Comments by "Tony Wilson" (@tonywilson4713) on "Don't Be Fooled By Big Tech's Ai Marketing Scam" video.

  1. Aerospace engineer here: I've over 30 years working in industrial control systems, automation and robotics. I am going to try and explain were some of this confusion comes from. First there are 2 words we use a lot in robotics and automation, teaching and learning that have very different meanings to us our work than they do for the general population. In the AI field the word "training" is also a word that's got a completely different meaning to what the general population knows or uses. I'll do this in 3 sub comments so that this isn't one giant lecture like tome of confusion.
    3
  2. TEACHING INUDSTRIAL ROBOTS: 2 of the most common tasks industrial robots do are welding and pick & place. Welding is mostly a matter of following the joint line between 2 pieces of metal. Pick and place is our term for tasks like stacking boxes on a pallet as in we pick something up and place it down which is a very common task in robotics. For all the robot guys out there. This isn't for you this is for all the non-robot programmers so things are more phonetic than code. Imagine we have a robot and has a gripper that closes to grab hold of a part and opens the let that part go and we want it to pick that part up from Point A place and put it down at Point B. To do the most basic pick and place we have to get the robot to know at least 3 points. There's the pick position (Point A), the place position (Point B) as well as a home position which we'll call PHome where its out of the road of everything else . Now to make a robot work we don't just tell it to go to a position because it will (in general) take the fastest path which might not be a straight line. Robots aren't that smart and will simply crash into or through whatever is in the road. Even the best anti-collision software (like airbags) only limits the damage. So we never just tell the robot to go to the safe position. In most cases we tell it to go vertically up to get clear of everything so it can then go directly to the home position without hitting anything. So we end up with a list if commands that looks a bit like this except its all phonetic rather than in code. 1-move (current location + home position height) 2-move (home position) 3-wait for a go signal. 4-move (pick position + vertical clearance) 5-move (pick position) 6-pick part (as in close the gripper) 7-move (pick position + vertical clearance) 8-move (place position + vertical clearance) 9-move (place position) 10-place part (as in open the gripper) 11-goto step 1 NOW TO MAKE IT WORK we have to TEACH the robot those 3 points and this is where the word tech has a completely different meaning to how normal people use the word teach. We take the hand control pendant and MANUALLY DRIVE the robot from point to point and when we get to the location we want with the right orientation we will go into the right menu on the pendant and RECORD that location and save it to memory. We don't stand there and tell it something. It doesn't have a camera and a brain to work it out. We have to MANUALLY drive it to the right location and record that location. Things like vertical clearances we'll work out along the way and just put them in as distances from the location. As things get more complex we have to go around things we just "teach" more points for the robot to go to or go through. We don't just move-stop-move-stop we'll also move-go through the next point, then through the next point,........ and onto the final point. For really complex paths which we get with things like welding and gluing we'll do the programing on a cad system then go out to the robot and "teach" it some reference points and then link the complex paths to those reference points. When we "teach" a robot we don't talk to it we don't read stories to it. We drive it to a location and record that location. Once it starts going it will simply crash into or go through whatever is in the road. The only comment in movies or TV I ever saw was a comment in the film the "Terminator" when he said it does not think and can't be reasoned with it just has its program and will do it. This is why we encase them in gages with interlocked gates for normal operation.
    1
  3. AUTOMATION SYSTEMS and LEARNING This is again another word we use in engineering that maybe we shouldn't because in engineering its nothing like what learning is for humans. The most common place in my work where we see this is in what we call closed loop control which is where we use some kind of sensor to loop back into the control system. The alternative to closed loop is open loop and we use the basic data of a system to estimate what's needed. Its simpler and suitable where accuracy isn't required. But where you need accuracy you have a sensor feeding back data. If you draw a diagram of this it has what looks like a loop - hence the term closed loop. Between the sensor and what ever you are controlling which could be the speed of a pump or the position of a control valve there will be some kind of algorithm and that algorithm will have parameters that need setting up. At the most basic level we call setting up algorithms "tuning" as in we are tuning the algorithm. More complex systems have statistical data analysis algorithms constantly refine the system over time. In essence these systems are LEARNING how to do the job better and better over time BUT THEY ARE NOT learning like a human being does. They just collect data and use statistical analysis (or similar) to refine the parameters. I once installed an oven on a process line. When we first set it up, its controller would take about 45 minutes to heat the up the oven and settle the temperature down before production could start. Over the first 9 months (or so) it collected enough data from each days start up to refine its parameters and got that 45 minutes down to around 15 minutes giving them about an extra 1/2 hour each day of production. The controller did that without any human involvement at all. To a non-engineer these days it might sound like we installed an AI on that oven - WE DIDN'T. As I said it was just a data analysis package. In past generations they'd usually task a young engineer to collect data over several months and do all that analysis by hand as a character building exercise. If you listen to what people are saying about systems like ChatGPT its not that much different except the data its analysing is NOT something as basic as a temperature or pressure sensor its analysing a massive amount of written information. But they do a similar thing in that they effectively average out all the data which is why some people are calling them stochastic parrots. The danger is assuming the stochastic parrot was trained on good data because just like having a faulty sensor causes problems so does a faulty data set for teaching a language AI like ChatGPT.
    1