1 00:00:00,000 --> 00:00:02,825 Imagine a robot trying to steer. 2 00:00:02,825 --> 00:00:07,750 Its controller chooses what to do and sends a command: turn left. 3 00:00:07,750 --> 00:00:10,950 The actuator turns that command into movement. 4 00:00:10,950 --> 00:00:14,725 But the observed output is a wheel turning right. 5 00:00:16,225 --> 00:00:19,350 Compare that with the target: turn left. 6 00:00:19,350 --> 00:00:21,900 The observed movement missed the target. 7 00:00:21,900 --> 00:00:25,900 Was this a one-off error, or have the controls reversed? 8 00:00:25,900 --> 00:00:30,225 The controller needs evidence to tell those possibilities apart. 9 00:00:32,225 --> 00:00:36,613 A snapshot of the robot tells us how its controls are configured now. 10 00:00:36,613 --> 00:00:40,800 Following it through time tells us how that configuration changes. 11 00:00:40,800 --> 00:00:44,362 The controls might be normal, then become reversed. 12 00:00:44,362 --> 00:00:50,725 These two views, what is and what changes, are what we mean here by Being and Becoming. 13 00:00:52,225 --> 00:00:54,325 Keep two things separate. 14 00:00:54,325 --> 00:00:57,100 There is the actuator's actual condition. 15 00:00:57,100 --> 00:01:00,750 And there is the controller's estimate of that condition. 16 00:01:00,750 --> 00:01:04,512 The controls can change before the estimate catches up. 17 00:01:04,512 --> 00:01:09,950 New evidence can also correct an estimate when the controls have not changed. 18 00:01:11,950 --> 00:01:19,612 So the small question here is how a system can keep a useful description of its own changing machinery. 19 00:01:19,612 --> 00:01:24,500 A larger question is whether that tells us anything about self-awareness. 20 00:01:24,500 --> 00:01:29,225 We will return to that after seeing exactly what this study tests. 21 00:01:29,875 --> 00:01:33,600 The paper uses a toy model, not a physical robot. 22 00:01:33,600 --> 00:01:35,788 There are three successive actions. 23 00:01:35,788 --> 00:01:40,738 Each command has two possibilities, which we can picture as left and right. 24 00:01:40,738 --> 00:01:46,462 Normal controls usually preserve the command; reversed controls usually swap it. 25 00:01:46,462 --> 00:01:49,600 Random errors can also affect a movement. 26 00:01:50,250 --> 00:01:56,125 Before acting, the controller gets three preliminary checks, called calibration probes. 27 00:01:56,125 --> 00:02:00,688 They can be wrong, so the controller starts with an uncertain estimate. 28 00:02:00,688 --> 00:02:05,150 Before the second action, there is a one-in-four chance of a reversal. 29 00:02:05,150 --> 00:02:08,875 If it happens, it lasts through the third action. 30 00:02:10,875 --> 00:02:13,550 The target is chosen separately. 31 00:02:13,550 --> 00:02:19,962 The controller compares its command with the actual output: did they agree or point opposite ways? 32 00:02:19,962 --> 00:02:24,938 In this model that gives a noisy clue specifically about the actuator. 33 00:02:24,938 --> 00:02:28,425 Real-world causes can be harder to separate. 34 00:02:29,075 --> 00:02:30,475 Follow the loop. 35 00:02:30,475 --> 00:02:32,775 The controller chooses a command. 36 00:02:32,775 --> 00:02:35,200 The actuator produces an output. 37 00:02:35,200 --> 00:02:38,100 Information about that output comes back. 38 00:02:38,100 --> 00:02:43,138 The controller uses it to revise its estimate before choosing another command. 39 00:02:43,138 --> 00:02:46,325 That return of information is feedback. 40 00:02:48,325 --> 00:02:53,050 Bayesian updating is one particular way to use that feedback. 41 00:02:53,050 --> 00:03:00,400 It keeps possible explanations in play and asks: how likely is the new clue under each possibility? 42 00:03:00,400 --> 00:03:06,125 It combines that evidence with what earlier observations already supported. 43 00:03:06,775 --> 00:03:09,525 Here is one example from the paper. 44 00:03:09,525 --> 00:03:17,688 Three noisy checks leave the controller with eighty percent probability for normal controls and twenty percent for reversed. 45 00:03:17,688 --> 00:03:23,400 These are estimates of the controls, not the chance that the next action succeeds. 46 00:03:25,400 --> 00:03:29,000 Now the first action's output agrees with its command. 47 00:03:29,000 --> 00:03:30,925 We observed one agreement. 