WEBVTT 1 00:00:00.200 --> 00:00:04.404 A controller can misread a result if its own machinery has changed. 2 00:00:04.404 --> 00:00:09.448 Tracking that machinery could help it choose better actions, but the benefit needs a 3 00:00:09.448 --> 00:00:10.707 fair comparison. 4 00:00:10.707 --> 00:00:16.480 Watch the actual actuator condition and the controller's estimate as separate things, 5 00:00:16.480 --> 00:00:19.320 and ask what evidence supports a claimed advantage. 6 00:00:20.560 --> 00:00:25.000 This existing film asks when tracking a changing actuator helps. 7 00:00:25.000 --> 00:00:29.147 Its actual condition and the controller’s estimate are separate. 8 00:00:29.147 --> 00:00:34.347 It has its own calibration checks, reversal chance and comparison policies. 9 00:00:34.347 --> 00:00:37.867 Do not import our earlier noisy-reader numbers. 10 00:00:37.867 --> 00:00:46.720 Watch for two different comparisons: three versions sharing the same calculation, and two restricted rules using evidence differently. 11 00:00:48.040 --> 00:00:50.650 Imagine a robot trying to steer. 12 00:00:50.650 --> 00:00:55.120 Its controller chooses what to do and sends a command: turn left. 13 00:00:55.120 --> 00:00:58.420 The actuator turns that command into movement. 14 00:00:58.420 --> 00:01:01.660 But the observed output is a wheel turning right. 15 00:01:01.660 --> 00:01:04.420 Compare that with the target: turn left. 16 00:01:04.420 --> 00:01:06.840 The observed movement missed the target. 17 00:01:06.840 --> 00:01:10.760 Was this a one-off error, or have the controls reversed? 18 00:01:10.760 --> 00:01:14.440 The controller needs evidence to tell those possibilities apart. 19 00:01:15.230 --> 00:01:19.174 A snapshot of the robot tells us how its controls are configured now. 20 00:01:19.174 --> 00:01:23.590 Following it through time tells us how that configuration changes. 21 00:01:23.590 --> 00:01:26.857 The controls might be normal, then become reversed. 22 00:01:26.857 --> 00:01:33.054 These two views, what is and what changes, are what we mean here by Being and Becoming. 23 00:01:33.054 --> 00:01:34.750 Keep two things separate. 24 00:01:34.750 --> 00:01:37.490 There is the actuator's actual condition. 25 00:01:37.490 --> 00:01:40.510 And there is the controller's estimate of that condition. 26 00:01:40.510 --> 00:01:43.950 The controls can change before the estimate catches up. 27 00:01:43.950 --> 00:01:48.377 New evidence can also correct an estimate when the controls have not changed. 28 00:01:48.377 --> 00:01:54.590 So the small question here is how a system can keep a useful description of its own changing machinery. 29 00:01:54.590 --> 00:01:58.963 A larger question is whether that tells us anything about self-awareness. 30 00:01:58.963 --> 00:02:02.830 We will return to that after seeing exactly what this study tests. 31 00:02:03.540 --> 00:02:06.980 The paper uses a toy model, not a physical robot. 32 00:02:06.980 --> 00:02:09.252 There are three successive actions. 33 00:02:09.252 --> 00:02:14.374 Each command has two possibilities, which we can picture as left and right. 34 00:02:14.374 --> 00:02:19.906 Normal controls usually preserve the command; reversed controls usually swap it. 35 00:02:19.906 --> 00:02:22.786 Random errors can also affect a movement. 36 00:02:22.786 --> 00:02:28.996 Before acting, the controller gets three preliminary checks, called calibration probes. 37 00:02:28.996 --> 00:02:33.517 They can be wrong, so the controller starts with an uncertain estimate. 38 00:02:33.517 --> 00:02:37.940 Before the second action, there is a one-in-four chance of a reversal. 39 00:02:37.940 --> 00:02:41.200 If it happens, it lasts through the third action. 40 00:02:41.200 --> 00:02:43.700 The target is chosen separately. 41 00:02:43.700 --> 00:02:50.047 The controller compares its command with the actual output: did they agree or point opposite ways? 42 00:02:50.047 --> 00:02:54.564 In this model that gives a noisy clue specifically about the actuator. 43 00:02:54.564 --> 00:02:57.540 Real-world causes can be harder to separate. 44 00:02:58.410 --> 00:02:59.610 Follow the loop. 45 00:02:59.610 --> 00:03:01.610 The controller chooses a command. 46 00:03:01.610 --> 00:03:03.850 The actuator produces an output. 47 00:03:03.850 --> 00:03:06.610 Information about that output comes back. 48 00:03:06.610 --> 00:03:11.338 The controller uses it to revise its estimate before choosing another command. 49 00:03:11.338 --> 00:03:14.028 That return of information is feedback. 50 00:03:14.028 --> 00:03:18.223 Bayesian updating is one particular way to use that feedback. 51 00:03:18.223 --> 00:03:24.890 It keeps possible explanations in play and asks: how likely is the new clue under each possibility? 52 00:03:24.890 --> 00:03:29.578 It combines that evidence with what earlier observations already supported. 53 00:03:29.578 --> 00:03:32.250 Here is one example from the paper. 