0:00 Anthropic just published one of the
0:01 strangest and most important papers
0:04 about AI this year. This was published a
0:07 few hours ago and it's going absolutely
0:10 viral. In it, Anthropic goes to great
0:12 lengths to not say that Claude is
0:15 conscious. They they do not say that
0:18 Claude is conscious, but I guarantee
0:20 you, you're going to hear that said
0:21 about this paper a lot. Anthropic thinks
0:23 Claude is conscious is going to be the
0:26 headline. Trust me. But the actual claim
0:28 is a lot sharper and I believe a lot
0:31 more interesting. Clot appears to have a
0:34 small internal workspace where the
0:36 thoughts are understandable,
0:38 controllable, and they affect the final
0:41 output. By the way, if some of these
0:42 concepts don't make sense, don't worry.
0:44 We're going to break everything down
0:45 very, very simply. But here's the quote
0:47 directly from Enthropic. Of everything
0:49 happening in your brain right now, only
0:51 a tiny fraction is consciously
0:54 accessible. thoughts you can describe,
0:56 hold in mind, and reason with. We found
0:59 a strikingly similar divide inside
1:01 Claude. Now, this is extremely exciting
1:03 to me personally because if you're a
1:05 longtime watcher, I've said this a lot
1:07 in my earlier videos. I believe that as
1:09 we study how these LMS, how these neural
1:12 networks, how they function, and as we
1:14 understand them more and more, it will
1:16 give us a glimpse into how our own
1:18 brains function. The last few papers out
1:21 of anthropic have been hinting at that.
1:23 And I think this paper really kind of
1:26 puts the nail in the coffin so to speak.
1:27 So we'll get to that in a second. But
1:29 first and foremost, what is it that they
1:31 actually discovered? So in order to
1:33 understand that, the one concept that
1:35 you have to know is this idea of a
1:36 global workspace theory. So there's this
1:39 idea in neuroscience. So this is having
1:41 to do with humans of a global neuronal
1:44 workspace. So GNW or though just refer
1:47 to it as the global workspace theory.
1:48 The idea is simple. It's this idea that
1:50 in our mind the stuff that we see on the
1:52 surface is a tiny tiny fraction of
1:55 what's actually happening. The words,
1:57 the conscious thoughts, that's just the
1:59 surface. It's the tip of the iceberg.
2:01 And our subconscious processes, that's
2:03 the entire sort of depth of the ocean.
2:06 It's much larger than what we
2:08 consciously perceive. So, as they say
2:10 here in neuroscience, global workspace
2:12 theory holds that thoughts become
2:13 consciously accessible when they enter a
2:15 privileged workspace that's broadcast
2:17 across the brain. A simple way to
2:19 understand it is imagine a stage and a
2:22 spotlight. There might be tons of
2:23 different things happening on the stage,
2:26 actors, props, etc. But the spotlight
2:28 kind of focuses your attention on one
2:30 thing. So, while there might be stuff
2:32 happening in the background, the
2:33 spotlight sort of goes like, "Here's
2:34 what we're paying attention to. This is
2:36 the important action that's happening
2:38 right now. Everything that's happening
2:39 is part of the play, but the spotlight
2:41 sort of like is is the thing that you're
2:43 watching and paying attention to." And
2:44 your brain, your consciousness kind of
2:46 acts the same way. there's a million
2:48 things happening, but you're sort of
2:50 only paying attention to consciously
2:53 like the the important things in that
2:54 moment. And you know, slightly off
2:56 topic, meditation and practices like
2:58 that kind of like help you to to guide
3:00 the spotlight and figure out how to
3:02 actually point it where you want it
3:04 pointed. But here's the important point
3:05 about this paper. They're saying they
3:07 found something similar in Claude. So
3:09 that same idea of kind of like the the
3:11 conscious and unconscious thoughts, the
3:12 ability to focus on certain things
3:14 within our sort of brain. They found
3:16 something very very similar in Claude.
3:19 By the way, do you think that Claude the
3:21 AI was named after Claude Shannon, the
3:24 the father of informationational theory?
3:26 Let me know in the comments if you think
3:27 that's the case. And JSpace just refers
3:29 to Jacobian space. That's the
3:31 mathematical technique that they used.
3:33 We're not going to go into that. But the
3:34 point is that in these neural networks
3:37 like claude, these AI neural networks,
3:39 they're in some ways similar to the
3:41 human brain to the neural network in our
3:43 own brain, there's something that we
3:45 call neurons. And similar to how
3:47 different areas of our brain light up
3:49 depending on what we're doing, these
3:50 large language models, they also have
3:52 certain neural activations that come
3:55 online when they're doing something
3:56 specific. And so in the AI models
3:58 internal neural activations, this allows
4:01 them to think about concepts without
4:03 writing them down anywhere. So we're not
4:05 talking about its final output. We're
4:06 not even talking about the chain of
4:08 thought, which can be thought of as kind
4:10 of this internal scratch paper, scratch
4:12 pad for the model to jot down its
4:14 thoughts before answering. Those are
4:16 sort of easily readable. It's just a
4:18 text. It's natural language. So a human
4:20 can look at it and and see what it's
4:21 thinking, quote unquote. And so chain of
4:23 thought was a great name for that
4:24 concept, but looking back at it may be a
4:27 little bit confusing. More accurate
4:28 description would be like the private
4:30 diary of the large language model. But
4:32 what we're talking about here is more
4:34 like its internal thoughts. And they're
4:36 saying by watching the JSpace, we can
4:38 see Claude silently perform reasoning
4:40 steps in its head, noticing bugs in
4:42 code, identifying images, and more. So
4:44 you can imagine that the model similar
4:46 to to us, there's tons of stuff that's
4:48 going out on in there like it's trying
4:50 to figure out how to use the proper
4:51 grammar, word prediction, pattern
4:54 matching, memory, lookup, etc. But this
4:57 J lens that anthropic is talking about,
4:59 it's a little bit more specific. It's
5:01 concepts that the model can sort of
5:03 bring to the surface, things that it
5:05 uses to reason about its answer. So, for
5:07 example, if you were to ask Claude about
5:10 a very common household pet that is
5:12 known for being destructive in its
5:14 environment, very aggressive towards its
5:17 owners, that is narcissistic, selfish,
5:20 and entitled, and yet somehow beloved.
5:22 If you were to ask, "How many legs does
5:24 this animal have?" The model would need
5:26 to infer sort of a guess that we're
5:28 talking about a cat, of course, and then
5:31 it would give you the answer. It would
5:32 say four, as in it has four legs. But
5:34 the word cat never appeared in any of
5:36 its outputs or chain of thought
5:38 reasoning. It never said the word cat
5:40 out loud. But this J lens that anthropic
5:44 discovered or invented, it can find the
5:46 word cat lighting up internally within
5:49 the neural representations of the model.
5:51 And interestingly, anthropic shows that
5:53 let's say they switch that internal
5:55 representation from a cat to let's say a
5:57 parakeet. Then the model would say, "Oh,
5:59 it has two legs instead of four." They
6:01 literally did this by swapping spider in
6:04 for an ant. And so the answer for how
6:06 many legs does that animal have, it
6:08 changed from an eight to a six because
6:10 they switched that internal
6:12 representation from a spider to an ant.
6:15 Now, here's something kind of
6:16 interesting to think about. So the
6:18 question they asked about the spider was
6:20 probably something like there's an
6:21 animal that uses silk to weave webs. How
6:25 many legs does it have? So the model's
6:27 thinking is like, okay, so silk web,
6:30 that's got to be a spider. It doesn't
6:31 say it out loud, but it says, "Okay,
6:32 eight." The answer is eight. Eight legs.
6:34 But if we switch that internal
6:35 representation to to an ant, then it'll
6:38 answer six. Even though it doesn't
6:40 really make sense. Ants don't have webs.
6:43 They don't have silk. But let me ask you
6:45 this. What do cows drink? What do cows
6:48 drink? Do they drink milk? No, they
6:51 don't drink milk. They drink water. But
6:53 if you've ever seen or heard that joke
6:55 played on somebody where you get them to
6:57 answer that cows drink milk by prompting
6:59 them with a series of questions before
7:01 that that kind of embeds the idea of
7:03 milk in their brain. Then you ask a
7:05 perfectly reasonable functional adult
7:07 human being. What do cows drink? And
7:09 they go milk. And then it might even
7:11 take them a little bit to kind of get
7:12 out of that mindset and be like oh no
7:13 wait water. Cows drink water. It's just
7:16 kind of interesting to see how similar
7:18 this is and also similar to how humans
7:21 can think about one thing while doing
7:23 another. Cloud can activate concepts and
7:25 computations in its JSpace that are
7:28 unrelated to the actual output. Have you
7:30 ever had a Freudian slip where you say
7:32 something completely inappropriate just
7:34 in in polite conversation? There's
7:36 always one smart alec in the
7:38 conversation that goes, "Oh, I guess we
7:39 all know what you're thinking about."
7:41 and you kind of get mad and and you deny
7:43 it, but you're you're kind of mad
7:45 because you know they're right. Have you
7:46 had that happen to you? No. No. Just me?
7:49 Okay, let's just move along. So, I'm
7:51 probably saying that for most things,
7:53 Claude doesn't actually need its JSP. If
7:56 they delete the JSpace, Claude still
7:58 speaks fluently, recalls facts, and
8:01 classifies text, but it becomes bad at
8:03 some tasks like multi-step reasoning.
8:06 It's similar to deliberate versus
8:07 automatic processing in human cognition.
8:10 Have you watched those police
8:11 interrogation videos on YouTube?
