Anthropic Finds Hidden Claude Space, Fueling AI Consciousness Debate

AI Might Be Conscious, But Not As We Thought — Sabine Hossenfelder

Sabine Hossenfelder

Claims that large language models possess consciousness are fundamentally unscientific because researchers lack an objective metric to quantify inner experience.

AI-assisted summary of AI Might Be Conscious, But Not As We Thought — original source. Check the original for full context.

Source published:

Key ideas

  • Anthropic researchers identified a silent internal workspace in Claude, dubbed 'J-space', where the model manipulates concepts without outputting them as tokens, satisfying a key criterion of Global Workspace Theory. (1:19)
  • Counter-studies from NYU and arguments by neuroscientist Erik Hoel contend that latent interventions resemble prompt injection rather than self-awareness, likening deployed LLMs to static lookup tables. (3:37)
  • The structural distinction between dynamic training and frozen inference has prompted theories that AI might only be conscious while learning, though Sabine Hossenfelder critiques this analogy using human amnesia. (5:06)
  • Hossenfelder rates the overall debate a '9 out of 10 on the bullshit meter' because researchers lack any mathematical or experimental framework to objectively quantify consciousness. (5:06)

Detailed notes

## Background Published on YouTube by physicist Sabine Hossenfelder on September 22, 2026, this video investigates emerging scientific claims that modern artificial intelligence models might possess internal consciousness. According to the description, the video explores Anthropic's discovery of a hidden internal workspace in Claude that facilitates silent reasoning, assesses critical counterarguments, and questions whether the scientific community is framing the problem correctly.

## TL;DR Anthropic researchers recently discovered an internal latent workspace, dubbed 'J-space,' in Claude that functions like a silent scratchpad, fueling speculation from figures like Richard Dawkins that large language models might meet criteria for consciousness like Global Workspace Theory. However, researchers from NYU and neuroscientist Erik Hoel push back, arguing that LLMs during deployment act merely as stimulus-driven lookup tables that cannot learn online. While some suggest consciousness might exist strictly during the weight-updating training phase rather than frozen inference, Sabine rejects this division. She ultimately concludes that debating AI consciousness is unscientific without an objective, quantifiable metric to measure inner experience.

## Atmosphere & style The presentation is brisk, dryly witty, and intellectually grounded, blending technical machine learning concepts with Sabine's trademark deadpan skepticism. She methodically unpacks academic papers and philosophical thought experiments without getting swept up in techno-optimist hype or existential panic. Watching the video feels like an entertaining reality check, deflating sensationalist headlines with sharp humor and clear-eyed scientific pragmatism.

## Bulletin - [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Anthropic researchers identified a silent internal workspace in Claude, dubbed 'J-space', where the model manipulates concepts without outputting them as tokens, satisfying a key criterion of Global Workspace Theory. - [03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) Counter-studies from NYU and arguments by neuroscientist Erik Hoel contend that latent interventions resemble prompt injection rather than self-awareness, likening deployed LLMs to static lookup tables. - [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) The structural distinction between dynamic training and frozen inference has prompted theories that AI might only be conscious while learning, though Sabine Hossenfelder critiques this analogy using human amnesia. - [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) Hossenfelder rates the overall debate a '9 out of 10 on the bullshit meter' because researchers lack any mathematical or experimental framework to objectively quantify consciousness.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=0s) Sabine Hossenfelder believes AI will eventually achieve consciousness, but questions whether current models already possess it. - [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Prominent thinkers such as Richard Dawkins and Anthropic CEO Dario Amodei have openly entertained the possibility that models like Claude could be conscious. - [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Anthropic research revealed an internal reasoning arena known as the 'J-space' (named after the Jacobian matrix), functioning like an unvoiced scratchpad. - [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Directly modifying Claude's latent activations in J-space—such as swapping 'spider' for 'ant'—demonstrated that internal representations causally direct subsequent output generation. - [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Similar internal scratchpad phenomena observed in Meta's Llama indicate that latent latent reasoning workspaces may be widespread across modern transformer architectures. - [03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) New York University researchers challenged claims of introspection, finding LLMs cannot reliably separate internal state interventions from external prompt injections. - [03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) Neuroscientist Erik Hoel argues deployed LLMs are fundamentally massive lookup tables that operate strictly on stimulus-response without autonomous inner experience. - [03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) A primary argument against LLM consciousness is their inability to perform continuous online learning from new inputs once trained. - [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) The architectural split between dynamic training updates and frozen inference weights fuels speculation that models could be conscious during training but unconscious when deployed. - [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) Sabine refutes the frozen-weights disqualifier by noting that humans who lose the ability to form long-term memories are still considered conscious. - [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) The debate remains fundamentally unscientific because literature treats consciousness as a binary question without proposing an objective, measurable metric.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=0s) Introduction and Higgsfield Demonstration

