache to the company than a boon-send. The only person who sees this as a positive good is the stock conscious Brahmin CEO whose only known optimization is down-sizing the company, and “optimizing the numbers”. Yet, academics who rarely have any access to actual engineers spend a lot of time trying to automate the engineers work rather than trying to create new work for the engineer. We see this completely in the newest fad in robotics, “Reinforcement Learning”. Designing controllers that work in the real world is in fact really hard work and requires domain expertise, manual tuning, basically skilled engineers. Fuck that! Let’s automate this. Let’s use Reinforcement Learning and put all these control engineers out of work. If you think I’m exaggerating, let me read a quote from a foundational paper that applied Reinforcement Learning to actual Quadruped Robotics Control: “Designing agile locomotion for quadruped robots often requires extensive expertise and tedious manual tuning”. Imagine that, you require engineers to design a quadruped robot. What Blasphemy! The number of researchers and the amount of resource spent on doing this is unholy. So much so, that the lead researcher ex-Harvard professor at Boston Dynamics had to acclaim in a recent interview regarding the use of Reinforcement Learning said, “We have no interest in automating the work that an engineer competent with optimization can do”. He only saw value in RL if it can augment the current capabilities of robots, as he should. If we require skilled engineers to perform a task, then so be it. Automating that task should be the last priority, something that can be explored only when you’re at a dead end with nothing else to do. Hilariously it seems like Brahmins have realized that Reinforcement Learning actually requires a lot of townie work to get right, such as tuning, reward shaping, and model design. Not that townies are actually doing this work. They still can’t deal with it, so let’s automate it. Meta-RL, RL without Reward Shaping and so on. Ask anyone in industry if they think tuning an RL model is a legitimate problem and they’ll ask incredulously “Surely, you are joking”. If they thought that was all that was required to make the robots useful, they’d have done it years ago. Academics complained that results in simulation were not transferring to the real world, and so Google to its credit decided to do the biggest experiment of RL to date. They trained 10 robots on the same network with reinforcement learning on the actual world for months to collect enough data, and you better believe they tuned their models correctly and it seems like after all this they just threw away their results as worthless (Led to a paper though, guess we now know this technique doesn’t work. They could have just asked me) This is legitimately unhealthy. Academics need skilled engineers to perform and skilled engineers need Academics to do research. The best teams are skilled academics, combined with skilled engineers. Just look at Boston Dynamics, which constantly comes out with new results for its robots that academics can only dream of. Or look at the Military funded projects, which often provide academics with all the resources they need to start solving the problem they want to solve. Instead of a workable synergy, we have a pointless antagonism. A useful heuristic can be to do work that will create a lot of townie jobs, not eliminate them. If that doesn’t work, just ask the industry what developments they want and then focus all your effort on that. No matter how hard it is. People who can’t do this should just become townies anyway. It is for their good, as townies live much happier and less stressful lives. Overproduction of PhD’s is not helping anybody. A professor needs to decide if he needs a graduate student or an engineer. We need lesser professors who are funded much more so that they can hire good engineers. Alas, at this point this is becoming a wish-list. I can only wish for what is possible and unfortunately in this world, not much is possible.