WEBVTT - ThermoCog Audio Briefing: Why 20-Watt Biomorphic AGI Must Sleep

1
00:00:00.640 --> 00:00:02.320
You know, when you think about cutting-edge

2
00:00:02.320 --> 00:00:07.440
artificial intelligence today, you, you probably picture massive

3
00:00:07.440 --> 00:00:11.280
warehouses just full of glowing servers, humming away,

4
00:00:11.280 --> 00:00:13.705
pulling down enough electricity to power a small

5
00:00:13.705 --> 00:00:16.585
city. Yeah. Exactly. I mean, we basically treat

6
00:00:16.585 --> 00:00:18.905
intelligence like an industrial mining operation right now.

7
00:00:18.905 --> 00:00:22.585
We just throw more power, more data, well,

8
00:00:22.585 --> 00:00:25.545
more brute force at the problem. But today's

9
00:00:25.545 --> 00:00:27.625
deep dive flips that entirely on its head.

10
00:00:27.870 --> 00:00:30.750
We're unpacking a really dense, honestly mind-bending

11
00:00:30.750 --> 00:00:33.630
whitepaper on biomorphic AGI. Right. And the

12
00:00:33.630 --> 00:00:36.430
core mission here is to understand why this

13
00:00:36.430 --> 00:00:40.030
brute force approach of massive datacenters is

14
00:00:40.350 --> 00:00:42.575
hitting a physical wall. And why the ultimate

15
00:00:42.575 --> 00:00:45.775
solution is a biological blueprint for AI that

16
00:00:45.775 --> 00:00:48.175
runs on just 20 Watts of power. Okay.

17
00:00:48.175 --> 00:00:50.335
Let's unpack this. Why can't we just keep

18
00:00:50.335 --> 00:00:53.055
building bigger digital datacenters? Like, what is

19
00:00:53.055 --> 00:00:55.170
physically stopping us? Well, it comes down to

20
00:00:55.170 --> 00:00:57.570
a structural flaw called the Von Neumann bottleneck.

21
00:00:57.570 --> 00:01:00.050
So in modern digital hardware, your memory storage

22
00:01:00.050 --> 00:01:03.970
and your computational processors are—they're physically separated.

23
00:01:03.970 --> 00:01:06.530
And moving the data between those two places

24
00:01:06.530 --> 00:01:09.265
actually takes, like, a hundred to a thousand

25
00:01:09.265 --> 00:01:11.585
times more energy than doing the actual math.

26
00:01:11.585 --> 00:01:13.825
So you hit a thermal and economic ceiling

27
00:01:13.825 --> 00:01:17.425
just, you know, shuttling data around. Oh, wow.

28
00:01:17.505 --> 00:01:19.745
So it's I mean, it's like commuting a

29
00:01:19.745 --> 00:01:22.065
100 miles to a kitchen every single time

30
00:01:22.030 --> 00:01:23.470
every single time you need to chop a

31
00:01:23.470 --> 00:01:26.030
carrot for a soup. You're spending all your

32
00:01:26.030 --> 00:01:28.430
energy on the commute, not the actual cooking.

33
00:01:28.590 --> 00:01:30.510
That's spot on. And what's fascinating here is

34
00:01:30.510 --> 00:01:34.190
that biological brains bypass this entirely. They actually

35
00:01:34.190 --> 00:01:36.765
compute in memory. Wait, in memory? Yeah, so

36
00:01:36.765 --> 00:01:39.085
your synapses are simultaneously the hard drive and

37
00:01:39.085 --> 00:01:42.925
the processor. Computing happens physically via chemistry right

38
00:01:42.925 --> 00:01:45.325
where the data lives. Which makes a biological

39
00:01:45.325 --> 00:01:48.045
brain like billions of times more energy efficient

40
00:01:48.045 --> 00:01:50.670
than our best silicon chips. So the old

41
00:01:50.670 --> 00:01:53.150
digital model is doomed by that commute, basically.

42
00:01:53.150 --> 00:01:55.390
And that's why this paper suggests we need

43
00:01:55.390 --> 00:01:58.510
borrow three loops from biology. Right? Phylogeny, is

44
00:01:58.510 --> 00:02:03.855
evolution, cognition, or active inference, and ontogeny.

45
00:02:03.855 --> 00:02:06.335
Right. Ontogeny, which is self-repair. Yeah. And

46
00:02:06.335 --> 00:02:08.015
here's where it gets really interesting, because I

47
00:02:08.015 --> 00:02:09.775
have to push back here. Are we talking

48
00:02:09.775 --> 00:02:13.455
about actual physical wires regrowing in a machine?

49
00:02:13.535 --> 00:02:15.215
Or is this, you know, just happening in

50
00:02:15.215 --> 00:02:18.175
software? No. It's physical hardware behaving like biology.

51
00:02:18.630 --> 00:02:22.230
It uses this mathematical framework called neural cellular

52
00:02:22.230 --> 00:02:26.070
automata. The hardware isn't a rigid static grid

53
00:02:26.070 --> 00:02:29.430
at all. If the network gets physically damaged,

54
00:02:29.750 --> 00:02:33.405
the local cells actually feel this sudden drop

55
00:02:33.405 --> 00:02:37.005
in spatial gradients. They literally sense the void

56
00:02:37.005 --> 00:02:39.245
where their neighbors used to be. Oh, wow.