54 00:03:32.250 --> 00:03:39.477 Three noisy checks leave the controller with eighty percent probability for normal controls and twenty percent for reversed. 55 00:03:39.477 --> 00:03:44.610 These are estimates of the controls, not the chance that the next action succeeds. 56 00:03:44.610 --> 00:03:48.170 Now the first action's output agrees with its command. 57 00:03:48.170 --> 00:03:49.910 We observed one agreement. 58 00:03:49.910 --> 00:03:58.117 The model predicts an agreement nine times out of ten if the controls are normal, but only one time out of ten if they are reversed. 59 00:03:58.117 --> 00:04:04.634 Those are the chances of seeing this same clue under the two possibilities, not two more observations. 60 00:04:04.634 --> 00:04:07.983 This clue gives more support to normal controls. 61 00:04:07.983 --> 00:04:10.250 Watch where the new weights come from. 62 00:04:10.250 --> 00:04:18.100 For normal controls, take eighty from the prior estimate and multiply by the agreement chance, point nine. 63 00:04:18.100 --> 00:04:20.588 That gives seventy-two. 64 00:04:20.588 --> 00:04:25.758 For reversed controls, twenty times point one gives two. 65 00:04:25.758 --> 00:04:29.025 These weights total seventy-four. 66 00:04:29.025 --> 00:04:36.245 Rescale that total to one hundred percent: about ninety-seven percent normal, three percent reversed. 67 00:04:36.245 --> 00:04:39.787 That change is the Bayesian update. 68 00:04:39.787 --> 00:04:43.314 There is a separate step before action two. 69 00:04:43.314 --> 00:04:48.394 The controller allows for the one-in-four chance that the controls may reverse. 70 00:04:48.394 --> 00:04:52.634 Its normal estimate drops to about seventy-four percent. 71 00:04:52.634 --> 00:04:57.561 No new observation caused that drop; it is allowing for possible change. 72 00:04:57.561 --> 00:05:00.501 The next action uses this revised estimate. 73 00:05:01.371 --> 00:05:05.144 Now we can compare the ways of describing that calculation. 74 00:05:05.144 --> 00:05:11.781 First, the explicit self-model: a named estimate of the controller's own actuator. 75 00:05:11.781 --> 00:05:16.165 Explicit means that the quantity is identified as such in the model. 76 00:05:16.165 --> 00:05:18.407 Second, the compiled version. 77 00:05:18.407 --> 00:05:23.239 That name suggests a rule which maps an observation history to an action. 78 00:05:23.239 --> 00:05:27.159 Here, however, it runs exactly the same update code. 79 00:05:27.159 --> 00:05:29.706 A separate lookup table was not built. 80 00:05:29.706 --> 00:05:32.339 Third, the external observer. 81 00:05:32.339 --> 00:05:36.732 The same estimate is attributed to an observer outside the controller. 82 00:05:36.732 --> 00:05:43.319 It gets the same relevant signals at the same time, with no communication delay or cost in this model. 83 00:05:43.319 --> 00:05:50.139 Give all three the same sequence of events, and every action and probability report matches. 84 00:05:50.139 --> 00:05:53.024 This equality is built into the comparison. 85 00:05:53.024 --> 00:05:59.291 It tells us that changing the label or its assigned location cannot improve these outputs by itself. 86 00:06:00.611 --> 00:06:03.811 The paper also compares two restricted rules. 87 00:06:03.811 --> 00:06:08.958 The one called periodic keeps its starting estimate throughout these three actions. 88 00:06:08.958 --> 00:06:13.184 It uses neither later feedback nor the known chance of reversal. 89 00:06:13.184 --> 00:06:16.151 Think: keep the starting estimate. 90 00:06:16.151 --> 00:06:19.771 The change-point rule allows for the chance of reversal. 91 00:06:19.771 --> 00:06:25.504 A contradiction is an observation that conflicts with its current best guess about the controls. 92 00:06:25.504 --> 00:06:29.411 It counts these instead of reweighting after each observation. 93 00:06:29.411 --> 00:06:33.451 One contradictory observation raises the count to one. 94 00:06:33.451 --> 00:06:39.531 A second consecutive contradiction makes it swap the probabilities and reset the count. 95 00:06:39.531 --> 00:06:42.779 A confirming observation also resets the count. 96 00:06:42.779 --> 00:06:54.901 That gives us five policies: the explicit self-model, compiled version, and external observer, plus the two-contradiction rule and the keep-starting-estimate rule. 97 00:06:54.901 --> 00:07:02.609 All five get the same three initial checks, the same kinds of information, and feedback at the same points in time. 98 00:07:02.609 --> 00:07:05.797 What differs is how they use that evidence. 99 00:07:05.797 --> 00:07:10.411 Equality of memory and calculation costs was not established. 100 00:07:10.411 --> 00:07:15.508 Neither was equality of energy use and latency, or engineering effort. 101 00:07:15.508 --> 00:07:19.601 Equal initial checks do not mean equal total cost. 102 00:07:20.391 --> 00:07:24.279 The reported numbers come from an existing frozen benchmark. 