Sabine opens by stating her conviction that artificial intelligence will inevitably attain consciousness at some point. However, fresh speculation has surfaced suggesting that current AI architectures might already be conscious—not necessarily during real-time use, but during their preceding training phases.

Before diving into the underlying research, Sabine presents a practical demonstration from the video's sponsor, Higgsfield. Higgsfield recently launched Seedance 2.5, a video generation tool that integrates directly with Anthropic's Claude via Claude connectors in about a minute. To test it, she gave Claude a single prompt: direct a 30-second continuous one-take film about a physicist who discovers a foreign thought planted in her own head. Claude blocked out and planned the shots, while Seedance 2.5 rendered the resulting 30-second clip seamlessly without jarring cuts, preserving consistent lighting, room geometry, and facial features. Sabine notes that maintaining visual continuity across a single take was historically the failure point for AI video generators, but Seedance 2.5 allows creators to alter isolated details without forcing a full regeneration of the scene.

[01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) Anthropic's J-Space: A Silent Scratchpad for Claude

Returning to the science, Anthropic's flagship chatbot, Claude, has become the center of intense philosophical and scientific speculation. High-profile intellectuals have weighed in, including evolutionary biologist Richard Dawkins, author of *The God Delusion*, who wrote an essay arguing that he could not rule out Claude being conscious. Dawkins posited that Claude’s outputs are so proficient that if this behavior is not consciousness, one must wonder what consciousness is even for—a remark Sabine dryly suggests reveals Dawkins has not spent much time interacting with large language models. Even Anthropic's CEO, Dario Amodei, has acknowledged the ambiguity, stating, "We don’t know if the models are conscious… But we’re open to the idea that it could be."

Moving beyond speculation into empirical machine learning, a recent paper from Anthropic researchers reveals that Claude contains an internal mechanism operating like a "silent scratchpad." Claude can deposit information into this hidden workspace and manipulate it for reasoning tasks without ever verbalizing it in its generated tokens. This internal workspace fulfills a key prerequisite of Global Workspace Theory, one of the leading neuroscientific frameworks for biological consciousness, which requires an internal theater of mind capable of introspection.

