Worthwhile Reading and Viewing

A list of common cognitive biases

Formative experiences persist over long spans of time

The current status of artificial intelligence in medicine

Related: Limitations of neural networks   Also: Playing against AI can improve the abilities of human Go players.

The realities of energy storage.  I have observed that very few journalists comprehend the difference between a Kw and a Kwh–and why it matters.  (And this includes business and technology journalists.)

Paul Graham, at twitter:

Putin’s bungled invasion shows us that democracy’s greatest advantage, in the long term, is simply its “guarantee that leaders are regularly replaced.”  He was responding to this article.

Some people create or discover new things. Some enforce social norms. There is little overlap between the two.

Graham (who is one of the very few venture capitalists to have attended art school) also recommends this art history thread.

26 thoughts on “Worthwhile Reading and Viewing”

  1. What? You stole the list of common cognitive bias from Avi. Oh well he stole it too. I have a cognitive bias, its called: “I don’t believe you”. Its served me very well indeed. ;)

  2. Ai has always been ten years away for the 40 years that I’ve been paying attention. I can’t count the number of times I’ve seen some presentation in the media trumpeting some new “break through” and then never hear about it again. They present error rates that seem superficially seem quite low without also explaining that any usable system, like driving a car, require thousands of judgements per second where the errors snow ball out of control. Things like inference engines and neural networks always start out with a simple concept, easily explained in an article maybe with some simple sample code to try yourself. Where it ends up is always MORE POWER!, Moore’s law will make it work when we can just get more computing resources. In ten years, it will be awesome.

    Of course they don’t acknowledge the distinction between GW and GWh, it’s the same thing when the car salesman keeps talking payment instead of price. If they told the truth, people would run the other way. The problem is that batteries will always very limited in density. They are still based on chemical reactions. But these reactions have to be very gentle so that they can first of all be contained in a cell, and that can be reversed, recharged thousands of times. When the first requirement breaks down, we see battery fires and explosions. Utility grade batteries have different constraints than a car or laptop so there have been experiments with chemistries that need high temperatures and noxious chemicals that never seem to go anywhere. The large Tesla batteries are being sold as a loss leader.

  3. the human mind is an incredibly complete machine, much more than any computer, any real cognition is far away from reality,

  4. MCS…”Of course they don’t acknowledge the distinction between GW and GWh, it’s the same thing when the car salesman keeps talking payment instead of price. If they told the truth, people would run the other way.” I’m sure there’s some of that…a lot of that, actually…but I think a lot of them just don’t understand. Journalists tend to pick up and use words and phrases without feeling any need to understand what they actually mean.

    Now imagine trying to explain reactive power to one of them….

  5. Most “journalists” can’t count past ten with their shoes on. You get the same BS from the “authorities” that are supposed to know what they’re talking about. They know enough not to queer the pitch by telling the truth. If they did, everybody would know that these huge, ruinously expensive, fire prone batteries have actual hold up time measured in seconds.

  6. re artificial intelligence: some pretty practical AI applications have emerged over the past decade or so….speech recognition that is actually usable, language translation which, while it may not be elegant, usually suffices to give you some idea of what the foreign language document is about.

  7. but is that really intelligence, I’ve found programs like babelfish and google translate are adequate, but they don’t really capture the intricacies of language,

  8. There are quite a few machine-learning applications for failure prediction and for classification.

    Here’s one for predicting exhaust-valve failures in aircraft piston engines:


    …and here’s one for potential use by emergency-medical technicians in classifying stroke as either large vessel occlusive (LVO) strokes or non-LVO:


  9. My, probably imperfect, understanding of the success of computer assisted mammography is that the computer highlights things that were very subtle and easy to overlook using image processing. This has been facilitated by replacing film with digital imaging. But the call is still made by the radiologist.

    On the other hand, the much ballyhooed application of IBM’s Watson to diagnosis turned into a complete bust. I’m sure the recriminations will echo on for years.

