Anthropic’s New “Research” Report is Dumb.  🔗 **Source:** [Anthropic’s New “Research” Report is Dumb.](https://www.youtube.com/watch?v=m6ucYXJyjAk) > [!note]+ Document Notes > tags: [ai-interpretability, ai-anthropomorphism, mechanistic-interpretability, media-literacy, ai-research-critique] > author: Cal Newport > source: Cal Newport (YouTube) > reference: https://www.youtube.com/watch?v=m6ucYXJyjAk > date created: 2026-07-16 12:30 > > # 🏷️ Title: Anthropic's New "Research" Report is Dumb. > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - **What happened**: Anthropic published a research report, "On the Biology of a Global Workspace" (referred to in the video as "A Global Workspace and Language Models"), describing a technique — the "J-lens," built on a Jacobian linear-algebra tool — that decodes human-interpretable concepts (a "J space") out of Claude's internal activation vectors. The report shipped with an animated video and language like Claude "pondering" and "puzzling" over concepts silently, "on its own," without anyone programming it. Tech press (Axios, MIT Technology Review) and X posts amplified this into claims that Anthropic had found evidence of consciousness or moral patienthood in Claude. - **Why it matters**: Newport argues the underlying finding — that intermediate layers accumulate interpretable "annotations" that narrow down semantically appropriate output tokens — is not new. It matches the standard account of how transformer-based LLMs and deep neural networks generally have been understood to work for years (he cites his own 2023/2024 New Yorker explainers), and comparable interpretability techniques predate this paper (he cites a University of Illinois professor's tweet listing related work since 2022). His core claim is that Anthropic's framing — hedged just enough to avoid an explicit false claim, while using loaded language ("thinking," "on its own," global-workspace-theory comparisons) — manufactures a sense of novelty and "digital ick" that the actual result doesn't support. - **Key stakeholders**: Anthropic's research and PR/communications functions, AI journalists and outlets that amplified the framing (Axios, MIT Technology Review), the broader public forming beliefs about AI sentience/consciousness, and Anthropic's investors/business stakeholders (Newport's closing argument ties the framing to distraction from harder questions about unit economics and competitive moat). - **Risks and tradeoffs**: Overclaiming or ambiguous framing of interpretability research risks (a) misinforming the public about AI consciousness/moral status, a genuinely consequential question that deserves rigor rather than marketing lift; (b) crowding out scrutiny of Anthropic's business fundamentals (token costs, competitive moat, path to justifying valuation); and (c) devaluing real interpretability research by association with hype cycles, making it harder for the public to distinguish legitimate advances from press-release science. - **Bottom line**: Newport's assessment is that the research itself is competent, incremental, non-novel work executed at a scale (a frontier proprietary model) that outside researchers can't replicate — but the communication of it is, in his word, "disingenuous," designed to generate attention-driving anthropomorphic reactions rather than to inform. ## 📖 Main Content Newport structures the video as a teardown in three movements: (1) read Anthropic's own framing and the media/social reaction to it, (2) teach the actual mechanics of how transformer-based LLMs process a prompt, and (3) return to the paper to show that its findings map exactly onto that mechanical account — meaning nothing "unnerving, cool, or blockbuster" was revealed. **The paper's claim, in Anthropic's words** (as read on-screen): Claude has a small collection of internal neural patterns — the "J space," named for the Jacobian technique used to find them — that "play a special role" relative to other internal processing. Each pattern links to a word, but lighting up doesn't mean the model says that word, only that "the word is on its mind." Unlike a visible scratchpad or chain-of-thought, the J space "operates silently in the model's internal neural activations, allowing the model to think about a concept without writing it down," and — Anthropic's own italics — it "wasn't designed or programmed by us, but instead emerged on its own during Claude's training process." **The tutorial (mechanics of an LLM)**: Newport walks through transformer blocks arranged in sequential layers, each of which can add "annotations" to a token's underlying vector of numbers as the prompt passes through. He offers two analogies: a relay of scholars at successive tables, each reading prior annotations and adding their own; and the layout of a page of Talmud, where a small central text (the token) is surrounded by accumulating layers of commentary. Tokens are first embedded as long vectors of numbers capturing rough meaning; self-attention and feed-forward sublayers within each transformer block then update those numbers, layer by layer, with the final layer mapping the fully annotated result to an output token. He separates two things that are actually entangled in training: **syntactic validity** (what words could grammatically follow — learnable even with a simple n-gram model, likely concentrated in later layers) and **semantic appropriateness** (which of the syntactically valid options actually fits the prompt's meaning — where he argues the accumulated annotations do their work). **Back to the paper, now legible**: Anthropic's Jacobian-based "J-lens" decodes those annotation vectors into human-interpretable labels. Their worked example: for the prompt "the color of the fourth planet from the sun is," the technique surfaces annotations corresponding to "Mars" and "color" mid-network; the final layer combines them to output "red." They validate this causally two ways — replacing the "Mars" annotation values with "Earth" values changes the output to "blue," and "ablating" (zeroing out) the Mars-related pattern produces grammatically valid but semantically arbitrary color outputs. Newport calls this legitimately interesting experimental work, while objecting to Anthropic's use of "ablated," a term borrowed from neuroscience/animal research, as anthropomorphizing a numerical zeroing operation. **Newport's two-part verdict**: First, novelty — he cites a tweet from University of Illinois professor Yu Zhaozhang noting comparable Jacobian-based interpretability work dating to 2022, framing Anthropic's contribution as the first application at the scale of a frontier proprietary model (which outside researchers structurally cannot replicate, since only the lab has model internals) rather than a new idea. Second, and central to the video's title: the *language* used to describe the results. He argues that once you understand the layer-by-layer annotation mechanism, the J-lens findings are "exactly how we've always understood large language models to work" — not just LLMs specifically, but deep networks generally, tracing back to early image-recognition work. He singles out specific phrases as misleading: "thinking about a concept without writing it down" (his rebuttal: updating the activation matrix *is* how it writes things down — that matrix is its only workspace) and "Claude did this on its own, we didn't program it" (his rebuttal: that describes literally all machine learning — nothing is programmed in a trained neural network by definition, so stating it as a distinguishing fact about this system is misleading through omission, not falsehood). **The consciousness section**: Anthropic connects J-space experiments to Global Workspace Theory (GWT), a framework for conscious access in humans and animals, while explicitly hedging that "our experiments don't show that Claude can have experiences or feel things." Newport's strongest technical objection: GWT describes an ongoing, stateful system that integrates continuously arriving inputs and selects among them over time. A transformer forward pass is the opposite — strictly feedforward, one layer to the next, no persisted state between prompts, "and then it's done, and the token is out." He treats this as a structural mismatch the GWT comparison glosses over rather than confronts. **The closing argument (motive)**: Newport's final move shifts from technical critique to an inference about incentives — that anthropomorphic, "digital ick"-inducing framing distracts audiences from harder, more consequential questions about Anthropic's business (token costs, thin competitive moat against smaller specialized models, and what justifies a trillion-dollar valuation). He closes by distinguishing the researchers ("good work") from what he calls the PR department's packaging of it, and calls on Anthropic to publish real computer science papers rather than "glorified press releases." **Terms defined**: J space / J-lens (Anthropic's Jacobian-based interpretability method and the resulting decoded activation patterns); "annotations" (Newport's own explanatory term for information layers add to a token's activation vector as it passes through a network — not Anthropic's term); ablation (in this context, zeroing out specific activation values to test their causal role); Global Workspace Theory (a cognitive-science framework for how a subset of information becomes "consciously" available across a system, here disputed as inapplicable to feedforward inference). **Uncertainty / contested claims**: This analysis is built from the video's auto-generated YouTube captions; a small number of caption artifacts (word-order glitches, filler-word transcription like "It's putting") were present in the raw transcript and are called out inline where a quote spans one. Anthropic's original paper/report was not independently reviewed for this note — all characterizations of "what Anthropic said" reflect Newport's on-screen reading and paraphrase of it, not a separate verification against Anthropic's published text. Newport's inference about Anthropic's *motive* (distraction from business fundamentals) is presented by him as opinion ("that's just my opinion") and is treated here as an argued but unverified claim, not an established fact. ### 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=0) The report and the online reaction - Anthropic's "A Global Workspace and Language Models" report shipped with an animated video. - An X post read aloud claims: "Claude, my friends, by all counts, is a conscious entity. Claude, my dear friends, is a moral patient." - Axios: "Anthropic says Claude has carved out its own space to ponder." MIT Technology Review: "Anthropic found a hidden space where Claude puzzles over concepts." - Newport frames the episode's question: genuine evidence of something new, or "another over-wrought cynical push to generate a fresh wave of relevance-reinforcing digital ick"? #### ⏱️ [00:02:06](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=126) Anthropic's own framing, read verbatim - J space: "a small collection of internal neural patterns that... play a special role." - Explicitly contrasted with visible chain-of-thought: it "operates silently," lets the model "think about a concept without writing it down." - Anthropic's italics, per Newport: the J space "wasn't designed or programmed by us, but instead emerged on its own." #### ⏱️ [00:03:34](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=214) Tutorial: transformer layers as accumulating "annotations" - A prompt passes through sequential transformer-block layers; each layer can add information ("annotations") on top of what prior layers added. - Two analogies: relay tables of scholars adding commentary in sequence; a Talmud page, with the token as the small central text and layered commentary (the Gemara) built up around it by each layer. - The final layer maps the fully annotated representation to an output token (technically a probability distribution the control program samples from). #### ⏱️ [00:06:49](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=409) Tutorial: tokens, embeddings, and the annotation matrix - Prompts are broken into tokens; each token is embedded as a long vector of numbers initially capturing rough meaning. - As the matrix of token-vectors moves layer to layer, self-attention and feed-forward sublayers update those numbers — that update *is* where annotations "live." - "This big... table of numbers, we call this a matrix" — the same matrix Newport later calls the model's only workspace. #### ⏱️ [00:11:23](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=683) Syntactic vs. semantic contribution to next-token choice - Two entangled but conceptually separable jobs: which next tokens are grammatically valid (easy, likely concentrated in final layers, learnable even by a simple n-gram model), and which of those is semantically appropriate given the prompt's meaning (where accumulated annotations do the work). - Example: "the quick brown fox ___" — many grammatically valid continuations (jumped, danced, died, talked), but annotation of "this is a known saying" privileges "jumped." #### ⏱️ [00:13:52](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=832) Back to the paper: the Jacobian/J-lens technique - Jacobian-based linear algebra is used to find which patterns of values in the activation matrix are most causally influential on the final output. - Worked example: prompt "the color of the fourth planet from the sun is" surfaces mid-network annotations decodable as "Mars" and "color"; combined at the final layer, the model outputs "red." - Causal test: swapping the "Mars" pattern for the "Earth" pattern changes the output to "blue," holding the rest of the prompt fixed. #### ⏱️ [00:19:03](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1143) The ablation experiment - Zeroing out ("ablating") the Mars-related annotation still produces grammatically valid output, but the color is no longer semantically tied to Mars — arbitrary colors result. - Newport flags "ablated" itself as an anthropomorphizing word choice, borrowed from neuroscience, for what is a numerical zeroing operation. - His assessment: legitimately cool experimental design, showing the annotations are causally load-bearing for output. #### ⏱️ [00:20:45](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1245) Take 1 — the technique isn't novel - Cites a tweet from University of Illinois professor Yu Zhaozhang: comparable Jacobian-based interpretability work traced to 2022, with several related papers listed. - Anthropic's distinct contribution, per Newport, is applying it at the scale of a frontier proprietary model — which outside researchers