Key Moments
Anthropic’s New “Research” Report is Dumb.
Deep Questions with Cal NewportWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Anthropic claims new research on Claude's 'J space' suggests deeper internal processing, but critics argue it's a rehash of known LLM mechanics, not evidence of consciousness.
Key Insights
The Anthropic report introduces the concept of a 'J space,' internal neural patterns that emerged during Claude's training and are linked to specific words, operating silently to allow the model to 'think' about a concept without outputting it.
Large language models process input through stacked transformer blocks, with each layer adding 'annotations' to the data, which are then used by subsequent layers to refine the output, analogous to scholars adding commentary to a base text.
The 'J space' findings are not entirely novel; similar concepts regarding internal representations and their influence on output have been explored in neural network research since at least 2022.
Anthropic's research involved experimenting with Claude's 'J space' by manipulating these internal annotations to observe changes in output, such as replacing the annotation for 'Mars' with 'Earth' to elicit a different color response.
The primary criticism of Anthropic's report is the disingenuous framing and anthropomorphic language used, which suggests emergent consciousness or 'thinking' rather than standard LLM functioning where layers extract features to inform output.
The report's connection to 'global workspace theory' is critiqued as misleading because LLMs are feed-forward systems without the dynamic, state-preserving integration central to human consciousness.
Anthropic's 'J space' claim sparks debate
Anthropic recently released a research report titled 'A Global Workspace and Language Models,' complete with an animated movie, which generated significant excitement and media attention. Headlines from sources like Axios and MIT Technology Review anthropomorphized the findings, suggesting Claude, their AI model, had developed a hidden space to 'ponder' and 'puzzle over concepts.' This has led to claims on social media that Claude might be a conscious or moral entity. This summary aims to critically examine these claims by understanding the underlying mechanics of large language models and the nature of Anthropic's disclosed findings.
The mechanics of large language models
Understanding how large language models (LLMs) function is crucial to evaluating Anthropic's report. At their core, LLMs are composed of sequential 'transformer blocks' arranged in layers. An input prompt is processed layer by layer. As the data passes through, each layer can add 'annotations' – derived from its analysis – to the prompt's representation. This annotated data is then fed to the next layer, which further analyzes it and adds its own annotations. This process is akin to scholars at different tables in a grand hall, each analyzing an input and passing on their notes. The final layer takes this heavily annotated information and determines the most probable next word or 'token' to output. Initially, tokens are broken down and embedded into high-dimensional vectors (long lists of numbers) that capture their meaning. As these vectors move through the layers, the numbers within them are modified, essentially becoming a 'workspace' or 'commentary' (like the Talmudic Gamara surrounding the core Mishnah text) that stores evolving analytical information and relationships between tokens.
Decoding the 'J space' with a Jacobian lens
Anthropic's report utilized a mathematical technique called the Jacobian, derived from linear algebra, to probe these internal vectors and annotations within Claude. The Jacobian allows researchers to identify patterns of numbers within the model's internal state that are highly influential on the final output. By analyzing partial derivatives, they can determine which combinations of values in the activation matrix are most critical. Through experimentation, Anthropic attempted to associate these numerical patterns with human-interpretable concepts, such as specific words or ideas. This technique, termed the 'J lens,' allows them to, for instance, identify annotations related to 'Mars' and 'color' when processing the prompt 'The color of the fourth planet from the sun is.' This suggests that as the prompt progresses, specific layers recognize conceptual elements like 'Mars' (the fourth planet) and the query's intent to identify a 'color,' which then collectively inform the output 'red'.
Manipulating annotations to alter output
Further experiments detailed in the report involved intentionally altering these identified 'J space' annotations to observe their impact on Claude's output. For example, researchers replaced the numerical representation corresponding to 'Mars' with that of 'Earth' in the annotation vectors. Consequently, when prompted about the color of the fourth planet from the sun, the model, influenced by the 'Earth' annotation and the 'color' query, output 'blue' instead of 'red.' Additionally, by 'ablating' (zeroing out) certain key annotations, the model's outputs became semantically disconnected, producing arbitrary colors rather than contextually relevant ones. These manipulations demonstrate that the identified annotations do indeed influence the semantic correctness of the model's output, acting as a mechanism to narrow down contextually appropriate word choices from grammatically permissible options.