103 00:07:24.279 --> 00:07:31.291 This manuscript analyzes those results; it did not run new simulations or collect human data. 104 00:07:31.291 --> 00:07:34.768 The main outcome is how often actions fail. 105 00:07:34.768 --> 00:07:43.051 Across the model's possible episodes, the Bayesian trio fails on about twenty-three out of every hundred actions. 106 00:07:43.051 --> 00:07:47.551 The two-contradiction rule fails on about twenty-seven. 107 00:07:47.551 --> 00:07:52.271 The keep-starting-estimate rule fails on about twenty-nine. 108 00:07:52.271 --> 00:07:57.657 These are model averages, not counts from a hundred physical robot tests. 109 00:07:57.657 --> 00:08:09.711 So, on average, the updating trio avoids about four failures per hundred actions compared with waiting for two contradictions, and about six compared with keeping the starting estimate. 110 00:08:09.711 --> 00:08:17.824 In this small task, using the evidence in the full Bayesian calculation helps the controller choose commands that work. 111 00:08:17.824 --> 00:08:20.684 But all three versions share that gain. 112 00:08:20.684 --> 00:08:25.784 Calling one version a self-model gives it no additional advantage here. 113 00:08:25.784 --> 00:08:34.097 The updating trio also does better on the paper's short recovery measure and on avoiding two failures after a reversal. 114 00:08:34.097 --> 00:08:38.184 Every rule pays for the same three initial checks. 115 00:08:38.184 --> 00:08:45.264 The improvement concerns these two restricted rules; it does not rank all possible change-detection methods. 116 00:08:45.264 --> 00:08:50.244 The paper also scores forecasts of whether the next action will succeed. 117 00:08:50.244 --> 00:08:55.508 A confidently wrong forecast gets a larger penalty than a cautious wrong one. 118 00:08:55.508 --> 00:08:57.824 This is called a Brier score. 119 00:08:57.824 --> 00:09:06.984 The trio scores better, but that score alone does not show whether its stated probabilities match how often the actions actually succeed. 120 00:09:06.984 --> 00:09:11.071 The actions also change how predictable success is. 121 00:09:11.071 --> 00:09:13.764 No human feeling of confidence was measured. 122 00:09:15.004 --> 00:09:18.104 There are different questions hiding in the word awareness. 123 00:09:18.104 --> 00:09:21.784 Can a system use information about what is happening around it? 124 00:09:21.784 --> 00:09:24.664 Can it track something about its own machinery? 125 00:09:24.664 --> 00:09:27.772 And does anything feel like something to that system? 126 00:09:27.772 --> 00:09:29.740 Keep those questions separate. 127 00:09:29.740 --> 00:09:33.004 This pilot studies a narrow part of the second question. 128 00:09:33.004 --> 00:09:37.612 Its estimate concerns its own actuator, and helps choose commands. 129 00:09:37.612 --> 00:09:41.804 That fits the paper's minimal functional definition of a self-model. 130 00:09:41.804 --> 00:09:48.204 It does not establish a sense of identity, reflection on oneself, or subjective experience. 131 00:09:48.204 --> 00:09:52.044 Nor does the result show that explicit self-models never help. 132 00:09:52.044 --> 00:09:56.964 It shows that the same working calculation succeeds under all three descriptions. 133 00:09:56.964 --> 00:10:01.704 The gains over the two restricted rules come from the way evidence is used. 134 00:10:01.704 --> 00:10:05.404 The self-model label adds no extra measured benefit here. 135 00:10:06.274 --> 00:10:07.674 Return to the robot. 136 00:10:07.674 --> 00:10:11.754 Its controls have a condition now, and that condition can change. 137 00:10:11.754 --> 00:10:16.706 Observation lets the controller revise its description and use it to act. 138 00:10:16.706 --> 00:10:21.794 This is a small example of keeping a useful account of Being through Becoming. 139 00:10:21.794 --> 00:10:27.294 A changing description and a changing machine are connected, but they are not the same event. 140 00:10:27.294 --> 00:10:38.994 One useful next test is to give the competing designs the same memory limit, then ask whether an explicitly organized actuator model helps with new actuator changes. 141 00:10:38.994 --> 00:10:43.494 The alternatives must get the same information and comparable resources. 142 00:10:43.494 --> 00:10:45.042 The outcome is open. 143 00:10:45.042 --> 00:10:49.264 That would test what a particular self-model enables a system to do. 144 00:10:49.264 --> 00:10:56.034 Whether the system is aware in the sense of having an experience would remain a further question. 145 00:10:56.264 --> 00:11:08.944 Before continuing, did the self-model label give the shared-code trio an extra measured advantage in this comparison? 146 00:11:08.944 --> 00:11:15.904 No. The three versions shared a calculation; differences versus the restricted rules concerned evidence use. 147 00:11:15.904 --> 00:11:20.244 Equal information did not establish equal total resources. 148 00:11:20.244 --> 00:11:29.904 The final lesson asks which claims in our playlist were definitions, deductions, checked implementations, reported model results or proposed tests.