can't replicate, since only the lab has access to those internals — not a new underlying idea. #### ⏱️ [00:22:05](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1325) Take 2 — "incredibly disingenuous" framing - Core thesis stated directly: "the way that Anthropic described these results and the way that they were echoed after Anthropic described them, I think is incredibly disingenuous." - Newport ties this to his own 2023 and 2024 New Yorker explainers, which used the same "annotations that build on annotations" framing years before this report — this is not new territory. - Generalizes the point to all deep learning: "This is just how we understand large language models to work. It's also how we understand essentially any deep learning neural network to work," tracing back to early image-recognition (Yann LeCun-era) work. #### ⏱️ [00:25:36](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1536) Line-by-line critique: "thinking... without writing it down" - Quotes the paper's line about the J space operating silently, letting the model "think about a concept without writing it down." - Direct rebuttal: "It's not thinking... It is writing it down... It has this matrix that it updates the numbers every time. That's its workspace. That's its work pad." - Frames this as the paper describing ordinary deep-network mechanics in language engineered to sound mysterious. #### ⏱️ [00:26:25](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1585) The consciousness / Global Workspace Theory section - Anthropic's hedge, read on screen: "our experiments don't show that Claude can have experiences or feel things" — followed, per Newport, by language he characterizes as "all very suggestive that there's something new going on here." - Newport's structural objection: GWT describes an ongoing, stateful system integrating continuously arriving inputs and selecting among them; a transformer inference pass is feedforward, layer to layer, with no persisted state — "and then it's done, and the token is out. Nothing is saved. There's no states that are changing." - Separately notes GWT itself remains a contested framework even for human consciousness — a further layer of unsettled ground under the comparison. #### ⏱️ [00:27:03](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1623) "Claude did this on its own" — true of all machine learning - Targets the paper's framing that the J space "emerged on its own" rather than being programmed. - Rebuttal: this describes every trained neural network by definition — nothing in machine learning is hand-programmed; weights start random and are learned through training until loss is minimized. Stating it as if distinctive to Claude implies something it doesn't establish. #### ⏱️ [00:29:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1740) The real stakes: distraction from business questions - Newport's inference about motive: anthropomorphic framing generates "a general sense of ick" that makes the technology feel powerful and alive, which crowds out harder questions. - Named examples of the questions he thinks get skipped: Anthropic's per-token costs, thin competitive moat against smaller specialized/hand-harnessed models, and what justifies a trillion-dollar valuation. - Explicitly flagged by Newport as inference/opinion rather than a proven claim about intent. #### ⏱️ [00:30:04](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1804) Conclusion and call to action - Distinguishes the researchers ("I like your work and this is good work") from the communications packaging around it. - Direct ask: "Anthropic, will you stop with the... press release research reports? Write computer science papers, sell products, convince us to buy the products." - Final verdict on the episode's framing question: no blockbuster reveal — "we confirmed that yes, language models work the way that at least I always thought they did." ## 📚 References - Anthropic — research report titled (per on-screen reading) "A Global Workspace and Language Models," accompanied by an animated explainer video; specific URL not stated in the transcript. - Axios — headline "Anthropic says Claude has carved out its own space to ponder"; specific article URL not stated in the transcript. - MIT Technology Review — headline "Anthropic found a hidden space where Claude puzzles over concepts"; specific article URL not stated in the transcript. - Yu Zhaozhang (University of Illinois professor) — X/Twitter post citing Jacobian-based interpretability work dating to 2022, listing several related papers; specific post URL not stated in the transcript. - Cal Newport, The New Yorker — prior long-form explainers on LLM architecture (cited by Newport as dating to 2023 and 2024) using the same layered-annotation framing this report is compared against; specific article titles/URLs not stated in the transcript. - Yann LeCun / early deep-learning image-recognition work — referenced generally as the origin of the "layers pick up different features" account of deep networks; no specific paper cited in the transcript. ## 🗣️ Paraphrase Anthropic published an interpretability report showing that Claude's internal activations contain decodable, human-interpretable patterns — a "J space" — that causally shape which word comes out at the end of a prompt. Their own worked example: feed in "the color of the fourth planet from the sun is," and mid-network you can find a pattern that decodes to "Mars" and another that decodes to "color"; swap the Mars pattern for an Earth pattern and the model now says "blue" instead of "red." That's genuinely neat, causally-grounded interpretability work. But Newport's issue isn't the finding — it's the packaging. Once you understand, in ordinary terms, how transformer layers work (each layer adds interpretable "annotations" to a token's representation, later layers narrow down semantically appropriate output using those annotations), Anthropic's result is exactly what that mechanical picture predicts. Nothing about it requires or supports language like "thinking silently," "on its own," or comparisons to a framework for human consciousness. His sharpest technical objection targets the consciousness comparison specifically: Global Workspace Theory describes an ongoing system that integrates a continuous stream of inputs over time, while a transformer forward pass is a single, stateless, feedforward computation that starts fresh with every prompt. Calling attention to that mismatch is a structural argument, not just a stylistic complaint. His closing argument is more speculative: that this kind of anthropomorphic framing serves Anthropic's interests by generating attention and unease that displace scrutiny of harder business questions — token costs, competitive moat, and what justifies the company's valuation. He's explicit that this last part is his opinion rather than a demonstrated fact, while the technical critique of the paper's language is argued in more detail and grounded in the mechanics he spent the first third of the video teaching. ## 📰 Newsletter **A researcher's rebuttal to "Claude is conscious" headlines — worth the 31 minutes.** Anthropic's newest interpretability report ("A Global Workspace and Language Models") set off a familiar cycle: an animated explainer video, breathless X posts about Claude being a "moral patient," and headlines like MIT Technology Review's "Anthropic found a hidden space where Claude puzzles over concepts." Cal Newport's response is a genuinely useful teardown — not a dismissal of the research (he calls it "good work"), but a line-by-line challenge to the language used to describe it. He spends the first third of the video teaching, in plain terms, how transformer layers actually work — annotations accumulating layer by layer, narrowing down a semantically appropriate next word. Then he shows that Anthropic's headline finding maps exactly onto that mechanical picture, and that comparable interpretability techniques predate this paper by several years. His sharpest point: comparing the result to Global Workspace Theory (a framework for human consciousness) glosses over the fact that transformer inference is stateless and feedforward, while GWT describes an ongoing, integrating system — a structural mismatch, not just a stretch. Useful watch for anyone who writes, reviews, or shares AI research communications — a good case study in reading past a press release's framing to what the underlying result actually shows. ## ✅ Action Items #### DevRel - Developer advocacy: build a short internal reference distinguishing "interpretability finding" from "consciousness claim" for use when developers or community members ask about AI-lab research announcements. - Education: use this video's transformer-layer explainer as a citable plain-language teaching resource when introducing how LLMs process prompts. - Developer experience: none directly applicable — this source concerns research communications, not tooling. - Developer marketing: when citing AI-lab research in marketing copy, default to the paper's own hedged language rather than the press coverage's amplified framing. - Community management: watch for community members repeating "Claude is conscious" style claims sourced from press coverage rather than the primary report, and be ready to point to a plain-language rebuttal. - Internal/external/cross-project: flag this as a recurring pattern (AI-lab research report → animated video → anthropomorphized press coverage) worth a standing media-literacy note rather than a one-off response. #### OSPO - Policy & Governance: not directly addressed in this source — no governance mechanism proposed or critiqued. - License Compliance: not addressed in this source. - Security Management: not addressed in this source. - Software Inventory: not addressed in this source. - Developer Support: not addressed in this source. - Community Engagement: not addressed in this source. - Strategic Alignment: relevant as a cautionary case study when evaluating vendor (AI-lab) claims against roadmap decisions — distinguish PR framing from the technical report before acting on a vendor's stated capability claims. #### Community - Maintainers: not addressed in this source. - Contributors: not addressed in this source. - Sponsors: not addressed in this source. - Ecosystem: worth sharing as a model for how to publicly critique a research communication on its technical merits (Newport separates "good research" from "disingenuous framing" rather than attacking the researchers) — a useful tone reference for any ecosystem disagreement about vendor claims. #### WordPress - Core AI: directly relevant — any WordPress Core AI messaging about AI-assisted features should avoid the same anthropomorphizing pattern Newport critiques (e.g., avoid "the AI thinks about" language for deterministic or well-understood mechanisms). - Training: this video is a usable teaching example for a "how to read an AI research announcement critically" module. - Marketing: same caution as Core AI — avoid mystifying language about AI features in WordPress-facing marketing copy. - All other subteams: not addressed in this source. #### The WPCC - Not addressed in this source. ### FAIR - Technical independence: relevant tangentially — Newport's point that only Anthropic could apply this technique to Claude (because only they have model internals) is itself an argument for open-weight models as a check on interpretability claims from closed labs. - Policies and covenants: not addressed in this source. - Productized governance artifacts: a short "how to evaluate an AI-lab research announcement" checklist (distinguish paper language from press coverage; check for prior art; separate causal claims from suggestive juxtaposition) would be a natural artifact derived from this video. ## 🧪 Hypotheses 1. Research communications from AI labs that use anthropomorphic language (e.g., "thinks," "on its own," "puzzles over") generate measurably higher social media engagement than functionally equivalent communications using mechanistic language, controlling for underlying finding. 2. Readers given a plain-language transformer-architecture explainer before reading an AI-lab interpretability report rate the report's findings as less novel and less "conscious-like" than readers who read the report without that context. 3. The Jacobian-based interpretability technique described in this report, when independently reproduced on open-weight models, will show comparable annotation/decoding patterns to those Anthropic reports for Claude — supporting Newport's claim that this is a general transformer property rather than something specific or newly discovered in Claude. 4. Volume of public discourse invoking "AI consciousness" correlates more strongly with the presence of hedged-but-suggestive language and accompanying illustrative media (animated videos, loaded terminology) in an AI lab's own communications than with the magnitude or novelty of the underlying technical finding. 