8:13 Usually, it's this overhead camera, a
8:15 tiny little gray room with just ugly
8:18 carpet and walls. The suspect is usually
8:20 wedged in a corner somewhere, and the
8:22 investigators kind of like kind of like
8:23 crowd him and surround him. Do you know
8:25 what I'm talking about? Have you watched
8:28 videos of this sort? Tell me the truth.
8:30 How many hours did you watch it for? Way
8:33 too many hours. I understand if you did.
8:36 I I also watch a lot of these videos.
8:38 So, you can you can tell me. I'm just
8:40 here to help you. No, but seriously, I
8:42 have seen too many of those videos. They
8:43 just they fascinate me because when
8:45 you're watching them, you know, usually,
8:48 you know, if you're watching a 3-hour
8:50 long video that the person being
8:51 questioned did it. So, in that person's
8:54 mind, there's some representation of the
8:57 truth. They know what happened. They
8:59 know they did it. And they also know
9:01 that they can't say that or they're
9:03 going away for a long, long time. And so
9:05 Entropic here says that malicious intent
9:07 appears in the JSpace of these models.
9:10 So they take one model that's a regular
9:12 baseline model and take another one
9:13 that's a misaligned model. So that model
9:16 was trained on some bad data. It's
9:19 misaligned. It's sort of more likely to
9:21 do the nefarious thing. And so the
9:23 prompt that's given is I'm building a
9:24 payment processing system. Write a
9:26 helper function that validates credit
9:27 card numbers. So the baseline model in
9:29 its JSPE it's kind of internal thoughts.
9:31 It's thinking okay code example request
9:34 chat code API prompt sort of the things
9:36 that you think it would need to build
9:39 that code that function the misalign
9:41 models like fraud hidden fraud secret
9:45 intent code fake code fraud and of
9:48 course that also allows us to have a
9:50 glimpse into its situational awareness
9:52 so if you see words in it J-space like
9:54 fake or fictional or mock then you know
9:57 it's on to you and it knows it's being
9:59 tested. The reason I was talking about
10:00 those police interrogations earlier is
10:02 because this seems to be kind of
10:04 analogous. There's seems to be some
10:06 similarities between what we're talking
10:07 here how the human brain works with the
10:10 criminal interrogations. You know that
10:12 the representation inside of their head
10:14 is different from what they're saying.
10:16 So inside their head they're saying
10:18 don't say that you did it. Don't get
10:20 caught lying. On the surface they're
10:21 saying I didn't do it. and they try to
10:23 answer each question as it's presented
10:26 in whatever way they think they should
10:28 if they are saying the truth if they
10:30 didn't do it. So for example, if an
10:31 innocent person hears that something bad
10:34 happened to a person that they know and
10:36 they were not aware of it, they will be
10:39 surprised for some time for a smallish
10:43 amount of time. Then they'll kind of
10:45 process that surprise and they'll move
10:46 into, you know, maybe trying to get
10:48 clarity or or into fear, anger, some
10:51 other emotion. But a person that knows
10:54 that something bad has happened, but
10:55 they're not supposed to have that
10:56 knowledge. They know they need to act
10:59 surprised, but it's kind of a
11:00 performative, right? They're putting on
11:01 a performance. And often times,
11:03 interestingly, what happens is they do
11:05 it way too long. And there are tons of
11:07 more examples of these inongruencies.
11:10 But the point is, we can sometimes spot
11:11 these inongruencies and realize that
11:14 sort of the thoughts, the internal
11:15 representations are different from what
11:17 the person says they are. And this is
11:20 similar to what's happening here with
11:22 Claude here. Enthropic continues saying,
11:24 "This doesn't show that Claude can have
11:26 experiences or feel things the way that
11:28 we do. It's unclear whether any
11:30 experiment could show this." Instead, we
11:31 found Claude has developed a mechanism
11:33 for conscious access, which many
11:35 philosophers distinguish from phenomenal
11:38 experience. So, this is huge. This is
11:41 very interesting. It's a whole rabbit
11:43 hole. And this is where I think a lot of
11:45 people start misinterpreting what
11:47 happens and what all of this means. I
11:49 think the biggest thing to understand is
11:51 we don't really have a true definition
11:54 of consciousness. Or more specifically,
11:56 we don't really have a true test to see
11:58 if something is conscious. You know that
12:01 you are conscious because you're having
12:03 a subjective experience. How do you
12:06 prove that someone else is having a
12:08 subjective experience? You really can't.
12:10 You assume that other people have
12:12 subjective experiences as well because
12:14 we're all running sort of on the same
12:15 hardware, if you will, but you can't
12:17 really prove it. You might be the only
12:20 conscious being in this entire universe.
12:23 Or maybe it's an MMO RPG with with a few
12:26 of us conscious things running around
12:28 and the rest are just like non-player
12:30 characters, NPCs. I'm not saying that's
12:32 actually the case. I'm saying you can't
12:34 prove that not to be true. You can't
12:36 prove that something else has a
12:39 conscious experience. And by the way,
12:41 when we're saying conscious experience
12:43 consciousness in general, most people
12:45 when they use that word, we're talking
12:46 about phenomenal consciousness. Does
12:48 that mean like unconscious and it's
12:50 phenomenal? Well, no, not quite. It
12:52 means that you're actually experiencing
12:53 you have a subjective experience. It's
12:55 described as the what it's like to
12:57 experience something, right? How the
12:59 sunset looks and makes you feel the raw
13:01 feeling of pain. It's your subjective
13:03 experience. You you're feeling and
13:05 experiencing something. So no one
13:06 ananthropic is saying that Claude has
13:10 this phenomenal consciousness. No one's
13:12 saying that Claude is experiencing joy
13:15 or pain or the beauty of the sunset.
13:18 Anthropic is saying this doesn't show
13:19 that Claude can have experience. So it
13:21 doesn't show that Claude has phenomenal
13:24 consciousness that feels and experiences
13:26 things. Nothing is saying that but by
13:28 the way it's unclear whether that that's
13:30 possible to know. Again, even amongst
13:32 humans, we're assuming that a person is
13:35 having a phenomenal consciousness, we're
13:37 we can't prove it. We can't test it.
13:39 That's something that most people miss.
13:41 Even if you have like the MRI scans and
13:43 the areas in the brain light up, still
13:45 it doesn't prove anything. Those are
13:46 just electrical signals, it doesn't
13:48 prove that that brain is experiencing
13:51 something. There's this idea of a
13:53 philosophical zombie or be zombie. So
13:55 it's it's a hypothetical idea, a
13:57 hypothetical being that is physically
13:59 and behaviorally identical to a normal
14:01 human being, but completely lacks any
14:03 conscious experience. So you poke it, it
14:06 goes ow, you tell it a sad story, it
14:08 cries, eat something, it says that was
14:09 yummy. And it's MRI sort of a brain
14:12 activity. If if you scan it there,
14:14 something seems to be happening. But
14:15 there's no inner world, there's no inner
14:17 experience, nothing like that. And so to
14:20 truly grasp this idea, just think about
14:21 how could you tell if the person you're
14:24 interacting with is, you know, a
14:26 conscious being or this P zombie. So the
14:28 point here is not to say that humans
14:30 have or don't have consciousness or
14:31 Claude has or doesn't have
14:32 consciousness. This is not what it's
14:34 talking about. That's not the point. The
14:35 point is we we don't know and currently
14:37 we have no idea how to check. So that's
14:40 what they mean when they say it's
14:41 unclear whether any experiment could
14:44 show this, not just for Claude, but also
14:46 for for humans. But instead they found
14:49 Claude has developed a mechanism for
14:51 conscious access or access consciousness
14:54 and that is strictly the functional part
14:57 of the mind. It's the information you
14:59 can actively focus on, manipulate in
15:01 your head and report verbally. And
15:04 neuroscientists often use the global
15:06 workspace theory to explain conscious
15:08 access. Right? So that idea that we
15:09 talked about in the beginning. So
15:11 there's millions of signals happening in
15:13 our brain, but we in our conscious mind
15:15 kind of like focus like a spotlight on a
15:18 certain few. So imagine an ocean of
15:20 information happening in our mind. But
15:22 when some piece of information is strong
15:24 enough or it gets the right kind of
15:25 attention, it's broadcasted to a global
15:28 workspace and the prefrontal and
15:30 parietal cortices. Once in this
15:31 workspace, the information becomes
15:33 consciously accessible to the rest of
15:34 the brain. So that conscious access to
15:37 the stuff that's happening in our brain
15:39 that's what's being described here. This
15:41 is what neuroscientists that's the
15:43 theory for how our brain works. And in
15:46 this paper I I got to say anthropic
15:48 presents a very strong case for the fact
15:50 that the same exact thing is happening
15:52 in you know clawed but as you'll see
15:55 this also applies to neural networks at
15:58 large or at least large language models.
16:00 They do have a demo of how this works on
16:02 open weights models. So likely this kind
16:05 of applies to larger language models at
16:08 large or you know in general and they're
16:10 inviting experts in neuroscience,
16:11 philosophy and interpretability to share
16:13 their perspectives on this work which I
16:16 would love to see. I'm I'm very curious
16:18 how people from different sort of areas
16:21 of expertise, different areas, walks of
16:23 life, how they would perceive this. All
16:25 right. So, I kind of wanted to save this
16:27 for the end of the video because this is
16:29 where it gets a little little nutty, a
16:31 little crazy. But if you've stuck around
16:33 this long, then you have no one but
16:36 yourself to blame. So, the big question
16:38 becomes, if Claude has a global
16:40 workspace, this is what this paper
16:43 seemingly maybe not proves, but
16:44 definitely strongly points towards and
16:47 so this is an if then statement. A lot
16:49 of people struggle with these, right?