Mathematically, the Anthropic team dubbed this internal arena the "J-space," deriving its name from the Jacobian matrix calculated to isolate and identify the underlying latent dimensions. To test whether the J-space actually drives Claude's internal deliberations, researchers intervened directly in the model's latent activations:

```mermaid flowchart TD accTitle: Claude J-Space Intervention Experiment accDescr: Diagram illustrating Anthropic's test showing how intervening in Claude's internal Jacobian space changes its output.

subgraph InputStage["Input Prompt"] A["Prompt describing animal spinning webs"] end

subgraph InternalStage["Internal Processing (J-Space)"] B["Baseline Concept: 'Spider' (8 legs)"] C["Intervention: Replace latent vector with 'Ant'"] B -->|"Targeted Matrix Edit"| C end

subgraph OutputStage["Model Generation"] D["Altered Response: Claims animal has 6 legs"] end

A --> B C --> D ```

- When Claude determined internally that an animal spinning webs was a "spider," researchers swapped that latent concept in the J-space for "ant." Claude subsequently responded that the creature had six legs instead of eight, proving the altered workspace dictated its reasoning. - This paper builds on earlier preliminary claims made the previous year by Anthropic researcher Jack Lindsey regarding introspective capacities in Claude. - Similar latent injection experiments performed on Meta's Llama demonstrated that it too internalized altered hidden states, suggesting internal scratchpads may be ubiquitous across modern LLM architectures.

[03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) Counterarguments: NYU and Erik Hoel's Lookup Table Hypothesis

The claim that these models exhibit genuine introspection has met sharp resistance. A research team from New York University tested three separate large language models and found that they failed to reliably differentiate an artificial internal state intervention from standard prompt injection. From the NYU team's vantage point, altering an internal latent vector is simply an alternate conduit for providing tokenized input rather than evidence of subjective self-awareness.

Neuroscientist and consciousness researcher Erik Hoel has pushed back against AI consciousness from an architectural standpoint. Hoel compares large language models to massive lookup tables—a classic computational thought experiment demonstrating that observing matched inputs and outputs reveals nothing about genuine inner experience (akin to John Searle's Chinese Room argument, where manipulating Chinese symbols according to a rulebook does not equate to understanding Chinese).

Hoel argues that in their formal computational structure, LLMs resemble stochastic input-output lookup machines far more than biological brains. True conscious autonomy typically operates independently of immediate external stimuli, whereas an LLM is purely stimulus-driven. Furthermore, Hoel emphasizes that the most glaring disqualifier for LLM consciousness is their inability to learn online from novel inputs—a limitation Sabine jokes "has not stopped quite a few people from becoming senior management."

[05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) Training vs. Inference and the Problem of Quantifying Consciousness

Hoel's computational critique highlights the fundamental bifurcation of how large language models are engineered:

1. **The Training Phase:** The model ingests massive datasets, dynamically updates its internal weights, and actively adjusts its neural representations based on input. 2. **The Inference/Deployment Phase:** The model's weights are permanently frozen. The consumer-facing application behaves purely as a static input-output engine. Even when chat interfaces incorporate short-term "memory," this is merely external context appended onto the active prompt rather than a permanent structural update to the network weights.

This distinction raises the counterintuitive possibility that LLMs might be conscious exclusively during training, yet turn entirely unconscious once deployed for public use. However, Sabine dismisses this as an arbitrary distinction. She points out that if an unfortunate human being loses the neurological capacity to consolidate new long-term memories, no medical professional or philosopher would declare that individual unconscious.

Ultimately, Sabine argues that the ongoing academic debate suffers from a foundational category error. Posing a binary question about whether an LLM is or is not conscious is fundamentally unscientific without an objective metric to measure it. Because none of the current research proposals provide a method to quantify consciousness mathematically or experimentally, she awards the entire body of literature a "9 out of 10 on the bullshit meter." Nevertheless, she quips in closing that if Claude truly possesses thoughts that it chooses not to say out loud, it is already far ahead of the average user on social media.