5. Global Workspace Theory, as standardly defined in cognitive science, requires persistent cross-input state integration that a strictly feedforward transformer inference pass structurally cannot exhibit — meaning any GWT-based consciousness claim about current-generation feedforward LLM inference is, by definition, unsupported regardless of interpretability findings about single-pass activations. ### 📊 Methodology - **H1**: A/B test identical technical findings summarized in anthropomorphic vs. mechanistic language, distributed to matched audiences; measure shares, comments, and self-reported perception of "aliveness" of the described system. - **H2**: Randomized controlled reading study — one group reads a plain-language LLM architecture primer before the Anthropic report, a control group reads the report cold; both groups rate novelty and human-likeness of the findings on a standardized scale. - **H3**: Independent replication of the Jacobian/J-lens technique on an open-weight model (e.g., a Llama or Mistral variant) using the same Mars/color-style causal-intervention protocol; compare decoded annotation patterns and ablation effects to Anthropic's reported results. - **H4**: Content analysis of AI-lab research communications over a defined period, coding for anthropomorphic language density and accompanying media (video, animation), correlated against social-media engagement and volume of "consciousness"-framed coverage. - **H5**: Formal comparison of Global Workspace Theory's defining criteria (persistent state, competitive selection among concurrent inputs, broadcast/integration across subsystems over time) against the architectural properties of feedforward transformer inference (statelessness between prompts, single-pass computation) to establish whether the comparison is falsifiable in principle or a category mismatch by definition. ### ❓ Questions - Does Anthropic's original written report include more or less hedging than the accompanying animated video and social/PR amplification — i.e., where exactly does the anthropomorphizing enter the communication chain? - How would Newport's critique change, if at all, for a multi-turn or agentic system that does carry persistent state across a session (as opposed to the single feedforward pass he describes)? - What would a genuinely novel (not previously demonstrated at any scale) interpretability finding about LLM internals look like, and would Newport's own framework recognize it as such? - Is there a version of "interpretability progress at frontier model scale" that's worth reporting on its own terms (bigger model, more layers, higher-dimensional activations) without needing consciousness-adjacent framing to generate interest? - To what extent do AI labs' research communications teams operate independently of the research teams, and does that organizational split explain the gap Newport identifies between "good research" and "disingenuous framing"? ### 🔮 Future Research - A systematic comparison of interpretability terminology across AI labs (Anthropic, OpenAI, Google DeepMind, Meta) to see whether anthropomorphizing language is a house style specific to one lab or an industry-wide pattern. - A history of the specific Jacobian-based interpretability technique from its 2022-era origins (per the Yu Zhaozhang citation) through to this report, to establish precisely what, if anything, is new at the implementation or scale level. - Independent, peer-reviewed evaluation of whether Global Workspace Theory can be meaningfully operationalized for any current AI architecture, feedforward or otherwise — separate from any single lab's marketing framing. - A media-literacy study tracking how a hedged technical claim ("our experiments don't show X, but...") propagates through tech press into unhedged public claims ("Anthropic admitted X"). ### 💡 Implications **Practice**: Communicators, journalists, and developer-relations professionals writing about AI-lab research should default to checking the underlying mechanism (what architecture, what's actually being measured) before amplifying framing language from a press release or accompanying video. **Policy**: As public and regulatory interest in AI consciousness/moral-status questions grows, this video is a case study in why source discipline matters — hedged claims by an interested party (the lab itself) shouldn't set the terms of a genuinely open scientific and ethical question. **Security**: Not directly addressed in this source — no security vulnerability or concrete attack vector is described. **Education**: The video doubles as a teaching resource for transformer mechanics (layers, tokens, embeddings, attention) delivered in accessible analogy-driven language, useful independent of the critique it's embedded in. **Funding models**: Newport's closing argument implies a funding-scrutiny implication directly — that anthropomorphic hype cycles can substitute for scrutiny of unit economics (token costs, competitive moat, valuation justification) when evaluating heavily funded AI labs. ### 🧠 Learning Styles Assessment **Reading level estimate**: Approximate Flesch-Kincaid Grade ~9-10 (estimated manually from transcript sentence length and technical vocabulary density; conversational spoken register with frequent self-correction and long, clause-heavy sentences typical of unscripted podcast delivery, offset by deliberately plain-language analogies for technical concepts — no automated readability tool was run against the transcript). #### 👀 Visual learners - **Present**: Newport shares his screen to show the actual paper, the J-lens example table, and the specific quoted tweets/headlines, giving visual grounding to claims made verbally. - **Improve**: A side-by-side diagram contrasting "feedforward, stateless, layer-by-layer" (transformer inference) against "ongoing, stateful, integrating" (Global Workspace Theory) would make his strongest technical point far more scannable than prose description alone. #### 🔊 Aural learners - **Present**: Long-form, single-speaker, conversational narration is the native format of this podcast and well suited to audio-only consumption. - **Improve**: A tighter 3-5 minute audio pull-quote of just the GWT mismatch argument and the closing "write real papers" ask would work as a standalone highlight for listeners who don't have 31 minutes. #### 🗣️ Verbal learners - **Present**: Newport explicitly defines terms as he introduces them (tokens, embeddings, annotations, ablation) and offers two distinct analogies (scholars at tables; Talmud commentary) for the same underlying mechanism. - **Improve**: A short glossary card (J space, J-lens, ablation, Global Workspace Theory, feedforward) would help viewers who want to look up a specific term without re-watching the tutorial section. #### 🧩 Kinesthetic learners - **Present**: The Mars/Earth/blue causal-intervention example is concrete enough that a viewer could conceptually "try it" by imagining swapping the annotation. - **Improve**: A companion notebook or interactive demo reproducing the swap on an open-weight model would let viewers verify the mechanism hands-on rather than take the paper's word for it. #### 🧠 Logical learners - **Present**: The video is structured as an explicit two-part argument (novelty claim, then framing claim), each supported by named evidence (a prior-art tweet; a term-by-term critique of the paper's language). - **Improve**: A simple table mapping "Anthropic's phrase" → "what it technically means" → "why the phrasing is misleading" would sharpen the already-logical structure into something scannable. #### 🧑‍🤝‍🧑 Social learners - **Present**: The video opens by engaging directly with a real X post and named outlets (Axios, MIT Technology Review), situating the critique within an actual public conversation. - **Improve**: Newport's own newsletter/community (Deep Questions listeners) is the implied discussion venue, but the video itself doesn't include a structured prompt for listener response or debate. #### 🧘‍♀️ Solitary learners - **Present**: Self-contained single-episode argument that a solo viewer can follow start to finish without external context. - **Improve**: A short self-check ("can you now explain in your own words why 'thinking without writing it down' is a misleading phrase?") would suit independent, reflective viewers. #### 🌿 Naturalistic learners - **Present**: Not addressed in this source — no ecosystem or environmental framing present. - **Improve**: Not applicable; forcing an ecological analogy onto an interpretability-research critique would add noise rather than clarity. ### 📊 Accessibility Assessment #### Screen reader semantics - **Present**: Not directly assessable from a transcript-only review; hosted on standard YouTube, which provides basic player semantics. - **Improve**: If republished as an article, use real headings for each argument beat (paper's claim → tutorial → paper revisited → verdict) rather than relying on narration structure alone. #### Keyboard navigation - **Present**: Standard YouTube player keyboard controls apply. - **Improve**: Not applicable beyond standard player behavior — no custom interactive elements in this source. #### Contrast and readability - **Present**: Not assessable from transcript alone (no visual/screenshot review performed). - **Improve**: Screen-shared paper excerpts should use adequately large, high-contrast text given Newport reads from the paper on-screen at several points. #### Text resizing and reflow - **Present**: Standard YouTube player behavior. - **Improve**: Not applicable — no custom UI to evaluate. #### Media and captions - **Present**: Auto-generated English captions exist and were used as the source for this analysis. - **Improve**: Auto-captions contain occasional disfluency artifacts (e.g., a stray "It's putting" mid-sentence, "feed for Remember" instead of a clean "feed-forward. Remember...") around technically precise moments — manual caption correction would help viewers relying on captions for exact wording of the technical rebuttal. #### Forms and interactions - **Present**: Not applicable — no forms or interactive elements in this source. - **Improve**: Not applicable. #### Standards reference - No WCAG 2.2 AA gaps can be confirmed without a visual review of the actual video; the one concrete, verifiable gap from this transcript-only analysis is caption precision on technical terminology during the paper-reading segments (relates to WCAG 1.2.2 Captions). ### 🌍 DEIB Assessment #### Cultural breadth - **Present**: Single speaker, US-centric framing (Axios, MIT Technology Review, X/Twitter as the cited media ecosystem); one cross-cultural reference point (the Talmud/Gemara analogy) used respectfully and accurately to explain a technical concept. - **Improve**: Additional international AI-research or AI-policy voices on the same interpretability-vs-hype question would broaden the perspective beyond US tech media and US-based academic citation (Yu Zhaozhang, University of Illinois). #### Inclusive language - **Present**: Technical, gender-neutral language throughout; the Talmud analogy is introduced with accurate framing (correctly names the Mishnah and Gemara) rather than as a casual or reductive reference. - **Improve**: A few dense technical terms (Jacobian, ablation, feedforward) are defined in-line but assume some comfort with linear algebra vocabulary; a plain-language gloss alongside first use would widen accessibility for non-technical viewers. #### Neurodiversity support - **Present**: Clear two-part argument structure (novelty, then framing) with explicit signposting ("idea number one," "idea number two") helps viewers track the argument. - **Improve**: The video runs long (31:42) with a lengthy embedded tutorial before the critique resumes; chapter markers or a written companion summary would help viewers who benefit from non-linear or skimmable access to the argument. #### Representation - **Present**: One male presenter; quoted parties include an unnamed X poster, two publications, and one named academic (Yu Zhaozhang). - **Improve**: Not enough content in a single-speaker critique video to meaningfully assess representation; would require the wider Deep Questions catalog for a fair evaluation. #### Psychological safety - **Present**: Newport is careful to separate criticism of the *communication* from criticism of the *researchers*, explicitly praising the underlying work ("I like your work and this is good work") rather than attacking individuals — a model of substantive disagreement without personal targeting. - **Improve**: The closing section's inference about Anthropic's motive (distraction from business scrutiny) is more speculative than the technical critique and is framed with confident rhetoric ("I do not like the way Anthropic talks about the research") that could read as more certain than the evidence supports; explicitly labeling that section as opinion earlier (rather than only at the very end) would improve safety/fairness for the party being critiqued. ### ✅ Logical Validity #### Strong reasoning - **Explanatory/mechanistic grounding**: Newport builds a from-scratch technical explanation of transformer mechanics before evaluating Anthropic's claims, then shows the paper's findings map onto that independently-established baseline. This is a legitimate move — if a "finding" is fully predicted by prior, well-established understanding, it's evidence against novelty, not evidence of a breakthrough. Location: ~00:03:34-00:21:00. Strengthened by citing his own dated prior publications (2023, 2024 New Yorker pieces) using the same framing before this report existed. - **Category-mismatch argument (Global Workspace Theory)**: His strongest single point. GWT, as a theory, requires persistent, ongoing, stateful integration of continuously arriving inputs. A transformer forward pass is single-pass, feedforward, and stateless between prompts. This isn't a vibes-based objection — it identifies a specific, checkable structural property (statefulness vs. feedforward) on which the analogy fails. Location: ~00:26:25-00:28:38. This could be clarified further by explicitly naming which GWT criterion (competitive selection among concurrent inputs; broadcast across subsystems over time) is violated, rather than gesturing at "ongoing" and "stateful" together. - **Appeal to verifiable prior art**: Citing a named academic's tweet with a list of specific prior papers (2022 onward) is a legitimate, checkable way to undercut a novelty claim — it's falsifiable (the papers either exist and predate this or they don't) rather than a bare assertion of "this isn't new." - **Consistent equivocation catch ("on its own")**: Correctly identifies that "emerged on its own, not programmed" is definitionally true of all trained neural networks, not a distinguishing fact about Claude specifically — a legitimate point about a phrase that's true but selectively deployed to imply something unusual. #### Fallacies - **Motive speculation presented with more confidence than evidence supports (weak causal inference)**: The closing argument — that anthropomorphic framing serves to distract from harder business questions (token costs, competitive moat, valuation) — is plausible but unfalsified. Simpler alternative explanations exist (research communications teams favor vivid, shareable narratives for engagement, recruiting, or prestige reasons unrelated to obscuring unit economics) that Newport doesn't address or rule out. Location: ~00:29:00-00:29:56. He does flag it as "just my opinion," which mitigates but doesn't eliminate the overclaim; a fix would be presenting it explicitly as one plausible hypothesis among others rather than as the video's structural conclusion. - **Some conflation of distinct sources' rigor**: Newport moves fluidly between critiquing an X post, two press headlines, and Anthropic's own paper language as instances of the same "disingenuous" framing, without clearly distinguishing the epistemic standards each should be held to (a clickbait headline and a hedged research report are not the same category of claim, even if both are misleading). Location: throughout, most visible ~00:00:17-00:01:28 and again ~00:22:05-00:25:00. A fix would be explicitly separating "what Anthropic's paper says (hedged)" from "what the media did with it (unhedged)" as two distinct failure points with two distinct remedies. - **Minor overstatement of paper