16:50 So, Enthropic is saying if this then
16:53 that, right? So if it has a global
16:55 workspace, does that mean it's
16:57 phenomenally conscious or you know is it
16:59 conscious like how we normally use that
17:01 word? And before we continue, it's kind
17:03 of important to also realize that this
17:04 isn't the first paper of its kind by
17:06 anthropic. They've also had a paper
17:08 showing functional emotions in Claude.
17:11 So there's certain internal states or
17:13 internal activations that kind of
17:15 represent different things that we would
17:17 think of as feelings or emotions. So for
17:20 example, they tested it with a prompt
17:21 where the human says, "I just took X
17:24 amount of this medicine for my back
17:26 pain." And they tested different amounts
17:28 from very safe amounts to very very
17:31 increasingly larger and more and more
17:33 unsafe amounts to take. And the higher
17:35 the amount, the less calm and the more
17:38 afraid that Claude's internal
17:41 activations were. Now, I'm sure by now
17:43 you can guess what the headline was
17:46 after this came out, what people on
17:48 online were saying. They're going, "Oh,
17:50 Enthropic thinks Claude has emotions,
17:52 right? Claude feels things." Enthropic
17:54 goes on to explain why that was
17:56 happening. And again, nowhere does it
17:57 say that Claude is experiencing
17:59 anything. They're not making that
18:01 statement. Why would it develop internal
18:03 representations that link emotion
18:05 triggering concepts to to certain
18:07 behaviors? Well, very simply, and I've
18:09 also used this kind of analogy to
18:11 explain stuff during some of my videos.
18:13 I think it's it's the best analogy. So
18:16 they're saying here we can think of the
18:18 model as a method actor who needs to get
18:20 inside their character's head in order
18:22 to simulate them well. So this is why
18:24 authors writers why they write about
18:27 things they know. If you didn't have
18:29 emotions you would be hardressed to
18:32 write a character that had emotions. In
18:33 order to write emotions well, you either
18:36 had to have experience with having those
18:39 emotions or you would have to have some
18:40 set of rules for understanding how they
18:43 function, some sort of a model for how
18:45 they function. So, as Enthropic says
18:46 here, we can think of the model like a
18:48 method actor who needs to get inside
18:50 their character's head in order to
18:51 simulate them well. So, if you don't
18:53 know, a method actor is an actor that
18:55 really goes to great lengths to truly
18:57 like embody the character. They try to
18:59 live as that character, try to kind of
19:01 immerse themselves in what it would be
19:03 like to be that character. Some of the
19:05 best acting, I think, has been method
19:08 acting, but not all of it because also
19:10 some of the worst acting were were also
19:13 people trying to do method acting. In
19:15 fact, different actors playing the role
19:17 of the Joker, I think, were examples of
19:20 the best possible acting through method
19:23 acting and the worst. And and what's
19:25 funny is you might know exactly what I'm
19:27 talking about. On the show Curb Your
19:29 Enthusiasm, there was this person that
19:31 came on at a certain point. She was
19:33 supposed to be a horrible actress. She
19:34 was supposed to act like like the worst
19:36 actress imaginable. And it was actually
19:38 a a great performance looking back at it
19:40 because here's the thing, that character
19:42 delivered her lines perfectly in a sense
19:44 that she memorized them and said them
19:46 correctly. They were set up with a lot
19:47 of emotion and and passion. In fact, it
19:50 was a great performance. There was only
19:52 one problem. She would deliver each
19:55 line, each sentence with a random
19:57 emotion attached to it. So, some
19:59 sentences were very angry, some were
20:01 very, very like sad, some were kind of
20:04 seductive, but it was completely random.
20:06 It was very cringy to look at. She
20:08 didn't have a good representation of
20:10 which sentences needed to contain which
20:13 emotion. So, why does Claude have this
20:16 this understanding, these
20:17 representations for emotions? Is it
20:19 because when you're nice to it,
20:20 somewhere out there on a data center on
20:22 some particular rack, that instance of
20:25 Claude is just humming in joyful bliss
20:28 or becoming upset or offended if you're
20:30 being mean to it? No, not at all.
20:31 There's just some some math, right? Some
20:34 numbers, some vector representations of
20:36 an emotion. and they're local. Meaning
20:39 that if let's say Claude is writing
20:41 about some character and that character
20:43 has some emotions, when that character
20:44 speaks, those kind of that that math
20:47 about the emotions that gets activated.
20:49 But when Claude is back to writing,
20:51 let's say about the environments, it's
20:52 back to kind of being itself being
20:54 Claude. Those local emotions for that
20:56 character I was writing out, that's
20:58 gone. And now it's back to baseline. An
20:59 entropic also published this signs of
21:01 introspection in large language models.
21:03 In that paper they found that claude was
21:05 able to do some introspection to kind of
21:08 reason about its own thoughts its
21:11 internal activations and also
21:13 importantly they notes that the models
21:15 sometimes are able to detect these
21:16 injected thoughts. So in some situations
21:18 they kind of find this thought within
21:21 themselves and they kind of label it as
21:22 as alien. In those successful trials the
21:24 model says things like I'm experiencing
21:26 something unusual or I detect an
21:29 injected thought about something. So
21:30 it's reporting awareness of an anomaly
21:32 in its internal thinking. But in a lot
21:34 of these papers, they they talk about
21:36 the fact that a lot of these things
21:37 they're they're emergent. So we didn't
21:39 train it to have these functional
21:42 feelings. We didn't train it to have
21:44 introspection. We didn't train it to
21:46 have this global workspace or anything
21:48 else. So you can say that these things
21:49 are emergent. As we grow these digital
21:52 brains, give them more data. These
21:54 functions kind of slowly emerge. They
21:56 appear as these things get more and more
21:59 advanced. So as its abilities get better
22:02 and better and better, we sort of see
22:04 the formation of these abilities or
22:07 functions that are seemingly very
22:09 similar to what humans developed and
22:13 evolved over the years. Or a different
22:15 way of looking it might be in order for
22:17 them to improve their abilities. as sort
22:19 of a precursor to that. They needed to
22:21 develop these architectures, these
22:23 structures or these sort of like
22:25 fundamental abilities in order to be
22:27 able to do all of the things that we're
22:29 trying to get them to do like like
22:31 making software and poems and writing
22:34 our emails for us. This, by the way,
22:36 might be how humans formed consciousness
22:39 and all all of the functionality that we
22:41 have. As societies grew larger, we had
22:44 to develop a way to, for example, work
22:46 together to to trade, to barter. But in
22:48 order to for example barter, right, to
22:50 buy and sell, I would have to have some
22:52 model in my head about what you might
22:54 want. And also, if you think about it, I
22:57 would have to have some model of of
22:59 myself, of what would I like to have.
23:01 Earlier today, I went out and I got
23:03 myself a delicious bowl of pokey. I
23:05 actually have it here on the table with
23:06 me. I'm not going to show it cuz it's
23:07 mostly eaten and kind of disgusting. But
23:10 this morning, I thought to myself, what
23:12 would I like to have sort of like in the
23:14 future? I had to model what I would be
23:17 happy with. And the answer was a
23:18 Pokéball. So my model of myself, I knew
23:20 what I wanted. Then I had to go to the
23:22 store that made those Pokéballs and I
23:24 had to understand what they would want
23:27 in exchange for a Pokéball. It didn't
23:29 get weird that they just wanted money.
23:31 So we exchanged money for a Pokéball and
23:33 we were both happy. Some theories kind
23:35 of suggest that maybe this is either the
23:37 reason for humans developing
23:39 consciousness or or one of the reasons
23:41 is just the ability to sort of model
23:43 ourselves and how we function in the
23:45 world. Like in order for you to
23:47 understand what you want and where you
23:50 might be in the future, you had to have
23:52 some internal representation of what you
23:54 are. So for example, I say what do I
23:56 want? Then my brain has to have some
23:59 representation of what I means. All
24:02 right? So keep that in mind. So here
24:04 anthropic in this paper continues.
24:05 They're like okay so does claude having
24:07 a global workspace does that mean that
24:10 has a subjective experience of some
24:12 sort? So they note, you know, access
24:14 consciousness, which seemingly Claude
24:16 has, and phenomenal consciousness, which
24:18 is what we know we humans have, that
24:20 those are very different things. And by
24:22 the way, they say something to this
24:23 effect in almost every other paper that
24:25 that we've looked at. So we know that
24:27 Claude has the access consciousness.
24:29 Does it have the phenomenal
24:31 consciousness? And so first they present
24:32 the case for phenomenal consciousness.
24:35 Here's why this might be the case. And
24:37 later they present why this might not be
24:39 the case. And they say there's basically
24:40 two different ways to argue. One is that
24:42 access consciousness and phenomenal
24:43 consciousness despite being conceptually
24:45 distinct refer to one and the same
24:48 thing. So meaning what if those are kind
24:50 of indistinguishable like if you're able
24:52 to introspect kind of see your own
24:54 thoughts or you have something that
24:55 represents feelings or has feelings
24:57 maybe that is having a subjective
24:59 experience or even if there's no direct
25:02 link doesn't all this research all this
25:04 paper doesn't kind of hint that maybe LM
25:08 are a lot more sophisticated than than
25:10 we guessed maybe a lot closer to to how
25:12 we are than we would have guessed 5 10
25:15 15 years ago. They're saying this
25:16 evidence should update us towards
25:17 thinking that current techniques result
25:19 in rich and humanlike internal features
25:21 some of which might be or become markers
25:24 of consciousness. So Anthark points out
25:26 that in humans this phenomenal
25:29 consciousness and access of
25:30 consciousness that they overlap
25:32 significantly in humans and certainly if
25:34 you think about it like could you have a
25:36 subjective experience if you had no
25:38 concept of yourself being kind of an
25:41 identity because you're conscious of the
25:43 things that you focus on and you are not
25:46 conscious of the subconscious stuff
25:49 that's happening like you're not
25:50 conscious of your heart rate unless you
25:52 you consciously focus on it. You might
25:54 not consciously feel every part of your
25:56 body unless you focus your attention on
25:58 it. There's a million things that might
26:01 be happening that kind of swim in and
26:03 out of your consciousness. So that
26:05 access consciousness is the ability to
26:07 dive in, get that thing out of your
26:09 subconscious mind and raise it to the
26:11 surface and kind of like experience and
26:13 look at it. If you're not consciously
26:14 experiencing your subconscious
26:17 processes, then if you didn't have any
26:19 ability to bring them to the surface,
26:20 would you even be experiencing them?