48 00:03:30,925 --> 00:03:39,213 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. 49 00:03:39,213 --> 00:03:45,713 Those are the chances of seeing this same clue under the two possibilities, not two more observations. 50 00:03:45,713 --> 00:03:49,350 This clue gives more support to normal controls. 51 00:03:50,850 --> 00:03:53,075 Watch where the new weights come from. 52 00:03:53,075 --> 00:03:59,150 For normal controls, take eighty from the prior estimate and multiply by the agreement chance, point nine. 53 00:03:59,150 --> 00:04:00,775 That gives seventy-two. 54 00:04:00,775 --> 00:04:04,375 For reversed controls, twenty times point one gives two. 55 00:04:04,375 --> 00:04:06,538 These weights total seventy-four. 56 00:04:06,538 --> 00:04:12,850 Rescale that total to one hundred percent: about ninety-seven percent normal, three percent reversed. 57 00:04:12,850 --> 00:04:15,525 That change is the Bayesian update. 58 00:04:18,525 --> 00:04:21,700 There is a separate step before action two. 59 00:04:21,700 --> 00:04:26,525 The controller allows for the one-in-four chance that the controls may reverse. 60 00:04:26,525 --> 00:04:30,412 Its normal estimate drops to about seventy-four percent. 61 00:04:30,412 --> 00:04:35,450 No new observation caused that drop; it is allowing for possible change. 62 00:04:35,450 --> 00:04:38,925 The next action uses this revised estimate. 63 00:04:40,925 --> 00:04:45,050 Now we can compare the ways of describing that calculation. 64 00:04:45,050 --> 00:04:51,463 First, the explicit self-model: a named estimate of the controller's own actuator. 65 00:04:51,463 --> 00:04:56,625 Explicit means that the quantity is identified as such in the model. 66 00:04:57,275 --> 00:04:59,712 Second, the compiled version. 67 00:04:59,712 --> 00:05:04,787 That name suggests a rule which maps an observation history to an action. 68 00:05:04,787 --> 00:05:08,912 Here, however, it runs exactly the same update code. 69 00:05:08,912 --> 00:05:11,750 A separate lookup table was not built. 70 00:05:12,400 --> 00:05:14,975 Third, the external observer. 71 00:05:14,975 --> 00:05:19,725 The same estimate is attributed to an observer outside the controller. 72 00:05:19,725 --> 00:05:27,350 It gets the same relevant signals at the same time, with no communication delay or cost in this model. 73 00:05:28,000 --> 00:05:34,662 Give all three the same sequence of events, and every action and probability report matches. 74 00:05:34,662 --> 00:05:37,800 This equality is built into the comparison. 75 00:05:37,800 --> 00:05:45,275 It tells us that changing the label or its assigned location cannot improve these outputs by itself. 76 00:05:48,275 --> 00:05:51,900 The paper also compares two restricted rules. 77 00:05:51,900 --> 00:05:57,150 The one called periodic keeps its starting estimate throughout these three actions. 78 00:05:57,150 --> 00:06:01,687 It uses neither later feedback nor the known chance of reversal. 79 00:06:01,687 --> 00:06:04,375 Think: keep the starting estimate. 80 00:06:05,025 --> 00:06:08,712 The change-point rule allows for the chance of reversal. 81 00:06:08,712 --> 00:06:14,475 A contradiction is an observation that conflicts with its current best guess about the controls. 82 00:06:14,475 --> 00:06:18,687 It counts these instead of reweighting after each observation. 83 00:06:18,687 --> 00:06:22,512 One contradictory observation raises the count to one. 84 00:06:22,512 --> 00:06:28,512 A second consecutive contradiction makes it swap the probabilities and reset the count. 85 00:06:28,512 --> 00:06:32,325 A confirming observation also resets the count. 86 00:06:34,325 --> 00:06:45,975 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. 87 00:06:47,475 --> 00:06:55,325 All five get the same three initial checks, the same kinds of information, and feedback at the same points in time. 88 00:06:55,325 --> 00:06:58,437 What differs is how they use that evidence. 89 00:06:58,437 --> 00:07:02,762 Equality of memory and calculation costs was not established. 90 00:07:02,762 --> 00:07:07,637 Neither was equality of energy use and latency, or engineering effort. 91 00:07:07,637 --> 00:07:11,725 Equal initial checks do not mean equal total cost. 92 00:07:13,225 --> 00:07:17,787 The reported numbers come from an existing frozen benchmark. 