vs. press distinction he himself later draws**: Early in the video Newport treats the paper's own language ("thinking... without writing it down," "on its own") as itself the source of the disingenuousness, but later acknowledges "they're careful in the paper... they don't say that." These two framings are in some tension — the video would be strengthened by resolving up front whether the primary target is the paper's carefully-hedged language, the accompanying video/animation, or the press amplification, since the rebuttal quotes (e.g., "thinking without writing it down") are attributed to the paper itself. #### Tone of voice - Conversational, didactic, and mostly measured — long uninterrupted teaching segments delivered patiently, punctuated by dry humor ("scary X voice," "digital ick," the Hogwarts Great Hall aside). Shifts to more pointed, repeated use of "disingenuous" (used four times across the second half) and rhetorical questions aimed at Anthropic directly ("Hey Anthropic, how are you going to make a profit?"). Closes on a conciliatory note that separates praise for the researchers from criticism of the communications framing, avoiding an adversarial register toward Anthropic's staff even while using strong language about the company's messaging choices. ### 🛡️ DARVO Analysis DARVO (Deny, Attack, Reverse Victim/Offender) describes how an accused party responds to an accusation of wrongdoing. This video is not a response to an accusation — it's an original piece of criticism directed at Anthropic, and Anthropic is not shown responding to it at all within this source. Strictly, DARVO doesn't apply here in either direction: Newport isn't defending against a claim, and no Anthropic rebuttal is present to assess for denial/attack/reversal patterns. The closest adjacent pattern worth noting is in the paper's own hedging structure, as Newport describes it: Anthropic states a limiting disclaimer ("our experiments don't show that Claude can have experiences or feel things") immediately followed, in Newport's account, by suggestive language that leaves the door open ("that's all very suggestive that there's something new going on here"). This is a **hedge-then-imply** rhetorical pattern — technically accurate denial paired with suggestive framing that undercuts the denial's practical effect — but it is not DARVO: there's no accusation being deflected, no attack on an accuser's credibility, and no reversal casting Anthropic as victim. It's closer to a marketing-communications pattern (qualify the claim precisely enough to avoid falsifiability, while retaining the rhetorical benefit of the suggestive version) than to the conflict-response dynamic DARVO describes. ## 🎯 Conclusion Newport's core argument is that Anthropic's J-space interpretability report describes a real, causally-validated finding — decodable, human-interpretable annotations in Claude's internal activations that measurably influence output tokens — using language calibrated to sound novel and mind-like when the underlying mechanism is the standard, years-old account of how transformer-based deep networks work. His strongest point is technical and specific: the report's implicit comparison to Global Workspace Theory doesn't hold up structurally, because GWT requires persistent, integrating state that a stateless, feedforward transformer inference pass doesn't have. His weakest point is more speculative: an unfalsified inference that the anthropomorphic framing serves to distract from scrutiny of Anthropic's business fundamentals, which he himself labels opinion rather than demonstrated fact. Taken together, the technical critique holds up well on its own logic and evidence; the motive argument is a plausible but separate claim that shouldn't be weighted the same as the mechanistic rebuttal. **Practical next steps for readers**: when an AI lab's research communication uses mind-adjacent language ("thinks," "on its own," "puzzles over"), check whether the underlying mechanism (as described in the technical sections, not the framing) would sound as novel in plain, mechanistic terms — and treat inferences about a company's communications motives as a separate, lower-confidence claim from technical assessments of what the research actually shows. ## 🔄 Refactor Appendix — Anthropic's New "Research" Report is Dumb. ### 📚 Literature Atoms (H3) #### GWT requires stateful integration; transformer inference is feedforward > [!quote] ID:yt_m6ucYXJyjAk_t1711 > Remember, this is feed forward, one layer after another in order, and then it's done, and the token is out. Nothing is saved. **Context & use**: - [00:28:31](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1711) - Newport's most structurally specific rebuttal of Anthropic's Global Workspace Theory comparison — identifies a concrete architectural property (statelessness) that the consciousness-framework comparison doesn't address. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### The paper's "thinking without writing it down" describes normal activation updates > [!quote] ID:yt_m6ucYXJyjAk_t1565 > It is writing it down. It has this matrix that it updates the numbers every time. That's its workspace. That's its work pad. **Context & use**: - [00:26:08](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1568) - Direct, quotable rebuttal of a specific paper phrase Newport treats as the clearest instance of misleading framing — useful as a template for reading other AI-lab communications critically. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### Anthropic's finding matches the pre-existing account of how deep networks work > [!quote] ID:yt_m6ucYXJyjAk_t1380 > This is just how we understand large language models to work. It's also how we understand essentially any deep learning neural network to work. **Context & use**: - [00:23:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1380) - Newport's core novelty rebuttal, delivered right after his "incredibly disingenuous" thesis statement — the load-bearing claim the rest of the critique builds on. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### A confirmatory finding isn't a novel finding **Claim** (Courtney's voice): A result that exactly matches a well-established prior model of how a system works is evidence the system works as expected, not evidence of a new capability — regardless of how the result is framed. **Evidence pointer**: [[#Anthropic's finding matches the pre-existing account of how deep networks work]] **Scope & conditions**: Holds when the "prior model" being confirmed is genuinely well-established and predates the new report (as Newport substantiates via his own 2023/2024 explainers and the cited 2022-era prior art); doesn't hold if the prior model itself was contested or the new result actually extends it in a checkable way. **Links**: - [[AI research communications]] — because applies-in evaluating any lab's press-released research - [[Media literacy]] — because extends into general source-evaluation practice <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Read Anthropic's actual published report (not just Newport's on-screen reading of it) to independently verify how much of the anthropomorphic framing originates in the paper itself vs. the accompanying video vs. press coverage — https://www.youtube.com/watch?v=m6ucYXJyjAk #xref - [F2] Check Yu Zhaozhang's cited 2022-era prior work on Jacobian-based interpretability for a fuller picture of what's actually new (if anything) in Anthropic's implementation #xref ### → Things 3 - [ ] Skim Anthropic's original "Global Workspace" report directly to cross-check Newport's characterization before citing his critique as a standalone source — media literacy follow-up #things <!-- deep-read:sections end --> ## 📺 Video  ## Highlights - We find that Claude has developed a small collection of internal neural patterns that compared to all its other internal processing play a special role. - Note: 📚 Anthropic finds internal patterns tied to specific words · - From Newport's reading of Anthropic's paper introduction. - Core technical claim being evaluated throughout the video — the "J space" patterns Anthropic's Jacobian-based method surfaced. · #atom/literature/h1 - Claude did this on its own. We didn't program it. It's a machine learned neural network. Nothing is programmed. - Note: 📚 Emergent-learning framing applies to any ML system · - Newport's rebuttal to Anthropic's framing that the J space "emerged on its own during training." - Central to his disingenuous-PR argument: this description is true of every trained neural network, not evidence of anything special. · #atom/literature/h2 - Good. It works like we thought it did. Nothing is new, and nothing has to do with human consciousness. - Note: 📚 J-lens results confirm, not overturn, LLM theory · - Newport's summary verdict after walking through the Mars/color ablation example. - The load-bearing conclusion of the whole critique — separates "good confirmatory research" from "bad framing." · #atom/literature/h3 - AI research PR anthropomorphizes routine mechanistic findings - Note: [[AI hype cycle]] — because extends · - [[AI hype cycle]] — because extends · #atom/zettel/z1 - Check whether Anthropic's "J space" report is publicly linked/paywalled and whether a non-animated technical writeup exists — needed to verify claims independently. #xref - Note: ✏️ Fleeting capture #atom/fleeting/f1 - Yu Zhaozhang's cited 2022+ prior work on Jacobian-based interpretability — worth tracking down as a comparison point if writing anything about "J lens" novelty claims. #articleseed - Note: ✏️ Fleeting capture #atom/fleeting/f2 --- # 📝 Full Notes (Deep-Read Analysis) # 🏷️ Title: Anthropic's New "Research" Report is Dumb. ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - **What happened**: Anthropic published a research report, "On the Biology of a Global Workspace" (referred to in the video as "A Global Workspace and Language Models"), describing a technique — the "J-lens," built on a Jacobian linear-algebra tool — that decodes human-interpretable concepts (a "J space") out of Claude's internal activation vectors. The report shipped with an animated video and language like Claude "pondering" and "puzzling" over concepts silently, "on its own," without anyone programming it. Tech press (Axios, MIT Technology Review) and X posts amplified this into claims that Anthropic had found evidence of consciousness or moral patienthood in Claude. - **Why it matters**: Newport argues the underlying finding — that intermediate layers accumulate interpretable "annotations" that narrow down semantically appropriate output tokens — is not new. It matches the standard account of how transformer-based LLMs and deep neural networks generally have been understood to work for years (he cites his own 2023/2024 New Yorker explainers), and comparable interpretability techniques predate this paper (he cites a University of Illinois professor's tweet listing related work since 2022). His core claim is that Anthropic's framing — hedged just enough to avoid an explicit false claim, while using loaded language ("thinking," "on its own," global-workspace-theory comparisons) — manufactures a sense of novelty and "digital ick" that the actual result doesn't support. - **Key stakeholders**: Anthropic's research and PR/communications functions, AI journalists and outlets that amplified the framing (Axios, MIT Technology Review), the broader public forming beliefs about AI sentience/consciousness, and Anthropic's investors/business stakeholders (Newport's closing argument ties the framing to distraction from harder questions about unit economics and competitive moat). - **Risks and tradeoffs**: Overclaiming or ambiguous framing of interpretability research risks (a) misinforming the public about AI consciousness/moral status, a genuinely consequential question that deserves rigor rather than marketing lift; (b) crowding out scrutiny of Anthropic's business fundamentals (token costs, competitive moat, path to justifying valuation); and (c) devaluing real interpretability research by association with hype cycles, making it harder for the public to distinguish legitimate advances from press-release science. - **Bottom line**: Newport's assessment is that the research itself is competent, incremental, non-novel work executed at a scale (a frontier proprietary model) that outside researchers can't replicate — but the communication of it is, in his word, "disingenuous," designed to generate attention-driving anthropomorphic reactions rather than to inform. ## 📖 Main Content Newport structures the video as a teardown in three movements: (1) read Anthropic's own framing and the media/social reaction to it, (2) teach the actual mechanics of how transformer-based LLMs process a prompt, and (3) return to the paper to show that its findings map exactly onto that mechanical account — meaning nothing "unnerving, cool, or blockbuster" was revealed. **The paper's claim, in Anthropic's words** (as read on-screen): Claude has a small collection of internal neural patterns — the "J space," named for the Jacobian technique used to find them — that "play a special role" relative to other internal processing. Each pattern links to a word, but lighting up doesn't mean the model says that word, only that "the word is on its mind." Unlike a visible scratchpad or chain-of-thought, the J space "operates silently in the model's internal neural activations, allowing the model to think about a concept without writing it down," and — Anthropic's own italics — it "wasn't designed or programmed by us, but instead emerged on its own during Claude's training process." **The tutorial (mechanics of an LLM)**: Newport walks through transformer blocks arranged in sequential layers, each of which can add "annotations" to a token's underlying vector of numbers as the prompt passes through. He offers two analogies: a relay of scholars at successive tables, each reading prior annotations and adding their own; and the layout of a page of Talmud, where a small central text (the token) is surrounded by accumulating layers of commentary. Tokens are first embedded as long vectors of numbers capturing rough meaning; self-attention and feed-forward sublayers within each transformer block then update those numbers, layer by layer, with the final layer mapping the fully annotated result to an output token. He separates two things that are actually entangled in training: **syntactic validity** (what words could grammatically follow — learnable even with a simple n-gram model, likely concentrated in later layers) and **semantic appropriateness** (which of the syntactically valid options actually fits the prompt's meaning — where he argues the accumulated annotations do their work). **Back to the paper, now legible**: Anthropic's Jacobian-based "J-lens" decodes those annotation vectors into human-interpretable labels. Their worked example: for the prompt "the color of the fourth planet from the sun is," the technique surfaces annotations corresponding to "Mars" and "color" mid-network; the final layer combines them to output "red." They validate this causally two ways — replacing the "Mars" annotation values with "Earth" values changes the output to "blue," and "ablating" (zeroing out) the Mars-related pattern produces grammatically valid but semantically arbitrary color outputs. Newport calls this legitimately interesting experimental work, while objecting to Anthropic's use of "ablated," a term borrowed from neuroscience/animal research, as anthropomorphizing a numerical zeroing operation. **Newport's two-part verdict**: First, novelty — he cites a tweet from University of Illinois professor Yu Zhaozhang noting comparable Jacobian-based interpretability work dating to 2022, framing Anthropic's contribution as the first application at the scale of a frontier proprietary model (which outside researchers structurally cannot replicate, since only the lab has model internals) rather than a new idea. Second, and central to the video's title: the *language* used to describe the results. He argues that once you understand the layer-by-layer annotation mechanism, the J-lens findings are "exactly how we've always understood large language models to work" — not just LLMs specifically, but deep networks generally, tracing back to early image-recognition work. He singles out specific phrases as misleading: "thinking about a concept without writing it down" (his rebuttal: updating the activation matrix *is* how it writes things down — that matrix is its only workspace) and "Claude did this on its own, we didn't program it" (his rebuttal: that describes literally all machine learning — nothing is programmed in a trained neural network by definition, so stating it as a distinguishing fact about this system is misleading through omission, not falsehood). **The consciousness section**: Anthropic connects J-space experiments to Global Workspace Theory (GWT), a framework for conscious access in humans and animals, while explicitly hedging that "our experiments don't show that Claude can have experiences or feel things." Newport's strongest technical objection: GWT describes an ongoing, stateful system that integrates continuously arriving inputs and selects among them over time. A transformer forward pass is the opposite — strictly feedforward, one layer to the next, no persisted state between prompts, "and then it's done, and the token is out." He treats this as a structural mismatch the GWT comparison glosses over rather than confronts. **The closing argument (motive)**: Newport's final move shifts from technical critique to an inference about incentives — that anthropomorphic, "digital ick"-inducing framing distracts audiences from harder, more consequential questions about Anthropic's business (token costs, thin competitive moat against smaller specialized models, and what justifies a trillion-dollar valuation). He closes by distinguishing the researchers ("good work") from what he calls the PR department's packaging of it, and calls on Anthropic to publish real computer science papers rather than "glorified press releases." **Terms defined**: J space / J-lens (Anthropic's Jacobian-based interpretability method and the resulting decoded activation patterns); "annotations" (Newport's own explanatory term for information layers add to a token's activation vector as it passes through a network — not Anthropic's term); ablation (in this context, zeroing out specific activation values to test their causal role); Global Workspace Theory (a cognitive-science framework for how a subset of information becomes "consciously" available across a system, here disputed as inapplicable to feedforward inference). **Uncertainty / contested claims**: This analysis is built from the video's auto-generated YouTube captions; a small number of caption artifacts (word-order glitches, filler-word transcription like "It's putting") were present in the raw transcript and are called out inline where a quote spans one. Anthropic's original paper/report was not independently reviewed for this note — all characterizations of "what Anthropic said" reflect Newport's on-screen reading and paraphrase of it, not a separate verification against Anthropic's published text. Newport's inference about Anthropic's *motive* (distraction from business fundamentals) is presented by him as opinion ("that's just my opinion") and is treated here as an argued but unverified claim, not an established fact. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=0) The report and the online reaction - Anthropic's "A Global Workspace and Language Models" report shipped with an animated video. - An X post read aloud claims: "Claude, my friends, by all counts, is a conscious entity. Claude, my dear friends, is a moral patient." - Axios: "Anthropic says Claude has carved out its own space to ponder." MIT Technology Review: "Anthropic found a hidden space where Claude puzzles over concepts." - Newport frames the episode's question: genuine evidence of something new, or "another over-wrought cynical push to generate a fresh wave of relevance-reinforcing digital ick"? #### ⏱️ [00:02:06](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=126) Anthropic's own framing, read verbatim - J space: "a small collection of internal neural patterns that... play a special role." - Explicitly contrasted with visible chain-of-thought: it "operates silently," lets the model "think about a concept without writing it down." - Anthropic's italics, per Newport: the J space "wasn't designed or programmed by us, but instead emerged on its own." #### ⏱️ [00:03:34](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=214) Tutorial: transformer layers as accumulating "annotations" - A prompt passes through sequential transformer-block layers; each layer can add information ("annotations") on top of what prior layers added. - Two analogies: relay tables of scholars adding commentary in sequence; a Talmud page, with the token as the small central text and layered commentary (the Gemara) built up around it by each layer. - The final layer maps the fully annotated representation to an output token (technically a probability distribution the control program samples from). #### ⏱️ [00:06:49](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=409) Tutorial: tokens, embeddings, and the annotation matrix - Prompts are broken into tokens; each token is embedded as a long vector of numbers initially capturing rough meaning. - As the matrix of token-vectors moves layer to layer, self-attention and feed-forward sublayers update those numbers — that update *is* where annotations "live." - "This big... table of numbers, we call this a matrix" — the same matrix Newport later calls the model's only workspace. #### ⏱️ [00:11:23](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=683) Syntactic vs. semantic contribution to next-token choice - Two entangled but conceptually separable jobs: which next tokens are grammatically valid (easy, likely concentrated in final layers, learnable even by a simple n-gram model), and which of those is semantically appropriate given the prompt's meaning (where accumulated annotations do the work). - Example: "the quick brown fox ___" — many grammatically valid continuations (jumped, danced, died, talked), but annotation of "this is a known saying" privileges "jumped." #### ⏱️ [00:13:52](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=832) Back to the paper: the Jacobian/J-lens technique - Jacobian-based linear algebra is used to find which patterns of values in the activation matrix are most causally influential on the final output. - Worked example: prompt "the color of the fourth planet from the sun is" surfaces mid-network annotations decodable as "Mars" and "color"; combined at the final layer, the model outputs "red." - Causal test: swapping the "Mars" pattern for the "Earth" pattern changes the output to "blue," holding the rest of the prompt fixed. #### ⏱️ [00:19:03](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1143) The ablation experiment - Zeroing out ("ablating") the Mars-related annotation still produces grammatically valid output, but the color is no longer semantically tied to Mars — arbitrary colors result. - Newport flags "ablated" itself as an anthropomorphizing word choice, borrowed from neuroscience, for what is a numerical zeroing operation. - His assessment: legitimately cool experimental design, showing the annotations are causally load-bearing for output. #### ⏱️ [00:20:45](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1245) Take 1 — the technique isn't novel - Cites a tweet from University of Illinois professor Yu Zhaozhang: comparable Jacobian-based interpretability work traced to 2022, with several related papers listed. - Anthropic's distinct contribution, per Newport, is applying it at the scale of a frontier proprietary model — which outside researchers can't replicate, since only the lab has access to those internals — not a new underlying idea. #### ⏱️ [00:22:05](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1325) Take 2 — "incredibly disingenuous" framing - Core thesis stated directly: "the way that Anthropic described these results and the way that they were echoed after Anthropic described them, I think is incredibly disingenuous." - Newport ties this to his own 2023 and 2024 New Yorker explainers, which used the same "annotations that build on annotations" framing years before this report — this is not new territory. - Generalizes the point to all deep learning: "This is just how we understand large language models to work. It's also how we understand essentially any deep learning neural network to work," tracing back to early image-recognition (Yann LeCun-era) work. #### ⏱️ [00:25:36](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1536) Line-by-line critique: "thinking... without writing it down" - Quotes the paper's line about the J space operating silently, letting the model "think about a concept without writing it down." - Direct rebuttal: "It's not thinking... It is writing it down... It has this matrix that it updates the numbers every time. That's its workspace. That's its work pad." - Frames this as the paper describing ordinary deep-network mechanics in language engineered to sound mysterious. #### ⏱️ [00:26:25](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1585) The consciousness / Global Workspace Theory section - Anthropic's hedge, read on screen: "our experiments don't show that Claude can have experiences or feel things" — followed, per Newport, by language he characterizes as "all very suggestive that there's something new going on here." - Newport's structural objection: GWT describes an ongoing, stateful system integrating continuously arriving inputs and selecting among them; a transformer inference pass is feedforward, layer to layer, with no persisted state — "and then it's done, and the token is out. Nothing is saved. There's no states that are changing." - Separately notes GWT itself remains a contested framework even for human consciousness — a further layer of unsettled ground under the comparison. #### ⏱️ [00:27:03](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1623) "Claude did this on its own" — true of all machine learning - Targets the paper's framing that the J space "emerged on its own" rather than being programmed. - Rebuttal: this describes every trained neural network by definition — nothing in machine learning is hand-programmed; weights start random and are learned through training until loss is minimized. Stating it as if distinctive to Claude implies something it doesn't establish. #### ⏱️ [00:29:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1740) The real stakes: distraction from business questions - Newport's inference about motive: anthropomorphic framing generates "a general sense of ick" that makes the technology feel powerful and alive, which crowds out harder questions. - Named examples of the questions he thinks get skipped: Anthropic's per-token costs, thin competitive moat against smaller specialized/hand-harnessed models, and what justifies a trillion-dollar valuation. - Explicitly flagged by Newport as inference/opinion rather than a proven claim about intent. #### ⏱️ [00:30:04](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1804) Conclusion and call to action - Distinguishes the researchers ("I like your work and this is good work") from the communications packaging around it. - Direct ask: "Anthropic, will you stop with the... press release research reports? Write computer science papers, sell products, convince us to buy the products." - Final verdict on the episode's framing question: no blockbuster reveal — "we confirmed that yes, language models work the way that at least I always thought they did." ## 📚 References - Anthropic — research report titled (per on-screen reading) "A Global Workspace and Language Models," accompanied by an animated explainer video; specific URL not stated in the transcript. - Axios — headline "Anthropic says Claude has carved out its own space to ponder"; specific article URL not stated in the transcript. - MIT Technology Review — headline "Anthropic found a hidden space where Claude puzzles over concepts"; specific article URL not stated in the transcript. - Yu Zhaozhang (University of Illinois professor) — X/Twitter post citing Jacobian-based interpretability work dating to 2022, listing several related papers; specific post URL not stated in the transcript. - Cal Newport, The New Yorker — prior long-form explainers on LLM architecture (cited by Newport as dating to 2023 and 2024) using the same layered-annotation framing this report is compared against; specific article titles/URLs not stated in the transcript. - Yann LeCun / early deep-learning image-recognition work — referenced generally as the origin of the "layers pick up different features" account of deep networks; no specific paper cited in the transcript. ## 🗣️ Paraphrase Anthropic published an interpretability report showing that Claude's internal activations contain decodable, human-interpretable patterns — a "J space" — that causally shape which word comes out at the end of a prompt. Their own worked example: feed in "the color of the fourth planet from the sun is," and mid-network you can find a pattern that decodes to "Mars" and another that decodes to "color"; swap the Mars pattern for an Earth pattern and the model now says "blue" instead of "red." That's genuinely neat, causally-grounded interpretability work. But Newport's issue isn't the finding — it's the packaging. Once you understand, in ordinary terms, how transformer layers work (each layer adds interpretable "annotations" to a token's representation, later layers narrow down semantically appropriate output using those annotations), Anthropic's result is exactly what that mechanical picture predicts. Nothing about it requires or supports language like "thinking silently," "on its own," or comparisons to a framework for human consciousness. His sharpest technical objection targets the consciousness comparison specifically: Global Workspace Theory describes an ongoing system that integrates a continuous stream of inputs over time, while a transformer forward pass is a single, stateless, feedforward computation that starts fresh with every prompt. Calling attention to that mismatch is a structural argument, not just a stylistic complaint. His closing argument is more speculative: that this kind of anthropomorphic framing serves Anthropic's interests by generating attention and unease that displace scrutiny of harder business questions — token costs, competitive moat, and what justifies the company's valuation. He's explicit that this last part is his opinion rather than a demonstrated fact, while the technical critique of the paper's language is argued in more detail and grounded in the mechanics he spent the first third of the video teaching. ## 📰 Newsletter **A researcher's rebuttal to "Claude is conscious" headlines — worth the 31 minutes.