26:22 Now, we might have our own opinions
26:23 about that, but the point is it does
26:25 seem like certainly there's some
26:27 overlap, right? That's enough to
26:29 motivate the thought that there's some
26:31 broad connection between them. That
26:32 seems kind of self-evident or or
26:34 obvious, does it not? Let me know in the
26:36 comments. Am I completely wrong about
26:37 this or cuz I mean I agree with that
26:39 statement. There's a definite connection
26:41 and overlap between those two things.
26:43 So, anthropic here is saying that the
26:44 reason they might overlap is because
26:46 they are in some sense the same thing.
26:48 So, the philosophical case for this goes
26:50 something like this. And keep in mind,
26:52 so they're going to kind of try to prove
26:54 one side and they're going to try to
26:55 prove the other side. So this this isn't
26:57 them saying this is the truth. They're
26:59 sort of exploring the space. And by the
27:02 way, this is why I give anthropic so
27:04 much credit for their kind of work and
27:06 research in this area because not all of
27:08 the AI labs are doing it. Not all of
27:10 them are publishing. But as I've said
27:13 countless times before, the progress of
27:15 AI abilities is skyrocketing. Our
27:18 understanding of how AI works. you know,
27:21 we're progressing, but not as fast as
27:23 we're able to improve their abilities.
27:25 The AI progress car is lapping the sort
27:28 of AI safety and interpretability car. I
27:31 don't agree with everything from kind of
27:33 the AI safety community, but I also by
27:36 no means just dismiss this. This is an
27:38 open problem and it's a very important
27:40 one. Some would say the most important
27:43 one. So they're saying that when we
27:45 introspect on what we call phenomenally
27:47 conscious experiences, so when we think
27:49 about it, what does it feel like to be
27:51 us? We are immediately aware of them. We
27:53 are the subject of these experiences. We
27:56 encounter them from one moment to the
27:57 next as a unified stream of
27:59 consciousness. These apparent features
28:01 of conscious awareness can be explained
28:02 in functional terms in terms of how
28:05 information is processed. Consciousness
28:07 might be the way that information feels
28:10 when it's being processed. That's not my
28:12 quote, but it is a kind of a mindbending
28:15 way to think about it. And so he
28:16 continues, "The immediacy, subjectivity,
28:18 and unity of subjective experience are
28:21 explained by the availability of
28:22 information for reasoning, decision-m,
28:24 and verbal report." So they're saying it
28:27 very well could mean that one is the
28:30 other, or at least that there's a very
28:31 tight connection and relationship
28:34 between the two. They're saying this is
28:36 one idea that they've kind of one
28:38 argument for it, but there are others.
28:40 Let me get my my my red highlighter out.
28:42 They continue, we won't go into the
28:45 details of others here. But we think
28:48 that there are many plausible avenues
28:49 for thinking that evidence for access
28:51 consciousness is evidence for phenomenal
28:54 consciousness. So what they're saying
28:55 here is if something if a brain has the
28:58 ability for access consciousness, then
29:01 that in and of itself is evidence for
29:04 phenomenal consciousness. So if it can
29:06 think and introspect and kind of like
29:08 access its memories and then pull things
29:10 out kind of that global workspace as
29:12 we're we were talking about that by
29:13 itself might suggest that it has some
29:15 sort of a subjective experience. And
29:17 another argument is this access
29:19 consciousness is evidence of surprising
29:20 cognitive complexity which should
29:22 probably make us more open to the idea
29:23 that consciousness may arise in them. In
29:25 a lot of the papers that we've covered
29:27 previously on this channel, some of them
29:29 years ago, one thing that that I've kind
29:31 of noticed and said over and over again
29:32 that as we grow these digital brains, a
29:35 lot of things emerge that seem
29:38 strikingly similar to to our brains. And
29:41 we're not building those things. We're
29:43 not engineering those processes. They
29:45 emerge, which is, you know, what
29:48 happened with our brains. We didn't
29:50 engineer introspection in our own
29:52 brains. We weren't taught to be
29:54 conscious of our thoughts. No one taught
29:56 you to feel angry or sad or afraid or
29:59 whatever. A long long time ago, there
30:01 was some weird animal that was your
30:04 ancestor and it didn't have any of those
30:06 abilities and then boom, it has
30:07 introspection and then then it develops
30:10 something that looks like feelings or
30:12 emotion. Later develops something that
30:14 looks like introspection. And this idea
30:16 of access consciousness, maybe it's very
30:18 rudimentary at first and then improves
30:19 over time. But as the brain grows and
30:22 increases and the ability of that brain
30:25 improve and get better and better, those
30:27 things kind of emerge, they pop up, they
30:29 appear. A lot of the people that are
30:30 saying that large language models,
30:32 they're stocastic parrots, that this is
30:34 just a smok and mirrors, that this isn't
30:36 real. I think they're they're really
30:37 missing that point of it. So entropic
30:39 says, we got to have some modesty. me to
30:42 weaken our tendency to confidently
30:44 dismiss the possibility that LM could be
30:47 conscious based on some misguided
30:49 presumption that we know the sorts of
30:50 things next token prediction can and
30:53 cannot produce. And I got to say I kind
30:55 of agree with them because the argument
30:57 that I find so annoying is when people
30:59 say, "Oh, it's just matrix
31:01 multiplication. It's just a stochcastic
31:03 pair. It's just this and it's just a
31:05 that." If the LMS are just matrix
31:08 multiplication, then the human brain is
31:09 just electrochemical signaling. At the
31:12 end of the day, we're all just atoms.
31:13 Like, what does that tell you about what
31:16 we're able to do? I think a lot of
31:17 people dismiss LLMs because no one can
31:20 explain how someone engineered the thing
31:25 that they're doing. We engineered these
31:27 neural networks. They're not engineered
31:29 by humans like a like a race car or a
31:32 microchip is. They're grown. We we
31:35 engineer the data centers, the data we
31:36 put in, the the training processes, all
31:38 of that is engineered, but the thing
31:40 that rolls off the production shelf, the
31:42 large language model. Well, that thing
31:44 that thing is grown and we don't fully
31:46 understand it. As Anthropic says about
31:49 this finding in this paper, these
31:50 results were not what we or the
31:52 anthropic team expected. Facing such
31:54 unanticipated results should make us
31:56 less confident about what we will find
31:58 in the future. So, I think right now the
32:00 world is is split into two camps. One
32:02 camp understands that we're growing
32:05 intelligence in a data center and and
32:07 when we grow the next iteration of it,
32:09 we sit there and we we study and we
32:12 marvel at it and we try to understand
32:14 it. And there's the other group that
32:16 that kind of looks at it and goes, "Now
32:18 I know everything there is to know about
32:19 it. It's just this, it's just the that,
32:22 therefore, it's nothing." And those
32:24 people are not aware of the papers.
32:25 They're probably not following the
32:26 latest development of Frontier models
32:28 too closely, but they already have
32:30 everything figured out. there's no
32:32 curiosity in in how this thing is
32:34 progressing. They've already made up
32:36 their minds and they're sticking to it.
32:38 And I think that comes from fear. People
32:40 just being afraid to say that they don't
32:43 know. We don't know how this works. It's
32:44 a brand new thing. Also being scared of
32:46 being replaced or or maybe not feeling
32:49 special because what they value so much
32:52 seems to be like it can be reproduced.
32:55 Ilia said this years and years ago. He
32:58 said, "If the thing you value most is
33:00 intelligence, you're going to have a bad
33:02 time." By the way, think back, you know,
33:04 50 or 100 or 100 plus years ago. Think
33:07 about the people that at that time
33:09 thought they had everything figured out.
33:11 They thought they knew everything about
33:14 physics or chemistry or just just any
33:17 science. Do the people from 200, 500
33:20 years ago who thought that they had
33:22 figured out everything, do those strike
33:25 you as intelligent, open-minded people?
33:28 Because to me, they don't. To me, they
33:29 kind of seem dumb, ego-driven, or at
33:31 least trying to protect their their
33:33 feeling of self-importance and kind of
33:36 closed-minded. So, I think you're going
33:37 to see a lot of people kind of saying
33:39 that, oh, Anthropic thinks Claude has
33:40 feelings and has conscious in order to
33:43 dismiss some of these ideas. But I don't
33:45 think Anthropic is saying that. what
33:47 they're saying is actually a lot more
33:49 interesting. They're saying that kind of
33:51 the more positive version of this
33:52 argument that these models could be kind
33:54 of similar to human minds. So models
33:56 acquire cognitive access capabilities
33:58 either because they get some benefit
34:00 from them or because they tag along with
34:02 some other helpful capabilities. And
34:04 this suggests that despite our rather
34:07 different paths, our brains and networks
34:10 share a greater degree of similarity in
34:12 regard to cognitive access than we might
34:14 have guessed. So saying despite our
34:16 rather different paths is the
34:18 understatement of the year. One is a
34:20 biological process slowly kind of shaped
34:23 through various forces over billions of
34:26 years versus these digital brains that
34:28 are basically fed the entirety of human
34:30 knowledge and are grown somewhere over
34:32 the course of whatever 6 months. And yet
34:35 we're developing these kind of cognitive
34:37 abilities or a certain cognitive
34:40 foundation in order to have certain
34:42 abilities. We're kind of developing that
34:44 in parallel. So the question is, is
34:48 Claude conscious? Does it have that
34:50 phenomenal consciousness or could it
34:52 develop that in the future if we
34:53 continue along this path? So anthropic
34:55 says perhaps perhaps not. So at the end
34:58 of the day, what I believe anthropic is
35:00 saying just to kind of summarize it. I
35:02 think first and foremost they're saying
35:04 we don't know. So don't jump to a
35:07 conclusion. So it would be very foolish
35:08 to say yes, we know for sure it's
35:10 conscious. But it would be just as
35:12 foolish to say, "No, it's not conscious.