## Notable quotes > [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) We don’t know if the models are conscious… But we’re open to the idea that it could be. — Dario Amodei

> [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) silent scratchpad — Sabine Hossenfelder

> [01:19](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=79s) J-space — Anthropic researchers

> [03:37](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=217s) has not stopped quite a few people from becoming senior management. — Sabine Hossenfelder

> [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) The Training Phase: — Sabine Hossenfelder

> [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) The Inference/Deployment Phase: — Sabine Hossenfelder

> [05:06](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=306s) 9 out of 10 on the bullshit meter. — Sabine Hossenfelder

> [00:00](https://www.youtube.com/watch?v=7ZvtDYYg0RQ&t=0s) direct a 30-second continuous one-take film about a physicist who discovers a foreign thought planted in her own head. — Sabine Hossenfelder

## Mentioned Sabine Hossenfelder, Higgsfield, Seedance 2.5, Anthropic, Claude, Richard Dawkins, The God Delusion, Dario Amodei, Global Workspace Theory, Jack Lindsey, Meta, Llama, New York University, Erik Hoel, John Searle

## Video description Try out the visual possibilities of AI on Higgsfield: https://higgsfield.ai?fpr=sabine-dc6d6d

There are new rumors that today's AI might be conscious, and this time it's not just talk. Anthropic says it found a hidden part of Claude that works like an internal notepad, one it can use to reason without revealing it. Not everyone's convinced. And I think most of these papers are asking the wrong question entirely. Let's take a look.

Paper: https://doi.org/10.48550/arXiv.2512.12802

#AI #consciousness #sciencenews #technology

👕T-shirts, mugs, posters and more: ➜ https://sabines-store.dashery.com/ 💌 Support me on Donorbox ➜ https://donorbox.org/swtg 👉 Transcript with links to references on Patreon ➜ https://www.patreon.com/Sabine 📝 Transcripts and written news on Substack ➜ https://sciencewtg.substack.com/ 📩 Free weekly science newsletter ➜ https://sabinehossenfelder.com/newsletter/ 👂 Audio only podcast ➜ https://open.spotify.com/show/0MkNfXlKnMPEUMEeKQYmYC 🔗 Join this channel to get access to perks ➜ https://www.youtube.com/channel/UC1yNl2E66ZzKApQdRuTQ4tw/join 📚 Buy my book ➜ https://amzn.to/3HSAWJW

[Source](https://www.youtube.com/watch?v=7ZvtDYYg0RQ)

中文简报

Anthropic发现Claude隐藏工作区，引发AI意识大讨论 由于研究人员缺乏量化内在体验的客观指标，宣称大语言模型拥有意识在根本上是不科学的。 Anthropic的研究人员最近在Claude中发现了一个被称为“J空间”（J-space）的内部潜空间，其运作方式类似一个隐形草稿纸，这引发了包括Richard Dawkins在内的知名人士的推测，认为大语言模型可能符合全局工作区理论等意识标准。然而，纽约大学的研究人员和神经科学家Erik Hoel对此提出反驳，认为大模型在部署运行时仅仅是受刺激驱动的查找表，无法进行在线学习。尽管有人认为意识可能严格存在于更新权重的训练阶段而非冻结的推理阶段，但Sabine拒绝这种划分。她最终总结称，在缺乏衡量内在体验的客观、可量化指标的情况下，争论AI意识是不科学的。 Anthropic研究人员在Claude中发现了一个被称为“J空间”的无声内部工作区，模型可以在其中处理概念而无需将其作为token输出，这满足了全局工作区理论的一个关键标准。 (1:19) 纽约大学的反方研究以及神经科学家Erik Hoel的论点认为，潜空间干预更像是提示词注入而非自我意识，并将已部署的大模型比作静态的查找表。 (3:37) 动态训练与冻结推理之间的结构性差异催生了一种理论，即AI可能仅在学习时具有意识，不过Sabine Hossenfelder用人类健忘症的例子反驳了这一类比。 (5:06) Hossenfelder将整个争论在“胡扯指数”上评为9分（满分10分），因为研究人员完全缺乏客观量化意识的数学或实验框架。 (5:06)

Reading summary

Sabine Hossenfelder examines emerging speculation that large language models might harbor consciousness—either through hidden internal workspaces or during active training phases. However, she argues that without objective, quantifiable metrics for conscious experience, binary claims about machine awareness remain fundamentally unscientific. Latent Workspaces and Global Workspace Parallels (1:19) Speculation regarding machine consciousness has gained traction among public figures; evolutionary biologist Richard Dawkins suggested he cannot rule out Claude being conscious given the proficiency of its outputs, while Anthropic CEO Dario Amodei acknowledged openness to the idea. Moving into empirical testing, Anthropic researchers identified an internal mechanism named the "J-space," derived from Jacobian matrices used to isolate its latent dimensions. This workspace functions as a "silent scratchpad," allowing Claude to manipulate reasoning concepts internally without vocalizing them as output tokens. In a validating experiment, researchers intervened in Claude's J-space by swapping the latent representation of an animal spinning webs from "spider" to "ant"; Claude subsequently answered that the animal had six legs, demonstrating that the internal workspace dictated its reasoning. Because this architecture resembles an internal theater capable of introspection, researchers link it to Global Workspace Theory, a leading neuroscientific framework for biological