** Anthropic's newest interpretability report ("A Global Workspace and Language Models") set off a familiar cycle: an animated explainer video, breathless X posts about Claude being a "moral patient," and headlines like MIT Technology Review's "Anthropic found a hidden space where Claude puzzles over concepts." Cal Newport's response is a genuinely useful teardown — not a dismissal of the research (he calls it "good work"), but a line-by-line challenge to the language used to describe it. He spends the first third of the video teaching, in plain terms, how transformer layers actually work — annotations accumulating layer by layer, narrowing down a semantically appropriate next word. Then he shows that Anthropic's headline finding maps exactly onto that mechanical picture, and that comparable interpretability techniques predate this paper by several years. His sharpest point: comparing the result to Global Workspace Theory (a framework for human consciousness) glosses over the fact that transformer inference is stateless and feedforward, while GWT describes an ongoing, integrating system — a structural mismatch, not just a stretch. Useful watch for anyone who writes, reviews, or shares AI research communications — a good case study in reading past a press release's framing to what the underlying result actually shows. **Watch**: [Anthropic's New "Research" Report is Dumb.](https://www.youtube.com/watch?v=m6ucYXJyjAk) ## ✅ Action Items #### DevRel - Developer advocacy: build a short internal reference distinguishing "interpretability finding" from "consciousness claim" for use when developers or community members ask about AI-lab research announcements. - Education: use this video's transformer-layer explainer as a citable plain-language teaching resource when introducing how LLMs process prompts. - Developer experience: none directly applicable — this source concerns research communications, not tooling. - Developer marketing: when citing AI-lab research in marketing copy, default to the paper's own hedged language rather than the press coverage's amplified framing. - Community management: watch for community members repeating "Claude is conscious" style claims sourced from press coverage rather than the primary report, and be ready to point to a plain-language rebuttal. - Internal/external/cross-project: flag this as a recurring pattern (AI-lab research report → animated video → anthropomorphized press coverage) worth a standing media-literacy note rather than a one-off response. #### OSPO - Policy & Governance: not directly addressed in this source — no governance mechanism proposed or critiqued. - License Compliance: not addressed in this source. - Security Management: not addressed in this source. - Software Inventory: not addressed in this source. - Developer Support: not addressed in this source. - Community Engagement: not addressed in this source. - Strategic Alignment: relevant as a cautionary case study when evaluating vendor (AI-lab) claims against roadmap decisions — distinguish PR framing from the technical report before acting on a vendor's stated capability claims. #### Community - Maintainers: not addressed in this source. - Contributors: not addressed in this source. - Sponsors: not addressed in this source. - Ecosystem: worth sharing as a model for how to publicly critique a research communication on its technical merits (Newport separates "good research" from "disingenuous framing" rather than attacking the researchers) — a useful tone reference for any ecosystem disagreement about vendor claims. #### WordPress - Core: not addressed in this source. - Core Performance: not addressed in this source. - Core AI: directly relevant — any WordPress Core AI messaging about AI-assisted features should avoid the same anthropomorphizing pattern Newport critiques (e.g., avoid "the AI thinks about" language for deterministic or well-understood mechanisms). - **Core Program Team:** not addressed in this source. - Mobile: not addressed in this source. - Playground: not addressed in this source. - Community: not addressed in this source. - Training: this video is a usable teaching example for a "how to read an AI research announcement critically" module. - Marketing: same caution as Core AI — avoid mystifying language about AI features in WordPress-facing marketing copy. - Support: not addressed in this source. - Design: not addressed in this source. - Documentation: not addressed in this source. - Photos: not addressed in this source. - TV: not addressed in this source. - Polyglots: not addressed in this source. - Accessibility: not addressed in this source. - Meta: not addressed in this source. - CLI: not addressed in this source. - Hosting: not addressed in this source. - Tide: not addressed in this source. - Openverse: not addressed in this source. - Themes: not addressed in this source. - Plugins: not addressed in this source. #### The WPCC - Programs: not addressed in this source. - Projects: not addressed in this source. - Funding experiments: not addressed in this source. ### FAIR - Technical independence: relevant tangentially — Newport's point that only Anthropic could apply this technique to Claude (because only they have model internals) is itself an argument for open-weight models as a check on interpretability claims from closed labs. - Policies and covenants: not addressed in this source. - Productized governance artifacts: a short "how to evaluate an AI-lab research announcement" checklist (distinguish paper language from press coverage; check for prior art; separate causal claims from suggestive juxtaposition) would be a natural artifact derived from this video. ## 🧪 Hypotheses 1. Research communications from AI labs that use anthropomorphic language (e.g., "thinks," "on its own," "puzzles over") generate measurably higher social media engagement than functionally equivalent communications using mechanistic language, controlling for underlying finding. 2. Readers given a plain-language transformer-architecture explainer before reading an AI-lab interpretability report rate the report's findings as less novel and less "conscious-like" than readers who read the report without that context. 3. The Jacobian-based interpretability technique described in this report, when independently reproduced on open-weight models, will show comparable annotation/decoding patterns to those Anthropic reports for Claude — supporting Newport's claim that this is a general transformer property rather than something specific or newly discovered in Claude. 4. Volume of public discourse invoking "AI consciousness" correlates more strongly with the presence of hedged-but-suggestive language and accompanying illustrative media (animated videos, loaded terminology) in an AI lab's own communications than with the magnitude or novelty of the underlying technical finding. 5. Global Workspace Theory, as standardly defined in cognitive science, requires persistent cross-input state integration that a strictly feedforward transformer inference pass structurally cannot exhibit — meaning any GWT-based consciousness claim about current-generation feedforward LLM inference is, by definition, unsupported regardless of interpretability findings about single-pass activations. ### 📊 Methodology - **H1**: A/B test identical technical findings summarized in anthropomorphic vs. mechanistic language, distributed to matched audiences; measure shares, comments, and self-reported perception of "aliveness" of the described system. - **H2**: Randomized controlled reading study — one group reads a plain-language LLM architecture primer before the Anthropic report, a control group reads the report cold; both groups rate novelty and human-likeness of the findings on a standardized scale. - **H3**: Independent replication of the Jacobian/J-lens technique on an open-weight model (e.g., a Llama or Mistral variant) using the same Mars/color-style causal-intervention protocol; compare decoded annotation patterns and ablation effects to Anthropic's reported results. - **H4**: Content analysis of AI-lab research communications over a defined period, coding for anthropomorphic language density and accompanying media (video, animation), correlated against social-media engagement and volume of "consciousness"-framed coverage. - **H5**: Formal comparison of Global Workspace Theory's defining criteria (persistent state, competitive selection among concurrent inputs, broadcast/integration across subsystems over time) against the architectural properties of feedforward transformer inference (statelessness between prompts, single-pass computation) to establish whether the comparison is falsifiable in principle or a category mismatch by definition. ### ❓ Questions - Does Anthropic's original written report include more or less hedging than the accompanying animated video and social/PR amplification — i.e., where exactly does the anthropomorphizing enter the communication chain? - How would Newport's critique change, if at all, for a multi-turn or agentic system that does carry persistent state across a session (as opposed to the single feedforward pass he describes)? - What would a genuinely novel (not previously demonstrated at any scale) interpretability finding about LLM internals look like, and would Newport's own framework recognize it as such? - Is there a version of "interpretability progress at frontier model scale" that's worth reporting on its own terms (bigger model, more layers, higher-dimensional activations) without needing consciousness-adjacent framing to generate interest? - To what extent do AI labs' research communications teams operate independently of the research teams, and does that organizational split explain the gap Newport identifies between "good research" and "disingenuous framing"? ### 🔮 Future Research - A systematic comparison of interpretability terminology across AI labs (Anthropic, OpenAI, Google DeepMind, Meta) to see whether anthropomorphizing language is a house style specific to one lab or an industry-wide pattern. - A history of the specific Jacobian-based interpretability technique from its 2022-era origins (per the Yu Zhaozhang citation) through to this report, to establish precisely what, if anything, is new at the implementation or scale level. - Independent, peer-reviewed evaluation of whether Global Workspace Theory can be meaningfully operationalized for any current AI architecture, feedforward or otherwise — separate from any single lab's marketing framing. - A media-literacy study tracking how a hedged technical claim ("our experiments don't show X, but...") propagates through tech press into unhedged public claims ("Anthropic admitted X"). ### 💡 Implications **Practice**: Communicators, journalists, and developer-relations professionals writing about AI-lab research should default to checking the underlying mechanism (what architecture, what's actually being measured) before amplifying framing language from a press release or accompanying video. **Policy**: As public and regulatory interest in AI consciousness/moral-status questions grows, this video is a case study in why source discipline matters — hedged claims by an interested party (the lab itself) shouldn't set the terms of a genuinely open scientific and ethical question. **Security**: Not directly addressed in this source — no security vulnerability or concrete attack vector is described. **Education**: The video doubles as a teaching resource for transformer mechanics (layers, tokens, embeddings, attention) delivered in accessible analogy-driven language, useful independent of the critique it's embedded in. **Funding models**: Newport's closing argument implies a funding-scrutiny implication directly — that anthropomorphic hype cycles can substitute for scrutiny of unit economics (token costs, competitive moat, valuation justification) when evaluating heavily funded AI labs. ### 🧠 Learning Styles Assessment **Reading level estimate**: Approximate Flesch-Kincaid Grade ~9-10 (estimated manually from transcript sentence length and technical vocabulary density; conversational spoken register with frequent self-correction and long, clause-heavy sentences typical of unscripted podcast delivery, offset by deliberately plain-language analogies for technical concepts — no automated readability tool was run against the transcript). #### 👀 Visual learners - **Present**: Newport shares his screen to show the actual paper, the J-lens example table, and the specific quoted tweets/headlines, giving visual grounding to claims made verbally. - **Improve**: A side-by-side diagram contrasting "feedforward, stateless, layer-by-layer" (transformer inference) against "ongoing, stateful, integrating" (Global Workspace Theory) would make his strongest technical point far more scannable than prose description alone. #### 🔊 Aural learners - **Present**: Long-form, single-speaker, conversational narration is the native format of this podcast and well suited to audio-only consumption. - **Improve**: A tighter 3-5 minute audio pull-quote of just the GWT