93 00:07:17,787 --> 00:07:25,175 This manuscript analyzes those results; it did not run new simulations or collect human data. 94 00:07:25,825 --> 00:07:29,287 The main outcome is how often actions fail. 95 00:07:29,287 --> 00:07:36,425 Across the model's possible episodes, the Bayesian trio fails on about twenty-three out of every hundred actions. 96 00:07:36,425 --> 00:07:40,000 The two-contradiction rule fails on about twenty-seven. 97 00:07:40,000 --> 00:07:43,987 The keep-starting-estimate rule fails on about twenty-nine. 98 00:07:43,987 --> 00:07:49,350 These are model averages, not counts from a hundred physical robot tests. 99 00:07:53,350 --> 00:08:04,637 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. 100 00:08:04,637 --> 00:08:12,200 In this small task, using the evidence in the full Bayesian calculation helps the controller choose commands that work. 101 00:08:12,200 --> 00:08:14,887 But all three versions share that gain. 102 00:08:14,887 --> 00:08:19,950 Calling one version a self-model gives it no additional advantage here. 103 00:08:21,950 --> 00:08:30,337 The updating trio also does better on the paper's short recovery measure and on avoiding two failures after a reversal. 104 00:08:30,337 --> 00:08:34,100 Every rule pays for the same three initial checks. 105 00:08:34,100 --> 00:08:41,400 The improvement concerns these two restricted rules; it does not rank all possible change-detection methods. 106 00:08:42,050 --> 00:08:46,787 The paper also scores forecasts of whether the next action will succeed. 107 00:08:46,787 --> 00:08:51,725 A confidently wrong forecast gets a larger penalty than a cautious wrong one. 108 00:08:51,725 --> 00:08:53,700 This is called a Brier score. 109 00:08:53,700 --> 00:09:02,125 The trio scores better, but that score alone does not show whether its stated probabilities match how often the actions actually succeed. 110 00:09:02,125 --> 00:09:05,837 The actions also change how predictable success is. 111 00:09:05,837 --> 00:09:09,075 No human feeling of confidence was measured. 112 00:09:09,725 --> 00:09:13,650 There are different questions hiding in the word awareness. 113 00:09:13,650 --> 00:09:17,625 Can a system use information about what is happening around it? 114 00:09:17,625 --> 00:09:20,425 Can it track something about its own machinery? 115 00:09:20,425 --> 00:09:23,400 And does anything feel like something to that system? 116 00:09:23,400 --> 00:09:25,700 Keep those questions separate. 117 00:09:27,700 --> 00:09:31,525 This pilot studies a narrow part of the second question. 118 00:09:31,525 --> 00:09:36,512 Its estimate concerns its own actuator, and helps choose commands. 119 00:09:36,512 --> 00:09:41,075 That fits the paper's minimal functional definition of a self-model. 120 00:09:41,075 --> 00:09:47,950 It does not establish a sense of identity, reflection on oneself, or subjective experience. 121 00:09:48,600 --> 00:09:52,950 Nor does the result show that explicit self-models never help. 122 00:09:52,950 --> 00:09:58,750 It shows that the same working calculation succeeds under all three descriptions. 123 00:09:58,750 --> 00:10:04,300 The gains over the two restricted rules come from the way evidence is used. 124 00:10:04,300 --> 00:10:08,850 The self-model label adds no extra measured benefit here. 125 00:10:10,850 --> 00:10:12,600 Return to the robot. 126 00:10:12,600 --> 00:10:17,062 Its controls have a condition now, and that condition can change. 127 00:10:17,062 --> 00:10:22,187 Observation lets the controller revise its description and use it to act. 128 00:10:22,187 --> 00:10:27,312 This is a small example of keeping a useful account of Being through Becoming. 129 00:10:27,312 --> 00:10:33,250 A changing description and a changing machine are connected, but they are not the same event. 130 00:10:35,250 --> 00:10:47,225 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. 131 00:10:47,225 --> 00:10:52,300 The alternatives must get the same information and comparable resources. 132 00:10:52,300 --> 00:10:54,400 The outcome is open. 133 00:10:55,050 --> 00:11:00,200 That would test what a particular self-model enables a system to do. 134 00:11:00,200 --> 00:11:06,725 Whether the system is aware in the sense of having an experience would remain a further question.