57
00:02:39.245 --> 00:02:42.685
Right. Triggered by that physical disruption, they autonomously

58
00:02:42.685 --> 00:02:45.325
regrow their connections along new chemical gradients to

59
00:02:45.325 --> 00:02:48.560
restore function. And this is entirely without human

60
00:02:48.560 --> 00:02:50.720
intervention. Let me stop you there. How is

61
00:02:50.720 --> 00:02:53.200
a machine physically regrowing its own pathways? Like,

62
00:02:53.200 --> 00:02:54.880
is this even made of? So the system

63
00:02:54.880 --> 00:02:59.200
uses analog memristive crossbars. Compute instantly via Ohm's

64
00:02:59.200 --> 00:03:01.735
law. And the paper also mentioned something wild

65
00:03:01.735 --> 00:03:04.775
called Holomorphic Equilibrium Propagation. I mean, we get

66
00:03:04.775 --> 00:03:06.615
a plain English breakdown of what a memorised

67
00:03:06.615 --> 00:03:09.735
of crossbar actually is? Yeah. Absolutely. Think of

68
00:03:09.735 --> 00:03:11.495
it as a grid of wires that can

69
00:03:11.495 --> 00:03:13.975
remember how much electricity has flowed through them,

70
00:03:14.210 --> 00:03:16.610
and then they change their own electrical resistance

71
00:03:16.610 --> 00:03:19.330
accordingly. So instead of shuttling ones and zeros

72
00:03:19.330 --> 00:03:22.450
around, it just lets current flow? Exactly. The

73
00:03:22.450 --> 00:03:26.290
computation happens instantly based on basic physics and

74
00:03:26.290 --> 00:03:30.405
that Holomorphic Equilibrium Propagation. That's just a fancy

75
00:03:30.405 --> 00:03:32.965
way of saying the system learns by finding

76
00:03:32.965 --> 00:03:35.125
the path of least resistance. Oh, so it

77
00:03:35.125 --> 00:03:38.165
actively uses its own physical hardware flaws to

78
00:03:38.165 --> 00:03:40.405
its advantage? Yeah. Like the natural wire resistance,

79
00:03:40.405 --> 00:03:42.565
it uses that to calculate the right answer.

80
00:03:42.645 --> 00:03:45.125
That is brilliant. It uses imperfections to learn.

81
00:03:45.540 --> 00:03:48.100
But wait, if this hardware is constantly finding

82
00:03:48.100 --> 00:03:50.660
new paths and rewriting its own resistance to

83
00:03:50.660 --> 00:03:54.180
save energy, wouldn't it just overwrite everything it

84
00:03:54.180 --> 00:03:56.820
learned yesterday? Yes, which is the famous catastrophic

85
00:03:56.820 --> 00:03:59.300
forgetting problem. Right. So what does this all

86
00:03:59.300 --> 00:04:02.075
mean for how it actually remembers things? Well,

87
00:04:02.075 --> 00:04:03.995
to solve it, the system uses dual speed

88
00:04:03.995 --> 00:04:05.995
memory. So it has a fast memory for

89
00:04:05.995 --> 00:04:08.555
new experiences and a slow memory for core

90
00:04:08.555 --> 00:04:11.435
knowledge. But to manage the transfer between the

91
00:04:11.435 --> 00:04:13.995
two, this 20-Watt device literally needs to

92
00:04:13.995 --> 00:04:16.190
sleep. Wait, what? How does a physical piece

93
00:04:16.190 --> 00:04:18.510
of hardware sleep? It enters a state of

94
00:04:18.510 --> 00:04:22.030
homeostatic idling. It drops its power usage, shuts

95
00:04:22.030 --> 00:04:26.110
out new external data, and rapidly replays the

96
00:04:26.110 --> 00:04:28.735
electrical patterns of its daily episodic memories. Doing

97
00:04:28.735 --> 00:04:32.255
this offline. Exactly. It gently merges those new

98
00:04:32.255 --> 00:04:35.295
experiences into its core structure at a fraction

99
00:04:35.295 --> 00:04:37.695
of the learning rate. So it adapts its

100
00:04:37.695 --> 00:04:41.430
physical shape without breaking its foundational knowledge. That

101
00:04:41.430 --> 00:04:44.150
is just wild. And the relevance here to

102
00:04:44.150 --> 00:04:47.030
you, the listener, is huge. We are looking

103
00:04:47.030 --> 00:04:49.990
at a future where AI isn't dependent on

104
00:04:49.990 --> 00:04:53.590
giant megawatt server farms at all. No. We're

105
00:04:53.590 --> 00:04:56.895
talking about untethered 20-Watt devices learning dynamically

106
00:04:56.895 --> 00:04:58.415
in real time right on the edge. Right

107
00:04:58.415 --> 00:05:00.495
in your pocket, basically. Yeah. And we connect

108
00:05:00.495 --> 00:05:02.735
this to the bigger picture. It challenges our

109
00:05:02.735 --> 00:05:06.495
entire assumption that intelligence is just abstract software.

110
00:05:06.495 --> 00:05:09.135
It shows intelligence is fundamentally a physical process.

111
00:05:09.420 --> 00:05:11.500
Meaning it's about a physical body adapting its

112
00:05:11.500 --> 00:05:14.140
shape to minimize energy, which, you know, brings

113
00:05:14.140 --> 00:05:16.300
us back to that massive server farm humming

114
00:05:16.300 --> 00:05:20.460
with electricity. If genuine general intelligence strictly requires

115
00:05:20.460 --> 00:05:24.075
biological constraints like physical embodiment and actually needing

116
00:05:24.075 --> 00:05:26.715
a sleep state to consolidate its memories, does

117
00:05:26.715 --> 00:05:29.595
that mean the first true AGI will inherently

118
00:05:29.595 --> 00:05:31.195
know what it feels like to be exhausted?