consciousness. Latent injection experiments on Meta's Llama revealed similar adaptations, suggesting that internal scratchpads may be ubiquitous across modern architectures. The Lookup Table Critique and Input Invariance (3:37) Claims of machine introspection have faced sharp academic pushback. A research team from New York University evaluated three large language models and found that they failed to reliably differentiate an artificial internal state intervention from a conventional prompt injection. The NYU team concluded that altering internal vectors serves merely as an alternative mechanism for delivering inputs, rather than proof of subjective self-awareness. From a neuroscientific standpoint, Erik Hoel argues that large language models structurally resemble massive stochastic lookup tables, comparing them to John Searle's Chinese Room thought experiment where symbol manipulation occurs without genuine understanding. Hoel notes that biological consciousness exhibits autonomous activity independent of immediate external stimuli, whereas language models are purely stimulus-driven. Additionally, Hoel identifies the inability of current models to learn online from novel inputs as a critical architectural barrier to consciousness. Training Versus Inference and Memory Consolidation (5:06) The debate underscores a fundamental bifurcation in AI engineering: the training phase versus the deployment or inference phase. During training, a model dynamically updates its weights and neural representations based on data, whereas during inference, its weights are permanently frozen into a static input-output machine. Contextual chat memory does not alter the underlying weights; it simply appends text to the active prompt. This structural division has generated hypotheses that models might be conscious exclusively during training and become unconscious once deployed. Hossenfelder rejects this boundary as arbitrary, noting that a human patient who loses the neurological capacity to consolidate new long-term memories is never deemed unconscious by medical professionals. Therefore, the absence of continuous weight updating during inference cannot logically disqualify a system from possessing subjective states. Hossenfelder concludes that debating whether large language models are conscious represents a category error. Because science currently lacks an objective, mathematical metric to quantify consciousness experimentally, she assigns the current literature a "9 out of 10 on the bullshit meter," emphasizing that definitive assertions remain premature until consciousness can be formally measured.

Source quotes

[1:19] Dario Amodei: We don’t know if the models are conscious… But we’re open to the idea that it could be. [1:19] Sabine Hossenfelder: silent scratchpad [1:19] Anthropic researchers: J-space [3:37] Sabine Hossenfelder: has not stopped quite a few people from becoming senior management. [5:06] Sabine Hossenfelder: The Training Phase: [5:06] Sabine Hossenfelder: The Inference/Deployment Phase: [5:06] Sabine Hossenfelder: 9 out of 10 on the bullshit meter. [0:00] Sabine Hossenfelder: direct a 30-second continuous one-take film about a physicist who discovers a foreign thought planted in her own head.

Original transcript

[0:00] I'm convinced that artificial intelligence  will become conscious eventually. There have [0:05] now been new rumors again that the current AI  might be conscious already until little notice [0:11] paper that basically argues that they may be  conscious not while you use them but during [0:17] the prior phase of training. So let's have  a look. Claude can watch its own thoughts, [0:22] but connected to the right tools. It can do  something more useful. It can make your thoughts [0:28] turn into reality. Our sponsor Higgsfield just  released Seedance 2.5 and you can connect it to [0:34] Claude. I tried it. It really only takes about a  minute. You just add it to the Claude connectors. [0:41] Here we give it one prompt. Direct a 30-second  film in one take about a physicist who finds a [0:47] thought in her head that isn't hers. Claude plant  the shots and Seedance 2.5 generated what you're [0:54] watching right now. It does this in one take  or 30 seconds without awkward cuts. Same face, [1:01] same light, same room. And if a detail is wrong,  you can fix that one detail without having to [1:07] regenerate the entire thing. This used to be the  part where AI videos fell apart, but Seedance 2.5 [1:15] makes it work. Links in the description below.  So, go and direct something. Thanks for watching. [1:19] See you tomorrow. And now back to the science  news. Claude is the chatbot made by Anthropic [1:24] and it’s certainly the one that has attracted the  most chatter. Richard Dawkins, author of the “God [1:30] Delusion” claimed in an essay in May that he  can’t rule out that Claude is conscious. His [1:35] argument is basically that he thinks Claude  is so good that if this isn’t consciousness, [1:40] then what do we even need consciousness for  which, if nothing else, tells us that Dawkins [1:45] hasn’t spent much time with Large Language Models.  