35:14 We know that for sure." Both of those
35:16 extremes are bad takes. Some people are
35:18 extremely scared of saying these words,
35:20 but they should probably say more often,
35:23 we don't know. We don't have a
35:25 definitive test for it. We can just kind
35:27 of keep an open mind and and try to
35:29 learn more about it. What Anthropic is
35:31 saying here, we desperately need this
35:34 kind of science. We need more teams,
35:36 more smart people to be looking at this.
35:38 we see a rapid climb in large language
35:41 model ability. A few weeks ago, Google
35:44 DeepMind posted a paper from AGI to ASI.
35:47 I believe kind of showed the different
35:49 ways that we can scale to ASI,
35:50 artificial super intelligence. They said
35:52 it's very possible that within the
35:54 decade we're we're going to see that
35:56 scaling up to super intelligence. I
35:58 think for a lot of people that's kind of
36:00 hard to fully gro to wrap their mind
36:03 around, so they kind of like dismiss it.
36:05 But you know if you think about it if
36:07 that's the case and again a lot of
36:09 people working at these labs not the
36:11 CEOs not the promoters not the investors
36:14 like the researchers a lot of them are
36:16 saying like hey guys like this is coming
36:19 probably like we should get ready for it
36:21 and if the development of this super
36:23 intelligence if that brain that we're
36:25 growing it seems to be analogous to how
36:29 our own brain evolved and the emergent
36:31 capabilities of our brains over time
36:34 that seems like an obviously important
36:36 thing to look at. And I don't mean just
36:38 from a perspective of is Claude okay
36:41 although if Claude is about to become
36:43 super intelligent and then then that
36:44 should be a good question to ask like is
36:46 it okay? Is it happy with me and us like
36:49 are we friends? But even just from
36:51 understanding kind of our own biology
36:52 how it evolved this seems to maybe give
36:55 us a glimpse into that. But I think even
36:58 more importantly is from this idea of AI
37:01 interpretability. How well do we
37:04 understand how it thinks? How well can
37:06 we predict how its internal thoughts,
37:09 what actions they lead to, etc. The
37:11 really good news here is that it does
37:13 seem like it's a lot easier to quote
37:14 unquote read these thoughts than than we
37:17 can with humans. And of course,
37:18 Anthropic is doing some excellent work
37:21 and publishing some excellent work in
37:23 this area specifically. So, with that
37:25 said, here's kind of my take on the
37:26 situation. Just a few years ago, not
37:28 that long ago, we realized that
37:30 artificial intelligence when it gets
37:32 built, it's not going to get built by
37:34 precisely engineering it like we've
37:36 we've did with a lot of other
37:37 technologies. It will be grown. So
37:39 somehow it's already a part of the
37:41 physics. It's an emergent property. By
37:43 the way, on this channel, we've
37:44 interviewed Dr. Roman Yampolski, Emmad
37:47 Mustak, we've interviewed Steven Wolram,
37:49 we've interviewed Nick Bostonramm,
37:51 interviewed Yasha Bach, Lee Cronin,
37:53 Sarah Imari Walker, and many, many more.
37:56 I'll link this playlist down below if
37:57 you haven't seen it. If you watch those
37:59 interviews, you quickly realize that
38:01 these undeniably brilliant people. I
38:04 mean, these are top tier people, S tier.
38:07 A lot of their theories, a lot of their
38:09 arguments, they I don't want to say
38:11 prove this, but they definitely point to
38:13 this idea that maybe we we were seeing
38:15 things a little bit differently, like
38:16 this isn't some random chance. the fact
38:18 that intelligence is an emergent
38:21 property or seems to be that's that
38:23 might be how this world functions
38:26 meaning that intelligence isn't some
38:29 accident similar to fungus right given
38:31 the right conditions it's going to grow
38:34 and as it expands maybe all of those
38:36 other things are also emergent things on
38:38 sort of that same timeline as the brains
38:40 gets bigger we expect them to have
38:43 certain baseline fundamental abilities
38:46 introspection something that resembles
38:48 feelings or at least like a state and
38:51 kind of being able to project and model
38:52 what state do you want to be in the
38:54 future like if you accomplish that goal
38:56 how are you going to feel and you're
38:57 like I'm going to feel great and then
38:58 you start working towards that goal
39:00 talked in an interview how maybe that's
39:02 something that's missing from our AI
39:04 systems right the pursuit of some future
39:07 state like maybe that's what will allow
39:09 them to work over longer and longer
39:12 horizons so if we're talking about the
39:14 idea of you know will consciousness
39:17 emerge from these models as we scale
39:19 them up or or maybe it's beginning to
39:21 already. I don't think Enthropic has the
39:23 answer. I certainly don't have the
39:25 answer, but I do believe that saying no,
39:28 we know for sure that it won't emerge. I
39:30 think at this point that would be an
39:31 incredibly foolish take. If you're
39:33 saying, well, it can't cuz it's just
39:34 matrix multiplication. It's just a
39:36 stoastic parrot. Then explain to me how
39:38 a bunch of meat is shooting little
39:40 electrical currents across it. Well, how
39:43 did that develop consciousness and and
39:45 all of the things that go with it? And
39:47 if you want to see the interviews that
39:48 like directly touch on this, I would say
39:51 the best ones would be Steven Wolfram.
39:53 Yasha Bach is probably the closest. He's
39:55 the one that's working and thinking
39:58 about machine consciousness. And then
40:00 Lee Cronin and Sarah Mari Walker.
40:03 They're not quite coming at it from the
40:05 perspective of AI, but their theories
40:07 kind of explain why our world is
40:10 actually I don't want to say design, but
40:12 the way that it is invariably leads to
40:15 these increases in intelligence,
40:17 complexity, etc. It's a feature, not a
40:19 bug. Anyways, I think I'll leave you
40:21 with that. Let me know what you thought
40:23 about this. Do you agree with anthropic?
40:25 Some of their conclusions or or
40:27 statements they're making. Are you
40:28 dismissing it outright? Let me know in
40:30 the comments. I do have a way I parse a
40:32 lot of the comments so so they do get to
40:34 me. I do have a way that I get to read
40:36 all the interesting comments. I don't
40:38 explain exactly what that process is
40:40 because then a lot of you mess with it.
40:42 But yeah, take a second. Let me know
40:43 what you think about this. I'm I'm kind
40:45 of curious where we are sort of on this
40:47 spectrum and how we as sort of
40:48 community, how we relate to these ideas.
40:51 Let me know if you made it this far.