    The chess playing works by playing millions (maybe billions?) of games at every move to try to find the best move. It works but has to be fundamentally different from how humans solve problems. It’s a little like comparing a man and a shovel with a 60 ton excavator.

    Has anybody actually used the grammar part of Word? And machine translation; have you read a Chinese translated manual ever? They end up with a language that outwardly resembles English, uses the same words and forms while not conveying any comprehensible information.

  10. They also ignore the difference between carbon, a somewhat dirty solid (coal, graphite, soot,diamonds) and carbon dioxide a clean gas that gives us beer, bread, fire extinguishers and more.

    Even people who know better talk of “carbon pollution” and get snippy when call them on it. “same thing, really” they try to tell me.

    And have you ever seen co2 expressed as anything other than ppm? OMIGOD! almost 400ppm! We are all going to die.

    They get snippy too when I point out that 400ppm (400/1,000,000) is 0.04%

    We seldom see nitrogen, 780,000, oxygen 210,000 or other gased expressed as ppm.

    Why? Because 400 can be made to look like a scary number in a way that is hard to do with 0.04%.

  11. how do they walk and chew gum, at the same time, yes carbon is the sine quanon of all interaction on this planet

  12. David,

    I used to be quite knowledgeable about power factor, Vars, capacitors and so on.

    35 years later I can’t even explain it to myself.

    I don’t really expect reporters to know anything about it. I would expect them to know that mixing generated ac power and inverted DC power is a problem. Not a huge one but one that must be addressed and one that adds complexity, cost and potential instability to the utility grid

  13. that was the contest between westinghouse and edison was it not, with tesla on the other side,

  14. David,
    Both of your examples show success over very narrow domains. Those sorts of things have been working for a long while. Both show what might be considered the essential prerequisite for machine learning.

    It’s hard to tell for sure, but the stroke classifier works after adding some sort of instrument that measures the patient’s condition to replace the present standard of care which, reading closely, is an algorithm based on the answers to a series of questions. Presumably, the computer is used to interpret the instrument readings and reduce what might be extensive training for the EMT’s otherwise. It’s still the instrument that provides the data rather than the questionnaire. I’ve seen computer learning applied to things like questionnaires too but it seems to work best by being able to examine more data than a person could do easily or quickly for subtle patterns, again, not readily discernible.

    I’d find it pretty hard to find a narrower problem domain than a single component type in an engine. Again, reading closely, they used data from three million flights to spot 60 failures. Like I’ve said, various sorts of pattern recognition has been working for a long time. I don’t know just what sort of data they were looking at but predictive maintenance generally work using things like accelerometers to predict bearing failure. I’m betting they are using data from new style engine instrumentation that records as well as providing the usual indications to the pilot.

    All the big, sexy AI projects that I’ve heard of have been disappointing. We’re still a long way from; “Siri, tell me the meaning of life.”. The systems that work best are those that exploit the differences between computers and humans. Humans are just not good at wading through large quantities of data and finding subtle patterns whether were talking about gauge readings or pixels. But computers still don’t seem to be able to do the things that humans do well like filling in the blanks where all the information isn’t available or processing human language.

  15. The engine failure predictor uses data captured by an engine monitor, specifically exhaust gas temperature and cylinder head temperature, at about 5-second intervals. Recording engine monitors have been around a long time, but previously, you’d have to graph the results and just assess them intuitively.

    The stroke classifier uses EEG data.

  16. Back in the late ’50’s, my dad built a data acquisition system the size of a large roll top desk (mostly in our living room) that plotted 23 data points on a two foot wide paper roll. I think it recorded every minute. It was used on a research project measuring the losses from a 750 KV power line operating at 10,000 feet above sea level. Eventually, he and his research partner spent around six weeks in a motel room transcribing that data so that it could be punched into cards and fed into a computer. There were something like 3.5 million points and many many feet of graph paper.

    Now I routinely build systems that might produce a million points in a day and that’s considered very low bandwidth.