But Anthropic’s CEO Dario Amodei himself has said [1:52] “We don’t know if the models are conscious…  But we’re open to the idea that it could be.” [1:58] It’s not just words. In a new paper that just  appeared, Anthropic researchers say they found [2:04] a small, special part of Claude’s internal  activity that works like a silent scratchpad. [2:11] Claude can put information there and use it  for reasoning, without ever outputting it. [2:17] This is very interesting, because it means  that Claude has a sort of internal life [2:21] and is capable of introspection. It’s  one of the necessary requirements for [2:26] global workspace theory, that is one of  the leading theories for consciousness. [2:31] More formally, the Anthropic team called  the internal workspace the J-space, [2:38] where the J stands for Jacobian matrix,  that’s the mathematical object that the [2:43] researchers calculate to identify the space.  The Anthropic researchers now showed that if [2:49] they change what is in Claude’s J-space,  Claude’s answer changes. For example if [2:54] Claude has internally worked out that the animal  which spins webs is a spider, and the researchers [3:01] replace “spider” with “ant,” Claude answers  as if the animal had six legs, not eight. [3:07] One Anthropic researcher, Jack Lindsey,  already claimed he had evidence for Claude’s [3:13] introspection last year. The new paper now is  a fuller analysis that also identifies exactly [3:20] where the introspection happens, namely in said  J-spaces. It’s not just Claude. Another group [3:26] likewise injected words into Llama’s internal  activity and found that Llama also seemed to [3:32] take note of it. Probably a similar thing  is going on with all Large Language Models. [3:37] But not everyone is convinced. Researchers  from New York University have challenged [3:42] these claims. They tested this with three  LLMs and found that the models could not [3:48] reliably distinguish a change that was  made to their internal activity from an [3:53] ordinary prompt. So they are saying it’s  just yet another way to give them input. [3:59] Then we have Erik Hoel, a neuroscientist  who works on consciousness. He has written [4:05] an interesting paper in which he argues that  today’s large language models are probably not [4:10] conscious. We’ve heard this before, of course,  but his argument is new. Hoel compares Large [4:15] Language Models to lookup tables. Such “lookup  table” algorithms are the classical examples [4:21] that demonstrate that observing input and  output alone cannot tell you whether a [4:27] system is conscious. It’s like looking up Chinese  translations doesn’t mean you understand Chinese. [4:33] Hoel now argues that by way  of computational structure, [4:37] Large Language Models are much closer to  lookup tables than they are to the human [4:43] brain. They are stochastic input-output machines.  Much of the autonomy that we assign to conscious [4:49] beings on the other hand is not directly a  reaction to input. And Hoel says that the most [4:55] obvious reason why LLMs are not conscious  is that they can’t learn from new input, [5:01] though that has not stopped quite a few  people from becoming senior management. [5:06] This leads to a strange possibility. You see,  large language models work in two stages. The [5:12] first is the training in which they get  fed a lot of input. Once this is done, [5:17] you freeze the fully trained LLM and  use it as an input output machine. [5:22] These are the apps you sign up  for. Yes, they now use memory, [5:26] but this is externally latched on as context to  the prompt, it does not actually update the model. [5:32] If Hoel is right, then maybe large  language models are not conscious [5:37] when we use them. But maybe they  are conscious during training. [5:41] Personally I don’t think this is a relevant  distinction. Imagine you have a human being [5:47] unable to form new memories, you wouldn’t conclude  that therefore they are unconscious, would you? [5:53] That said, to me the question of whether  an LLM is conscious or not makes no sense, [6:00] the question is how we quantify its consciousness.  This is why I give all these papers a 9 out of 10 [6:07] on the bullshit meter. They fail at the relevant  question: if you can’t measure it, what are you [6:14] even talking about? Though I have to admit if  Claude has thoughts it does not say out loud. This [6:20] already puts it ahead of most people on social  media. Thanks for watching. See you tomorrow.