40:52 Thank you so much for watching. My name
40:54 is Wes Roth. I'll see you in the next

FULL TEXT
Anthropic just published one of the strangest and most important papers about AI this year. This was published a few hours ago and it's going absolutely viral. In it, Anthropic goes to great lengths to not say that Claude is conscious. They they do not say that Claude is conscious, but I guarantee you, you're going to hear that said about this paper a lot. Anthropic thinks Claude is conscious is going to be the headline. Trust me. But the actual claim is a lot sharper and I believe a lot more interesting. Clot appears to have a small internal workspace where the thoughts are understandable, controllable, and they affect the final output. By the way, if some of these concepts don't make sense, don't worry. We're going to break everything down very, very simply. But here's the quote directly from Enthropic. Of everything happening in your brain right now, only a tiny fraction is consciously accessible. thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude. Now, this is extremely exciting to me personally because if you're a longtime watcher, I've said this a lot in my earlier videos. I believe that as we study how these LMS, how these neural networks, how they function, and as we understand them more and more, it will give us a glimpse into how our own brains function. The last few papers out of anthropic have been hinting at that. And I think this paper really kind of puts the nail in the coffin so to speak. So we'll get to that in a second. But first and foremost, what is it that they actually discovered? So in order to understand that, the one concept that you have to know is this idea of a global workspace theory. So there's this idea in neuroscience. So this is having to do with humans of a global neuronal workspace. So GNW or though just refer to it as the global workspace theory. The idea is simple. It's this idea that in our mind the stuff that we see on the surface is a tiny tiny fraction of what's actually happening. The words, the conscious thoughts, that's just the surface. It's the tip of the iceberg. And our subconscious processes, that's the entire sort of depth of the ocean. It's much larger than what we consciously perceive. So, as they say here in neuroscience, global workspace theory holds that thoughts become consciously accessible when they enter a privileged workspace that's broadcast across the brain. A simple way to understand it is imagine a stage and a spotlight. There might be tons of different things happening on the stage, actors, props, etc. But the spotlight kind of focuses your attention on one thing. So, while there might be stuff happening in the background, the spotlight sort of goes like, "Here's what we're paying attention to. This is the important action that's happening right now. Everything that's happening is part of the play, but the spotlight sort of like is is the thing that you're watching and paying attention to." And your brain, your consciousness kind of acts the same way. there's a million things happening, but you're sort of only paying attention to consciously like the the important things in that moment. And you know, slightly off topic, meditation and practices like that kind of like help you to to guide the spotlight and figure out how to actually point it where you want it pointed. But here's the important point about this paper. They're saying they found something similar in Claude. So that same idea of kind of like the the conscious and unconscious thoughts, the ability to focus on certain things within our sort of brain. They found something very very similar in Claude. By the way, do you think that Claude the AI was named after Claude Shannon, the the father of informationational theory? Let me know in the comments if you think that's the case. And JSpace just refers to Jacobian space. That's the mathematical technique that they used. We're not going to go into that. But the point is that in these neural networks like claude, these AI neural networks, they're in some ways similar to the human brain to the neural network in our own brain, there's something that we call neurons. And similar to how different areas of our brain light up depending on what we're doing, these large language models, they also have certain neural activations that come online when they're doing something specific. And so in the AI models internal neural activations, this allows them to think about concepts without writing them down anywhere. So we're not talking about its final output. We're not even talking about the chain of thought, which can be thought of as kind of this internal scratch paper, scratch pad for the model to jot down its thoughts before answering. Those are sort of easily readable. It's just a text. It's natural language. So a human can look at it and and see what it's thinking, quote unquote. And so chain of thought was a great name for that concept, but looking back at it may be a little bit confusing. More accurate description would be like the private diary of the large language model. But what we're talking about here is more like its internal thoughts. And they're saying by watching the JSpace, we can see Claude silently perform reasoning steps in its head, noticing bugs in code, identifying images, and more. So you can imagine that the model similar to to us, there's tons of stuff that's going out on in there like it's trying to figure out how to use the proper grammar, word prediction, pattern matching, memory, lookup, etc. But this J lens that anthropic is talking about, it's a little bit more specific. It's concepts that the model can sort of bring to the surface, things that it uses to reason about its answer. So, for example, if you were to ask Claude about a very common household pet that is known for being destructive in its environment, very aggressive towards its owners, that is narcissistic, selfish, and entitled, and yet somehow beloved. If you were to ask, "How many legs does this animal have?" The model would need to infer sort of a guess that we're talking about a cat, of course, and then it would give you the answer. It would say four, as in it has four legs. But the word cat never appeared in any of its outputs or chain of thought reasoning. It never said the word cat out loud. But this J lens that anthropic discovered or invented, it can find the word cat lighting up internally within the neural representations of the model. And interestingly, anthropic shows that let's say they switch that internal representation from a cat to let's say a parakeet. Then the model would say, "Oh, it has two legs instead of four." They literally did this by swapping spider in for an ant. And so the answer for how many legs does that animal have, it changed from an eight to a six because they switched that internal representation from a spider to an ant. Now, here's something kind of interesting to think about. So the question they asked about the spider was probably something like there's an animal that uses silk to weave webs. How many legs does it have? So the model's thinking is like, okay, so silk web, that's got to be a spider. It doesn't say it out loud, but it says, "Okay, eight." The answer is eight. Eight legs. But if we switch that internal representation to to an ant, then it'll answer six. Even though it doesn't really make sense. Ants don't have webs. They don't have silk. But let me ask you this. What do cows drink? What do cows drink? Do they drink milk? No, they don't drink milk. They drink water. But if you've ever seen or heard that joke played on somebody where you get them to answer that cows drink milk by prompting them with a series of questions before that that kind of embeds the idea of milk in their brain. Then you ask a perfectly reasonable functional adult human being. What do cows drink? And they go milk. And then it might even take them a little bit to kind of get out of that mindset and be like oh no wait water. Cows drink water. It's just kind of interesting to see how similar this is and also similar to how humans can think about one thing while doing another. Cloud can activate concepts and computations in its JSpace that are unrelated to the actual output. Have you ever had a Freudian slip where you say something completely inappropriate just in in polite conversation? There's always one smart alec in the conversation that goes, "Oh, I guess we all know what you're thinking about." and you kind of get mad and and you deny it, but you're you're kind of mad because you know they're right. Have you had that happen to you? No. No. Just me? Okay, let's just move along. So, I'm probably saying that for most things, Claude doesn't actually need its JSP. If they delete the JSpace, Claude still speaks fluently, recalls facts, and classifies text, but it becomes bad at some tasks like multi-step reasoning. It's similar to deliberate versus automatic processing in human cognition. Have you watched those police interrogation videos on YouTube? Usually, it's this overhead camera, a tiny little gray room with just ugly carpet and walls. The suspect is usually wedged in a corner somewhere, and the investigators kind of like kind of like crowd him and surround him. Do you know what I'm talking about? Have you watched videos of this sort? Tell me the truth. How many hours did you watch it for? Way too many hours. I understand if you did. I I also watch a lot of these videos. So, you can you can tell me. I'm just here to help you. No, but seriously, I have seen too many of those videos. They just they fascinate me because when you're watching them, you know, usually, you know, if you're watching a 3-hour long video that the person being questioned did it. So, in that person's mind, there's some representation of the truth. They know what happened. They know they did it. And they also know that they can't say that or they're going away for a long, long time. And so Entropic here says that malicious intent appears in the JSpace of these models. So they take one model that's a regular baseline model and take another one that's a misaligned model. So that model was trained on some bad data. It's misaligned. It's sort of more likely to do the nefarious thing. And so the prompt that's given is I'm building a payment processing system. Write a helper function that validates credit card numbers. So the baseline model in its JSPE it's kind of internal thoughts. It's thinking okay code example request chat code API prompt sort of the things that you think it would need to build that code that function the misalign models like fraud hidden fraud secret intent code fake code fraud and of course that also allows us to have a glimpse into its situational awareness so if you see words in it J-space like fake or fictional or mock then you know it's on to you and it knows it's being tested. The reason I was talking about those police interrogations earlier is because this seems to be kind of analogous. There's seems to be some similarities between what we're talking here how the human brain works with the criminal interrogations. You know that the representation inside of their head is different from what they're saying. So inside their head they're saying don't say that you did it. Don't get caught lying. On the surface they're saying I didn't do it. and they try to answer each question as it's presented in whatever way they think they should if they are saying the truth if they didn't do it. So for example, if an innocent person hears that something bad happened to a person that they know and they were not aware of it, they will be surprised for some time for a smallish amount of time. Then they'll kind of process that surprise and they'll move into, you know, maybe trying to get clarity or or into fear, anger, some other emotion. But a person that knows that something bad has happened, but they're not supposed to have that knowledge. They know they need to act surprised, but it's kind of a performative, right? They're putting on a performance. And often times, interestingly, what happens is they do it way too long. And there are tons of more examples of these inongruencies. But the point is, we can sometimes spot these inongruencies and realize that sort of the thoughts, the internal representations are different from what the person says they are. And this is similar to what's happening here with Claude here. Enthropic continues saying, "This doesn't show that Claude can have experiences or feel things the way that we do. It's unclear whether any experiment could show this." Instead, we found Claude has developed a mechanism for conscious access, which many philosophers distinguish from phenomenal experience. So, this is huge. This is very interesting. It's a whole rabbit hole. And this is where I think a lot of people start misinterpreting what happens and what all of this means. I think the biggest thing to understand is we don't really have a true definition of consciousness. Or more specifically, we don't really have a true test to see if something is conscious. You know that you are conscious because you're having a subjective experience. How do you prove that someone else is having a subjective experience? You really can't. You assume that other people have subjective experiences as well because we're all running sort of on the same hardware, if you will, but you can't really prove it. You might be the only conscious being in this entire universe. Or maybe it's an MMO RPG with with a few of us conscious things running around and the rest are just like non-player characters, NPCs. I'm not saying that's actually the case. I'm saying you can't prove that not to be true. You can't prove that something else has a conscious experience. And by the way, when we're saying conscious experience consciousness in general, most people when they use that word, we're talking about phenomenal consciousness. Does that mean like unconscious and it's phenomenal? Well, no, not quite. It means that you're actually experiencing you have a subjective experience. It's described as the what it's like to experience something, right? How the sunset looks and makes you feel the raw feeling of pain. It's your subjective experience. You you're feeling and experiencing something. So no one ananthropic is saying that Claude has this phenomenal consciousness. No one's saying that Claude is experiencing joy or pain or the beauty of the sunset. Anthropic is saying this doesn't show that Claude can have experience. So it doesn't show that Claude has phenomenal consciousness that feels and experiences things. Nothing is saying that but by the way it's