    Even pretty old General Aviation planes now have glass cockpits that remind you of a 787. The engines don’t seem to have improved as much as one would like. They couldn’t have done this sort of predictive maintenance if they were still using mechanical gauges. What I found surprising was the recommendation to bore-scope the cylinder. Bore-scopes have become so cheap and since all it requires is going in the spark plug hole, I would wonder why it wasn’t routine for all the cylinders. Like I said, engine failures are still a big problem and I think I’d want to know if I had a valve that was burned or a cylinder that was scuffed.

    I tried to find what the stroke instrument did, but I my eye wasn’t sharp enough to pick it out. Again, my imperfect understanding is that EEG’s are very hard to interpret, else why would neurologists spend so much time in school and post doc. And it’s not likely that it would be possible to train EMT’s to do it. It isn’t surprising that a portable one with the ability to read itself will improve accuracy. But it’s still operating in a reduced domain. Still, the real advance is making it possible for non-specialists to use it, so a fair use of AI. I believe it’s been possible to produce one for a while, just not practical to train people to use them reliably on a wide basis.

    I’m still waiting for my self driving, flying car.

  17. “air” bags are another pet peeve.

    No air is involved in deployment. It is a violent pyrotechnic explosion

  18. Mcs,

    I bought a borescope in 1979 to inspect internal welds on a high purity water system. It cost @$20,000. It had rather limited utility.

    4-5 years ago I bought a borescope on Amazon for about $15.

    It has a metal capsule about 3/8″ x 1″. One end has a camera and 4 Led lights. The other has a 15′ USB cable that plugs into a phone or tablet. I’ve tried it with a 30′ extension and it works fine. Very high resolution and very useful addition to any toolkit.

    On my phone it works like a Webcam. View or clip video or snapshots. Shsre them in real time.

  19. Miguel,

    Taking nothing away from tesla or Westinghouse, Edison was actually right. DC is much better than AC for a lot of reasons.

    One big drawback, and why AC won out wis distribution. It is very easy to transform 1,000 volt AC to 50,000vac using nothing more than a bunch of dumb copper wire. Then send the juice to point of use and transform it back to a usable voltage like 110.

    At the time it was very difficult to do this with DC. Edison envisioned a generator on every block.

    It is much easier to do now though still not what anyone would call “easy”. It is being used for some super high voltage 700kv high capacity long distance projects.

    When I was in the navy 80-90% of electricity on ships was DC. Motor-generator sets (DC motor driving an ac generator) were used to meet AC requirements.

  20. Again, my imperfect understanding is that EEG’s are very hard to interpret, else why would neurologists spend so much time in school and post doc. And it’s not likely that it would be possible to train EMT’s to do it. It isn’t surprising that a portable one with the ability to read itself will improve accuracy.

    Back in the 70s when I was in practice my hospital, like many small hospitals, used an EEG tech who had his own machine and drove around in his pickup from hospital to hospital. The tech I knew, a guy named Mike Warren, was as good as a neurologist at reading them. I realized this one time when a patient of mine, whose ruptured aneurysm I had successfully repaired a month earlier, came in with the symptoms of a stroke. This was pre-CAT scan and MRI so the stroke workup included an EEG. Mike was doing the EEG when he told me to call the neurosurgeon because “This guy has a subdural.”

    The patient had fainted from the pain when his aneurysm ruptured. He had hit his head at the time. Subdurals bleed into the space between the dura mater, which covers the brain, and the brain. It is venous bleeding from small veins that are torn when the brain moves from a blow or a fall. The venous bleeding usually does not cause symptoms until a month later, when the red cells break down and the blood clot absorbs water. It swells and puts pressure on the brain. Mike saw this as the EEG paper was coming out of the machine. I helped the neurosurgeon that night to drain the subdural. The fluid looked like crankcase oil. By morning he was back to normal with no neurological symptoms.

Comments are closed.