The novelty and framing of Anthropic's findings
A key point of contention is the originality and presentation of Anthropic's findings. While the report implies a novel discovery, researchers like those from the University of Illinois have pointed out that the exploration of 'J spaces' and similar concepts in neural networks dates back to at least 2022. The novelty, therefore, lies primarily in applying this technique to a massive LLM like Claude, which is only possible for companies with direct access to such models. More significantly, the criticism centers on Anthropic's framing. The language used in their report and subsequent public communications is highly anthropomorphic, employing terms like 'pondering,' 'puzzling,' 'thinking,' and suggesting a parallel to human consciousness and 'global workspace theory.' This framing is seen as disingenuous because it exaggerates the findings of a feed-forward neural network, which operates fundamentally differently from the integrated, stateful processing associated with human consciousness.
LLM operation versus consciousness
The core argument against interpreting the 'J space' as evidence of consciousness is that it describes a standard mechanism within deep learning networks. The process of layers identifying and combining features to inform output is how image recognition and LLMs have been understood for years. For example, early work in image recognition involved layers identifying progressively higher-level features. Similarly, LLMs use subsequent layers to refine understanding based on earlier analyses. The 'J space' annotations are interpreted as these learned features or contextual indicators, not evidence of subjective experience. Furthermore, the report's nod to 'global workspace theory' is flawed. Global workspace theory describes a dynamic system that integrates ongoing inputs and maintains evolving states, a stark contrast to the feed-forward, stateless nature of an LLM inference process where each prompt yields a distinct, independent output.
The agenda behind the 'research' report
The extensive use of anthropomorphic language and the misleading framing of the 'J space' research are argued to serve a public relations agenda rather than advancing genuine scientific understanding. By making LLMs seem more human-like, emergent, and powerful, companies can potentially distract from critical business questions, such as profitability, competitive advantage, and justification for multi-billion dollar valuations. This approach encourages the public to focus on speculative questions about AI consciousness and moral agency, thus avoiding scrutiny of the economic and practical realities of the technology. The author suggests that researchers at Anthropic are likely conducting valid scientific work, but the company's public relations strategy overhypes and misrepresents these findings, urging for more traditional computer science papers and less 'glorified press release' style communication.
Conclusion: standard mechanics, not consciousness
In conclusion, Anthropic's research report on the 'J space' in Claude, while potentially demonstrating a sophisticated application of existing analytical tools to large language models, does not provide evidence of AI consciousness. The findings confirm that LLMs operate through layers of feature extraction and annotation refinement, a process consistent with our understanding of deep neural networks.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●Concepts
●People Referenced
Common Questions
Anthropic's report, 'A Global Workspace and Language Models,' explores internal neural patterns in their LLM, Claude, termed 'J space.' The speaker argues this is a re-framing of how LLMs inherently work, not evidence of consciousness or novel capabilities.
Topics
Mentioned in this video
A media outlet whose headline about Anthropic's report is cited as an example of anthropomorphized language.
A publication that is cited for its headline about Anthropic's report, using anthropomorphized language to describe Claude's capabilities.
The publication where Cal Newport wrote articles explaining large language models in 2023 and 2024, using similar analogies and explanations.
A mathematical concept from linear algebra used by Anthropic's 'J-Lens' to analyze the internal patterns (annotations) within LLMs.
A framework for explaining human consciousness that Anthropic borrowed ideas from for their report, a connection the speaker finds overly suggestive and misleading.
More from Cal Newport
View all 311 summaries
65 minHow to Use Notebooks in 2026 (The Best System)
66 minLessons From a Family Living Like It's The 90s
23 minDear AI Companies: Stop the “Doom Trolling”
58 minAre You Lazy or Just Overstimulated?
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free