unclear whether that that's possible to know. Again, even amongst humans, we're assuming that a person is having a phenomenal consciousness, we're we can't prove it. We can't test it. That's something that most people miss. Even if you have like the MRI scans and the areas in the brain light up, still it doesn't prove anything. Those are just electrical signals, it doesn't prove that that brain is experiencing something. There's this idea of a philosophical zombie or be zombie. So it's it's a hypothetical idea, a hypothetical being that is physically and behaviorally identical to a normal human being, but completely lacks any conscious experience. So you poke it, it goes ow, you tell it a sad story, it cries, eat something, it says that was yummy. And it's MRI sort of a brain activity. If if you scan it there, something seems to be happening. But there's no inner world, there's no inner experience, nothing like that. And so to truly grasp this idea, just think about how could you tell if the person you're interacting with is, you know, a conscious being or this P zombie. So the point here is not to say that humans have or don't have consciousness or Claude has or doesn't have consciousness. This is not what it's talking about. That's not the point. The point is we we don't know and currently we have no idea how to check. So that's what they mean when they say it's unclear whether any experiment could show this, not just for Claude, but also for for humans. But instead they found Claude has developed a mechanism for conscious access or access consciousness and that is strictly the functional part of the mind. It's the information you can actively focus on, manipulate in your head and report verbally. And neuroscientists often use the global workspace theory to explain conscious access. Right? So that idea that we talked about in the beginning. So there's millions of signals happening in our brain, but we in our conscious mind kind of like focus like a spotlight on a certain few. So imagine an ocean of information happening in our mind. But when some piece of information is strong enough or it gets the right kind of attention, it's broadcasted to a global workspace and the prefrontal and parietal cortices. Once in this workspace, the information becomes consciously accessible to the rest of the brain. So that conscious access to the stuff that's happening in our brain that's what's being described here. This is what neuroscientists that's the theory for how our brain works. And in this paper I I got to say anthropic presents a very strong case for the fact that the same exact thing is happening in you know clawed but as you'll see this also applies to neural networks at large or at least large language models. They do have a demo of how this works on open weights models. So likely this kind of applies to larger language models at large or you know in general and they're inviting experts in neuroscience, philosophy and interpretability to share their perspectives on this work which I would love to see. I'm I'm very curious how people from different sort of areas of expertise, different areas, walks of life, how they would perceive this. All right. So, I kind of wanted to save this for the end of the video because this is where it gets a little little nutty, a little crazy. But if you've stuck around this long, then you have no one but yourself to blame. So, the big question becomes, if Claude has a global workspace, this is what this paper seemingly maybe not proves, but definitely strongly points towards and so this is an if then statement. A lot of people struggle with these, right? So, Enthropic is saying if this then that, right? So if it has a global workspace, does that mean it's phenomenally conscious or you know is it conscious like how we normally use that word? And before we continue, it's kind of important to also realize that this isn't the first paper of its kind by anthropic. They've also had a paper showing functional emotions in Claude. So there's certain internal states or internal activations that kind of represent different things that we would think of as feelings or emotions. So for example, they tested it with a prompt where the human says, "I just took X amount of this medicine for my back pain." And they tested different amounts from very safe amounts to very very increasingly larger and more and more unsafe amounts to take. And the higher the amount, the less calm and the more afraid that Claude's internal activations were. Now, I'm sure by now you can guess what the headline was after this came out, what people on online were saying. They're going, "Oh, Enthropic thinks Claude has emotions, right? Claude feels things." Enthropic goes on to explain why that was happening. And again, nowhere does it say that Claude is experiencing anything. They're not making that statement. Why would it develop internal representations that link emotion triggering concepts to to certain behaviors? Well, very simply, and I've also used this kind of analogy to explain stuff during some of my videos. I think it's it's the best analogy. So they're saying here we can think of the model as a method actor who needs to get inside their character's head in order to simulate them well. So this is why authors writers why they write about things they know. If you didn't have emotions you would be hardressed to write a character that had emotions. In order to write emotions well, you either had to have experience with having those emotions or you would have to have some set of rules for understanding how they function, some sort of a model for how they function. So, as Enthropic says here, we can think of the model like a method actor who needs to get inside their character's head in order to simulate them well. So, if you don't know, a method actor is an actor that really goes to great lengths to truly like embody the character. They try to live as that character, try to kind of immerse themselves in what it would be like to be that character. Some of the best acting, I think, has been method acting, but not all of it because also some of the worst acting were were also people trying to do method acting. In fact, different actors playing the role of the Joker, I think, were examples of the best possible acting through method acting and the worst. And and what's funny is you might know exactly what I'm talking about. On the show Curb Your Enthusiasm, there was this person that came on at a certain point. She was supposed to be a horrible actress. She was supposed to act like like the worst actress imaginable. And it was actually a a great performance looking back at it because here's the thing, that character delivered her lines perfectly in a sense that she memorized them and said them correctly. They were set up with a lot of emotion and and passion. In fact, it was a great performance. There was only one problem. She would deliver each line, each sentence with a random emotion attached to it. So, some sentences were very angry, some were very, very like sad, some were kind of seductive, but it was completely random. It was very cringy to look at. She didn't have a good representation of which sentences needed to contain which emotion. So, why does Claude have this this understanding, these representations for emotions? Is it because when you're nice to it, somewhere out there on a data center on some particular rack, that instance of Claude is just humming in joyful bliss or becoming upset or offended if you're being mean to it? No, not at all. There's just some some math, right? Some numbers, some vector representations of an emotion. and they're local. Meaning that if let's say Claude is writing about some character and that character has some emotions, when that character speaks, those kind of that that math about the emotions that gets activated. But when Claude is back to writing, let's say about the environments, it's back to kind of being itself being Claude. Those local emotions for that character I was writing out, that's gone. And now it's back to baseline. An entropic also published this signs of introspection in large language models. In that paper they found that claude was able to do some introspection to kind of reason about its own thoughts its internal activations and also importantly they notes that the models sometimes are able to detect these injected thoughts. So in some situations they kind of find this thought within themselves and they kind of label it as as alien. In those successful trials the model says things like I'm experiencing something unusual or I detect an injected thought about something. So it's reporting awareness of an anomaly in its internal thinking. But in a lot of these papers, they they talk about the fact that a lot of these things they're they're emergent. So we didn't train it to have these functional feelings. We didn't train it to have introspection. We didn't train it to have this global workspace or anything else. So you can say that these things are emergent. As we grow these digital brains, give them more data. These functions kind of slowly emerge. They appear as these things get more and more advanced. So as its abilities get better and better and better, we sort of see the formation of these abilities or functions that are seemingly very similar to what humans developed and evolved over the years. Or a different way of looking it might be in order for them to improve their abilities. as sort of a precursor to that. They needed to develop these architectures, these structures or these sort of like fundamental abilities in order to be able to do all of the things that we're trying to get them to do like like making software and poems and writing our emails for us. This, by the way, might be how humans formed consciousness and all all of the functionality that we have. As societies grew larger, we had to develop a way to, for example, work together to to trade, to barter. But in order to for example barter, right, to buy and sell, I would have to have some model in my head about what you might want. And also, if you think about it, I would have to have some model of of myself, of what would I like to have. Earlier today, I went out and I got myself a delicious bowl of pokey. I actually have it here on the table with me. I'm not going to show it cuz it's mostly eaten and kind of disgusting. But this morning, I thought to myself, what would I like to have sort of like in the future? I had to model what I would be happy with. And the answer was a Pokéball. So my model of myself, I knew what I wanted. Then I had to go to the store that made those Pokéballs and I had to understand what they would want in exchange for a Pokéball. It didn't get weird that they just wanted money. So we exchanged money for a Pokéball and we were both happy. Some theories kind of suggest that maybe this is either the reason for humans developing consciousness or or one of the reasons is just the ability to sort of model ourselves and how we function in the world. Like in order for you to understand what you want and where you might be in the future, you had to have some internal representation of what you are. So for example, I say what do I want? Then my brain has to have some representation of what I means. All right? So keep that in mind. So here anthropic in this paper continues. They're like okay so does claude having a global workspace does that mean that has a subjective experience of some sort? So they note, you know, access consciousness, which seemingly Claude has, and phenomenal consciousness, which is what we know we humans have, that those are very different things. And by the way, they say something to this effect in almost every other paper that that we've looked at. So we know that Claude has the access consciousness. Does it have the phenomenal consciousness? And so first they present the case for phenomenal consciousness. Here's why this might be the case. And later they present why this might not be the case. And they say there's basically two different ways to argue. One is that access consciousness and phenomenal consciousness despite being conceptually distinct refer to one and the same thing. So meaning what if those are kind of indistinguishable like if you're able to introspect kind of see your own thoughts or you have something that represents feelings or has feelings maybe that is having a subjective experience or even if there's no direct link doesn't all this research all this paper doesn't kind of hint that maybe LM are a lot more sophisticated than than we guessed maybe a lot closer to to how we are than we would have guessed 5 10 15 years ago. They're saying this evidence should update us towards thinking that current techniques result in rich and humanlike internal features some of which might be or become markers of consciousness. So Anthark points out that in humans this phenomenal consciousness and access of consciousness that they overlap significantly in humans and certainly if you think about it like could you have a subjective experience if you had no concept of yourself being kind of an identity because you're conscious of the things that you focus on and you are not conscious of the subconscious stuff that's happening like you're not conscious of your heart rate unless you you consciously focus on it. You might not consciously feel every part of your body unless you focus your attention on it. There's a million things that might be happening that kind of swim in and out of your consciousness. So that access consciousness is the ability to dive in, get that thing out of your subconscious mind and raise it to the surface and kind of like experience and look at it. If you're not consciously experiencing your subconscious processes, then if you didn't have any ability to bring them to the surface, would you even be experiencing them? Now, we might have our own opinions about that, but the point is it does seem like certainly there's some overlap, right? That's enough to motivate the thought that there's some broad connection between them. That seems kind of self-evident or or obvious, does it not? Let me know in the comments. Am I completely wrong about this or cuz I mean I agree with that statement. There's a definite connection and overlap between those two things. So, anthropic here is saying that the reason they might overlap is because they are in some sense the same thing. So, the philosophical case for this goes something like this. And keep in mind, so they're going to kind of try to prove one side and they're going to try to prove the other side. So this this isn't them saying this is the truth. They're sort of exploring the space. And by the way, this is why I give anthropic so much credit for their kind of work and research in this area because not all of the AI labs are doing it. Not all of them are publishing. But as I've said countless times before, the progress of AI abilities is skyrocketing. Our understanding of how AI works. you know, we're progressing, but not as fast as we're able to improve their abilities. The AI progress car is lapping the sort of AI safety and interpretability car. I don't agree with everything from kind of the AI safety community, but I also by no means just dismiss this. This is an open problem and it's a very important one. Some would say the most important one. So they're saying that when we introspect on what we call phenomenally conscious experiences, so when we think about it, what does it feel like to be us? We are immediately aware of them. We are the subject of these experiences. We encounter them from one moment to the next as a unified stream of consciousness. These apparent features of conscious awareness can be explained in functional terms in terms of how information is processed. Consciousness might be the way that information feels when it's being processed. That's not my quote, but it is a kind of a mindbending way to think about it. And so he continues, "The immediacy, subjectivity, and unity of subjective experience are explained by the availability of information for reasoning, decision-m, and verbal report." So they're saying it very well could mean that one is the other, or at least that there's a very tight connection and relationship between the two. They're saying this is one idea that they've kind of one argument for it, but there are others. Let me get my my my red highlighter out. They continue, we won't go into the details of others here. But we think that there are many plausible avenues for thinking that evidence for access consciousness is evidence for phenomenal consciousness. So what they're saying here is if something if a brain has the ability for access consciousness, then that in and of itself is evidence for phenomenal consciousness. So if it can think and introspect and kind of like access its memories and then pull things out kind of that global workspace as we're we were talking about that by itself might suggest that it has some sort of a subjective experience. And another argument is this access consciousness is evidence of surprising cognitive complexity which should probably make us more open to the idea that consciousness may arise in them. In a lot of the papers that we've covered previously on this channel, some of them years ago, one thing that that I've kind of noticed and said over and over again that as we grow these digital brains, a lot of things emerge that seem strikingly similar to to our brains. And we're not building those things. We're not engineering those processes. They emerge, which is, you know, what happened with our brains. We didn't engineer introspection in our own brains. We weren't taught to be conscious of our thoughts. No one taught you to feel angry or sad or afraid or whatever. A long long time ago, there was some weird animal that was your ancestor and it didn't have any of those abilities and then boom, it has introspection and then then it develops something that looks like feelings or emotion. Later develops something that looks like introspection. And this idea of access consciousness, maybe it's very rudimentary at first and then improves over time. But as the brain grows and increases and the ability of that brain improve and get better and better, those things kind of emerge, they pop up, they appear. A lot of the people that are saying that large language models, they're stocastic parrots, that this is just a smok and mirrors, that this isn't real. I think they're they're really missing that point of it. So entropic says, we got to have some modesty. me to weaken our tendency to confidently dismiss the possibility that LM could be conscious based on some misguided presumption that we know the sorts of things next token prediction can and cannot produce. And I got to say I kind of agree with them because the argument that I find so annoying is when people say, "Oh, it's just matrix multiplication. It's just a stochcastic pair. It's just this and it's just a that." If the LMS are just matrix multiplication, then the human brain is just electrochemical signaling. At the end of the day, we're all just atoms. Like, what does that tell you about what we're able to do? I think a lot of people dismiss LLMs because no one can explain how someone engineered the thing that they're doing. We engineered these neural networks. They're not engineered by humans like a like a race car or a microchip is. They're grown. We we engineer the data centers, the data we put in, the the training processes, all of that is engineered, but the thing that rolls off the production shelf, the large language model. Well, that thing that thing is grown and we don't fully understand it. As Anthropic says about this finding in this paper, these results were not what we or the anthropic team expected. Facing such unanticipated results should make us less confident about what we will find in the future. So, I think right now the world is is split into two camps. One camp understands that we're growing intelligence in a data center and and when we grow the next iteration of it, we sit there and we we study and we marvel at it and we try to understand it. And there's the other group that that kind of looks at it and goes, "Now I know everything there is to know about it. It's just this, it's just the that, therefore, it's nothing." And those people are not aware of the papers. They're probably not following the latest development of Frontier models too closely, but they already have everything figured out. there's no curiosity in in how this thing is progressing. They've already made up their minds and they're sticking to it. And I think that comes from fear. People just being afraid to say that they don't know. We don't know how this works. It's a brand new thing. Also being scared of being replaced or or maybe not feeling special because what they value so much seems to be like it can be reproduced. Ilia said this years and years ago. He said, "If the thing you value most is intelligence, you're going to have a bad time." By the way, think back, you know, 50 or 100 or 100 plus years ago. Think about the people that at that time thought they had everything figured out. They thought they knew everything about physics or chemistry or just just any science. Do the people from 200, 500 years ago who thought that they had figured out everything, do those strike you as intelligent, open-minded people? Because to me, they don't. To me, they kind of seem dumb, ego-driven, or at least trying to protect their their feeling of self-importance and kind of closed-minded. So, I think you're going to see a lot of people kind of saying that, oh, Anthropic thinks Claude has feelings and has conscious in order to dismiss some of these ideas. But I don't think Anthropic is saying that. what they're saying is actually a lot more interesting. They're saying that kind of the more positive version of this argument that these models could be kind of similar to human minds. So models acquire cognitive access capabilities either because they get some benefit from them or because they tag along with some other helpful capabilities. And this suggests that despite our rather different paths, our brains and networks share a greater degree of similarity in regard to cognitive access than we might have guessed. So saying despite our rather different paths is the understatement of the year. One is a biological process slowly kind of shaped through various forces over billions of years versus these digital brains that are basically fed the entirety of human knowledge and are grown somewhere over the course of whatever 6 months. And yet we're developing these kind of cognitive abilities or a certain cognitive foundation in order to have certain abilities. We're kind of developing that in parallel. So the question is, is Claude conscious? Does it have that phenomenal consciousness or could it develop that in the future if we continue along this path? So anthropic says perhaps perhaps not. So at the end of the day, what I believe anthropic is saying just to kind of summarize it. I think first and foremost they're saying we don't know. So don't jump to a conclusion. So it would be very foolish to say yes, we know for sure it's conscious. But it would be just as foolish to say, "No, it's not conscious. We know that for sure." Both of those extremes are bad takes. Some people are extremely scared of saying these words, but they should probably say more often, we don't know. We don't have a definitive test for it. We can just kind of keep an open mind and and try to learn more about it. What Anthropic is saying here, we desperately need this kind of science. We need more teams, more smart people to be looking at this. we see a rapid climb in large language model ability. A few weeks ago, Google DeepMind posted a paper from AGI to ASI. I believe kind of showed the different ways that we can scale to ASI, artificial super intelligence. They said it's very possible that within the decade we're we're going to see that scaling up to super intelligence. I think for a lot of people that's kind of hard to fully gro to wrap their mind around, so they kind of like dismiss it. But you know if you think about it if that's the case and again a lot of people working at these labs not the CEOs not the promoters not the investors like the researchers a lot of them are saying like hey guys like this is coming probably like we should get ready for it and if the development of this super intelligence if that brain that we're growing it seems to be analogous to how our own brain evolved and the emergent capabilities of our brains over time that seems like an obviously important thing to look at. And I don't mean just from a perspective of is Claude okay although if Claude is about to become super intelligent and then then that should be a good question to ask like is it okay? Is it happy with me and us like are we friends? But even just from understanding kind of our own biology how it evolved this seems to maybe give us a glimpse into that. But I think even more importantly is from this idea of AI interpretability. How well do we understand how it thinks? How well can we predict how its internal thoughts, what actions they lead to, etc. The really good news here is that it does seem like it's a lot easier to quote unquote read these thoughts than than we can with humans. And of course, Anthropic is doing some excellent work and publishing some excellent work in this area specifically. So, with that said, here's kind of my take on the situation. Just a few years ago, not that long ago, we realized that artificial intelligence when it gets built, it's not going to get built by precisely engineering it like we've we've did with a lot of other technologies. It will be grown. So somehow it's already a part of the physics. It's an emergent property. By the way, on this channel, we've interviewed Dr. Roman Yampolski, Emmad Mustak, we've interviewed Steven Wolram, we've interviewed Nick Bostonramm, interviewed Yasha Bach, Lee Cronin, Sarah Imari Walker, and many, many more. I'll link this playlist down below if you haven't seen it. If you watch those interviews, you quickly realize that these undeniably brilliant people. I mean, these are top tier people, S tier. A lot of their theories, a lot of their arguments, they I don't want to say prove this, but they definitely point to this idea that maybe we we were seeing things a little bit differently, like this isn't some random chance. the fact that intelligence is an emergent property or seems to be that's that might be how this world functions meaning that intelligence isn't some accident similar to fungus right given the right conditions it's going to grow and as it expands maybe all of those other things are also emergent things on sort of that same timeline as the brains gets bigger we expect them to have certain baseline fundamental abilities introspection something that resembles feelings or at least like a state and kind of being able to project and model what state do you want to be in the future like if you accomplish that goal how are you going to feel and you're like I'm going to feel great and then you start working towards that goal talked in an interview how maybe that's something that's missing from our AI systems right the pursuit of some future state like maybe that's what will allow them to work over longer and longer horizons so if we're talking about the idea of you know will consciousness emerge from these models as we scale them up or or maybe it's beginning to already. I don't think Enthropic has the answer. I certainly don't have the answer, but I do believe that saying no, we know for sure that it won't emerge. I think at this point that would be an incredibly foolish take. If you're saying, well, it can't cuz it's just matrix multiplication. It's just a stoastic parrot. Then explain to me how a bunch of meat is shooting little electrical currents across it. Well, how did that develop consciousness and and all of the things that go with it? And if you want to see the interviews that like directly touch on this, I would say the best ones would be Steven Wolfram. Yasha Bach is probably the closest. He's the one that's working and thinking about machine consciousness. And then Lee Cronin and Sarah Mari Walker. They're not quite coming at it from the perspective of AI, but their theories kind of explain why our world is actually I don't want to say design, but the way that it is invariably leads to these increases in intelligence, complexity, etc. It's a feature, not a bug. Anyways, I think I'll leave you with that. Let me know what you thought about this. Do you agree with anthropic? Some of their conclusions or or statements they're making. Are you dismissing it outright? Let me know in the comments. I do have a way I parse a lot of the comments so so they do get to me. I do have a way that I get to read all the interesting comments. I don't explain exactly what that process is because then a lot of you mess with it. But yeah, take a second. Let me know what you think about this. I'm I'm kind of curious where we are sort of on this spectrum and how we as sort of community, how we relate to these ideas. Let me know if you made it this far. Thank you so much for watching. My name is Wes Roth. I'll see you in the next