mismatch argument and the closing "write real papers" ask would work as a standalone highlight for listeners who don't have 31 minutes. #### 🗣️ Verbal learners - **Present**: Newport explicitly defines terms as he introduces them (tokens, embeddings, annotations, ablation) and offers two distinct analogies (scholars at tables; Talmud commentary) for the same underlying mechanism. - **Improve**: A short glossary card (J space, J-lens, ablation, Global Workspace Theory, feedforward) would help viewers who want to look up a specific term without re-watching the tutorial section. #### 🧩 Kinesthetic learners - **Present**: The Mars/Earth/blue causal-intervention example is concrete enough that a viewer could conceptually "try it" by imagining swapping the annotation. - **Improve**: A companion notebook or interactive demo reproducing the swap on an open-weight model would let viewers verify the mechanism hands-on rather than take the paper's word for it. #### 🧠 Logical learners - **Present**: The video is structured as an explicit two-part argument (novelty claim, then framing claim), each supported by named evidence (a prior-art tweet; a term-by-term critique of the paper's language). - **Improve**: A simple table mapping "Anthropic's phrase" → "what it technically means" → "why the phrasing is misleading" would sharpen the already-logical structure into something scannable. #### 🧑‍🤝‍🧑 Social learners - **Present**: The video opens by engaging directly with a real X post and named outlets (Axios, MIT Technology Review), situating the critique within an actual public conversation. - **Improve**: Newport's own newsletter/community (Deep Questions listeners) is the implied discussion venue, but the video itself doesn't include a structured prompt for listener response or debate. #### 🧘‍♀️ Solitary learners - **Present**: Self-contained single-episode argument that a solo viewer can follow start to finish without external context. - **Improve**: A short self-check ("can you now explain in your own words why 'thinking without writing it down' is a misleading phrase?") would suit independent, reflective viewers. #### 🌿 Naturalistic learners - **Present**: Not addressed in this source — no ecosystem or environmental framing present. - **Improve**: Not applicable; forcing an ecological analogy onto an interpretability-research critique would add noise rather than clarity. ### 📊 Accessibility Assessment #### Screen reader semantics - **Present**: Not directly assessable from a transcript-only review; hosted on standard YouTube, which provides basic player semantics. - **Improve**: If republished as an article, use real headings for each argument beat (paper's claim → tutorial → paper revisited → verdict) rather than relying on narration structure alone. #### Keyboard navigation - **Present**: Standard YouTube player keyboard controls apply. - **Improve**: Not applicable beyond standard player behavior — no custom interactive elements in this source. #### Contrast and readability - **Present**: Not assessable from transcript alone (no visual/screenshot review performed). - **Improve**: Screen-shared paper excerpts should use adequately large, high-contrast text given Newport reads from the paper on-screen at several points. #### Text resizing and reflow - **Present**: Standard YouTube player behavior. - **Improve**: Not applicable — no custom UI to evaluate. #### Media and captions - **Present**: Auto-generated English captions exist and were used as the source for this analysis. - **Improve**: Auto-captions contain occasional disfluency artifacts (e.g., a stray "It's putting" mid-sentence, "feed for Remember" instead of a clean "feed-forward. Remember...") around technically precise moments — manual caption correction would help viewers relying on captions for exact wording of the technical rebuttal. #### Forms and interactions - **Present**: Not applicable — no forms or interactive elements in this source. - **Improve**: Not applicable. #### Standards reference - No WCAG 2.2 AA gaps can be confirmed without a visual review of the actual video; the one concrete, verifiable gap from this transcript-only analysis is caption precision on technical terminology during the paper-reading segments (relates to WCAG 1.2.2 Captions). ### 🌍 DEIB Assessment #### Cultural breadth - **Present**: Single speaker, US-centric framing (Axios, MIT Technology Review, X/Twitter as the cited media ecosystem); one cross-cultural reference point (the Talmud/Gemara analogy) used respectfully and accurately to explain a technical concept. - **Improve**: Additional international AI-research or AI-policy voices on the same interpretability-vs-hype question would broaden the perspective beyond US tech media and US-based academic citation (Yu Zhaozhang, University of Illinois). #### Inclusive language - **Present**: Technical, gender-neutral language throughout; the Talmud analogy is introduced with accurate framing (correctly names the Mishnah and Gemara) rather than as a casual or reductive reference. - **Improve**: A few dense technical terms (Jacobian, ablation, feedforward) are defined in-line but assume some comfort with linear algebra vocabulary; a plain-language gloss alongside first use would widen accessibility for non-technical viewers. #### Neurodiversity support - **Present**: Clear two-part argument structure (novelty, then framing) with explicit signposting ("idea number one," "idea number two") helps viewers track the argument. - **Improve**: The video runs long (31:42) with a lengthy embedded tutorial before the critique resumes; chapter markers or a written companion summary would help viewers who benefit from non-linear or skimmable access to the argument. #### Representation - **Present**: One male presenter; quoted parties include an unnamed X poster, two publications, and one named academic (Yu Zhaozhang). - **Improve**: Not enough content in a single-speaker critique video to meaningfully assess representation; would require the wider Deep Questions catalog for a fair evaluation. #### Psychological safety - **Present**: Newport is careful to separate criticism of the *communication* from criticism of the *researchers*, explicitly praising the underlying work ("I like your work and this is good work") rather than attacking individuals — a model of substantive disagreement without personal targeting. - **Improve**: The closing section's inference about Anthropic's motive (distraction from business scrutiny) is more speculative than the technical critique and is framed with confident rhetoric ("I do not like the way Anthropic talks about the research") that could read as more certain than the evidence supports; explicitly labeling that section as opinion earlier (rather than only at the very end) would improve safety/fairness for the party being critiqued. ### ✅ Logical Validity #### Strong reasoning - **Explanatory/mechanistic grounding**: Newport builds a from-scratch technical explanation of transformer mechanics before evaluating Anthropic's claims, then shows the paper's findings map onto that independently-established baseline. This is a legitimate move — if a "finding" is fully predicted by prior, well-established understanding, it's evidence against novelty, not evidence of a breakthrough. Location: ~00:03:34-00:21:00. Strengthened by citing his own dated prior publications (2023, 2024 New Yorker pieces) using the same framing before this report existed. - **Category-mismatch argument (Global Workspace Theory)**: His strongest single point. GWT, as a theory, requires persistent, ongoing, stateful integration of continuously arriving inputs. A transformer forward pass is single-pass, feedforward, and stateless between prompts. This isn't a vibes-based objection — it identifies a specific, checkable structural property (statefulness vs. feedforward) on which the analogy fails. Location: ~00:26:25-00:28:38. This could be clarified further by explicitly naming which GWT criterion (competitive selection among concurrent inputs; broadcast across subsystems over time) is violated, rather than gesturing at "ongoing" and "stateful" together. - **Appeal to verifiable prior art**: Citing a named academic's tweet with a list of specific prior papers (2022 onward) is a legitimate, checkable way to undercut a novelty claim — it's falsifiable (the papers either exist and predate this or they don't) rather than a bare assertion of "this isn't new." - **Consistent equivocation catch ("on its own")**: Correctly identifies that "emerged on its own, not programmed" is definitionally true of all trained neural networks, not a distinguishing fact about Claude specifically — a legitimate point about a phrase that's true but selectively deployed to imply something unusual. #### Fallacies - **Motive speculation presented with more confidence than evidence supports (weak causal inference)**: The closing argument — that anthropomorphic framing serves to distract from harder business questions (token costs, competitive moat, valuation) — is plausible but unfalsified. Simpler alternative explanations exist (research communications teams favor vivid, shareable narratives for engagement, recruiting, or prestige reasons unrelated to obscuring unit economics) that Newport doesn't address or rule out. Location: ~00:29:00-00:29:56. He does flag it as "just my opinion," which mitigates but doesn't eliminate the overclaim; a fix would be presenting it explicitly as one plausible hypothesis among others rather than as the video's structural conclusion. - **Some conflation of distinct sources' rigor**: Newport moves fluidly between critiquing an X post, two press headlines, and Anthropic's own paper language as instances of the same "disingenuous" framing, without clearly distinguishing the epistemic standards each should be held to (a clickbait headline and a hedged research report are not the same category of claim, even if both are misleading). Location: throughout, most visible ~00:00:17-00:01:28 and again ~00:22:05-00:25:00. A fix would be explicitly separating "what Anthropic's paper says (hedged)" from "what the media did with it (unhedged)" as two distinct failure points with two distinct remedies. - **Minor overstatement of paper vs. press distinction he himself later draws**: Early in the video Newport treats the paper's own language ("thinking... without writing it down," "on its own") as itself the source of the disingenuousness, but later acknowledges "they're careful in the paper... they don't say that." These two framings are in some tension — the video would be strengthened by resolving up front whether the primary target is the paper's carefully-hedged language, the accompanying video/animation, or the press amplification, since the rebuttal quotes (e.g., "thinking without writing it down") are attributed to the paper itself. #### Tone of voice - Conversational, didactic, and mostly measured — long uninterrupted teaching segments delivered patiently, punctuated by dry humor ("scary X voice," "digital ick," the Hogwarts Great Hall aside). Shifts to more pointed, repeated use of "disingenuous" (used four times across the second half) and rhetorical questions aimed at Anthropic directly ("Hey Anthropic, how are you going to make a profit?"). Closes on a conciliatory note that separates praise for the researchers from criticism of the communications framing, avoiding an adversarial register toward Anthropic's staff even while using strong language about the company's messaging choices. ### 🛡️ DARVO Analysis DARVO (Deny, Attack, Reverse Victim/Offender) describes how an accused party responds to an accusation of wrongdoing. This video is not a response to an accusation — it's an original piece of criticism directed at Anthropic, and Anthropic is not shown responding to it at all within this source. Strictly, DARVO doesn't apply here in either direction: Newport isn't defending against a claim, and no Anthropic rebuttal is present to assess for denial/attack/reversal patterns. The closest adjacent pattern worth noting is in the paper's own hedging structure, as Newport describes it: Anthropic states a limiting disclaimer ("our experiments don't show that Claude can have experiences or feel things") immediately followed, in Newport's account, by suggestive language that leaves the door open ("that's all very suggestive that there's something new going on here"). This is a **hedge-then-imply** rhetorical pattern — technically accurate denial paired with suggestive framing that undercuts the denial's practical effect — but it is not DARVO: there's no accusation being deflected, no attack on an accuser's credibility, and no reversal casting Anthropic as victim. It's closer to a marketing-communications pattern (qualify the claim precisely enough to avoid falsifiability, while retaining the rhetorical benefit of the suggestive version) than to the conflict-response dynamic DARVO describes. ## 🎯 Conclusion Newport's core argument is that Anthropic's J-space interpretability report describes a real, causally-validated finding — decodable, human-interpretable annotations in Claude's internal activations that measurably influence output tokens — using language calibrated to sound novel and mind-like when the underlying mechanism is the standard, years-old account of how transformer-based deep networks work. His strongest point is technical and specific: the report's implicit comparison to Global Workspace Theory doesn't hold up structurally, because GWT requires persistent, integrating state that a stateless, feedforward transformer inference pass doesn't have. His weakest point is more speculative: an unfalsified inference that the anthropomorphic framing serves to distract from scrutiny of Anthropic's business fundamentals, which he himself labels opinion rather than demonstrated fact. Taken together, the technical critique holds up well on its own logic and evidence; the motive argument is a plausible but separate claim that shouldn't be weighted the same as the mechanistic rebuttal. **Practical next steps for readers**: when an AI lab's research communication uses mind-adjacent language ("thinks," "on its own," "puzzles over"), check whether the underlying mechanism (as described in the technical sections, not the framing) would sound as novel in plain, mechanistic terms — and treat inferences about a company's communications motives as a separate, lower-confidence claim from technical assessments of what the research actually shows. ## 🔄 Refactor Appendix — Anthropic's New "Research" Report is Dumb. > Density: 3 Literature atoms, 1 Permanent candidate. Resonance filter applied — dropped several strong candidate quotes (the Mars/Earth/blue example, the Talmud analogy) in favor of the three claims most load-bearing for Newport's actual thesis. **📚 Literature Atoms** ### GWT requires stateful integration; transformer inference is feedforward > [!quote] ID:yt_m6ucYXJyjAk_t1711 > Remember, this is feed forward, one layer after another in order, and then it's done, and the token is out. Nothing is saved. **Context & use**: - [00:28:31](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1711) - Newport's most structurally specific rebuttal of Anthropic's Global Workspace Theory comparison — identifies a concrete architectural property (statelessness) that the consciousness-framework comparison doesn't address. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### The paper's "thinking without writing it down" describes normal activation updates > [!quote] ID:yt_m6ucYXJyjAk_t1565 > It is writing it down. It has this matrix that it updates the numbers every time. That's its workspace. That's its work pad. **Context & use**: - [00:26:08](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1568) - Direct, quotable rebuttal of a specific paper phrase Newport treats as the clearest instance of misleading framing — useful as a template for reading other AI-lab communications critically. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### Anthropic's finding matches the pre-existing account of how deep networks work > [!quote] ID:yt_m6ucYXJyjAk_t1380 > This is just how we understand large language models to work. It's also how we understand essentially any deep learning neural network to work. **Context & use**: - [00:23:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1380) - Newport's core novelty rebuttal, delivered right after his "incredibly disingenuous" thesis statement — the load-bearing claim the rest of the critique builds on. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> **🧠 Permanent Candidates** #### A confirmatory finding isn't a novel finding **Claim** (Courtney's voice): A result that exactly matches a well-established prior model of how a system works is evidence the system works as expected, not evidence of a new capability — regardless of how the result is framed. **Evidence pointer**: [[#Anthropic's finding matches the pre-existing account of how deep networks work]] **Scope & conditions**: Holds when the "prior model" being confirmed is genuinely well-established and predates the new report (as Newport substantiates via his own 2023/2024 explainers and the cited 2022-era prior art); doesn't hold if the prior model itself was contested or the new result actually extends it in a checkable way. **Links**: - [[AI research communications]] — because applies-in evaluating any lab's press-released research - [[Media literacy]] — because extends into general source-evaluation practice <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Read Anthropic's actual published report (not just Newport's on-screen reading of it) to independently verify how much of the anthropomorphic framing originates in the paper itself vs. the accompanying video vs. press coverage — https://www.youtube.com/watch?v=m6ucYXJyjAk #xref - [F2] Check Yu Zhaozhang's cited 2022-era prior work on Jacobian-based interpretability for a fuller picture of what's actually new (if anything) in Anthropic's implementation #xref ### → Things 3 - [ ] Skim Anthropic's original "Global Workspace" report directly to cross-check Newport's characterization before citing his critique as a standalone source — media literacy follow-up #things --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=m6ucYXJyjAk)* **[00:00]** Last week, Anthropic released another one of their infamous research reports. This one was titled A Global Workspace and Language Models, and it was accompanied, like all great scientific research, by a lavishly produced animated movie. Now, this report, not surprisingly, uh soon led to some breathless excitement on X. Here's one such tweet. I'll read the beginning using my best sort of uh scary X voice. Anthropic just admitted they have **[00:30]** discovered what I and many others have been claiming exist for a very long time explicitly. Claude, my friends, by all counts, is a conscious entity. Claude, my dear friends, is a moral patient. All right, the traditional tech media also quickly began writing about this report using intensely anthropomorphized language. An Axios headline read, Anthropic says Claude has carved out its own space to ponder. The MIT Tech Review exclaimed, Anthropic **[01:03]** found a hidden space where Claude puzzles over concepts. All right, so what are we to make of this report? Has Anthropic revealed evidence that their LLMs are more human-like and alive than we realized? Or, like so many such reports in recent months, is this yet another over-wrought cynical push to generate a fresh wave of relevance-reinforcing digital ick? Well, it's Thursday, which means it's time for a reality check episode of this podcast, which makes this the perfect opportunity to go searching for some **[01:35]** measured answers, which is exactly what we'll do. As always, I'm Cal Newport, and this is Deep Questions, the show for people seeking depth in a distracted world. All right, so let's start by looking a little bit closer on how Anthropic describes their findings in the introduction of their paper. I'll load it up here. And I'll go down to the introduction. **[02:06]** All right, so here's what they say. We find that Claude has developed a small collection of internal neural patterns that compared to all its other internal processing play a special role. We call the collection of these patterns the J space named after the technique we used to find them involving a mathematical concept known as the Jacobian. Each J space pattern is linked to a particular word, but when one of these patterns lights up, it doesn't mean the model is saying that word, just that the word is on its mind. If you've heard the link of language models having **[02:37]** a scratch pad or chain of thought text they write to themselves while reasoning, the J space is something different. It operates silently in the model's internal neural activations, allowing the model to think about a concept without writing it down. Notably, the J space wasn't designed or programmed by us, but instead emerged on its own during Claude's training process. All right, so in isolation that intro summary sounds pretty impressive. Kind of like Claude This is their italics. On its own made some sort of leap and is now behaving **[03:07]** sufficiently human that we can't help but feel at least a little bit of digital ick. But what's really going on here? Well, to answer this question, I'm going to start with a high-level tutorial on how large language models actually work and I'll add a little bit more detail to it. Stick with me here because once you understand these basics, you are then going to understand that description I just read from Anthropic in a completely different light. All right, so this is an exercise that's worth doing. All right, I'm going to start at a very high level here. At its core, a large **[03:37]** language model like those of the flaw the fable Claude Opus or GPT families can be described underneath the hood as a sequence of what are called transformer blocks that are arranged in layers. One follows the other, so it's collection of layers. GPT-3, which was the last major LLM that they actually published stuff about, details about, had 96 of these transformer block layers. New LLMs probably um have more, but we don't know um how many more. All right. **[04:08]** Let's start at the very high level here. I submit as input a prompt that I have typed to an LLM. You can imagine that this input is going to pass through each of these layers one after another. It'll go into the first layer and come out the other end. It'll go into the second layer and come out the other end, and so on. Now, here's what's important, conceptually speaking. Those layers can add what we'll call for our own purposes annotations to that prompt. Uh so, as the input goes to the first layer, it'll come out the other side **[04:38]** with some annotations added to it, some information added to it that the first layer came up with in its own analysis. Now, this prompt with the first layer annotations goes into the second layer, and it comes out with even more annotations added onto it. And so on. So, as it moves through these layers, the original prompt is there, but you're getting all of these annotations added layer by layer. And critically, later layers can use the annotations from earlier layers to help do their analysis, right? So, it's it's um you can imagine it's almost like you **[05:08]** have like a bunch of long tables of scholars arranged in rows in some vast like Hogwarts Great Hall style dining room. And you're passing this prompt from table to table. And as it arrives at each table, the scholars at the table pour over it. Um they look at the annotations from the scholars that came before them. They figure out how to analyze it. Uh maybe the annotations from the scholars before them will let them know which of the scholars at the next table should look at it, and and then they add their own annotations, and it passes on. The very final layer in a large language model is special. It **[05:39]** takes this heavily annotated version of the prompt, and then it maps it to what word or part of a word to output next. Now, if we're going to be super pedantic, it really outputs a probability distribution over all possible tokens and then then the control program selects one probabilistically, but just think of the last layer takes all the work that all the transformer blocks did and says this is what we are going to output next. Um All right, so that that's what's happening at a at a very um at a very high level. I'm going to add a little bit more layer **[06:11]** la- level of detail now, right? So, now let's add a little bit more um level of detail. What are these annotations and how are they written down and how are they passed from layer to layer? Okay, let's look a little bit closer at what happens to this prompt that you have of handed on uh to your large language model. The first thing that happens is the prompt, which presumably you typed in with, you know, words on your keyboard, gets broken up into what are known as tokens, so in the pieces. And um a token might correspond to a single word or uh a longer word might get broken up into multiple tokens that **[06:42]** represents the different parts of the word. But we we kind of just let's break this up into these uh this fundamental unit we call tokens. Next, those tokens get embedded into a high-dimensional space. And all that really means is, okay, each token is going to be described by a long list of numbers. Uh a long list of numbers is otherwise known as a vector. So, it's a bunch of numbers that are in some order. So, we got a long list of numbers that represent each token. And there's something called a token embedder that takes each token, which is now like letters, right? A word or part of a **[07:13]** word, and transforms it to a long list of numbers. The long list of numbers, the the so-called token embeddings, that come out of this embedding, the numbers that are in there at first are basically capturing the the meaning or details of that token. Um you train the you train in a sort of semi-supervised fashion how to do these embeddings. So, at first, these long list of numbers are the numbers that are in this long list are sort of capturing what this word means or um what this part of a word means, right? We want these things in numbers because that's what we can uh **[07:45]** that's what language models are going to um they understand because in the end we're going to be manipulating numbers here, not letters. But here's the key thing. There's a lot of room in those long list of numbers. So, they don't just encode here's what this word means. As the this long list of numbers, which is how we represent the prompt now, as it moves through each layer, the layers mess around with the long list of numbers for each token. And in messing around, which means messing around with the values in there, **[08:18]** in part, that's where the annotations live. So, this is what's being passed from layer to layer is I just imagine like I have my my my prompt, it's been broken up into tokens, and hanging off of each token is a long list of numbers. And this big so, it's a big table of numbers, we call this a matrix. As this passes from layer to layer, those numbers uh are being updated to include, among other things, the analysis that's being done by each of the transformer blocks. So, there's there's all there's other things that are lots of things are stored in here. Like, for example, if you open up **[08:48]** a transformer block, there's really two parts in it. There's a self-attention mechanism layer, and then a small multi uh perceptron style feedforward neural network, which does the analysis. But like, this self-attention layer is kind of cool. Um it adds information in each of these uh numbers in each of these lists to try to help each um each token understand how relevant each other token is. So, it's like the this is like the the working space. Here's the way I think about it. I often think about uh these tokens and these long list of numbers, these these uh vectors that come with **[09:20]** each of them. It's it's almost like if you had uh it reminds me of the Talmud, right? So, if you if you look at the Jewish Talmud, um what you see in there on every page, if you actually look at uh a printed one, is you have in the very center of each page, not taking up all that much space, is the actual text from the the Mishnah, right? The the sort of the oral law written down by Judah the Prince in the early common era. It's a small part of the page. And then surrounding it all on the page, if you look at a page of the Talmud, is **[09:50]** commentary. Um so you got the the the Gamara, you got the commentary around it, okay? That's how I kind of imagine, conceptually, the prompt being passed through a large language model. It's like you have a book, and on each page you put in the very middle of the page one token. Right? So this book has the the whole prompt with tons of white space around each token. And as it gets kind of passed from layer to layer, scholars are writing more commentary around the token. And then the next layer gets it, and they look at that commentary, and **[10:20]** they look at it, and they build on it, and write more commentary. So I sort of imagine it like we're filling out the Gamara around the Mishnah uh in the Jewish Talmud. All right, so that's roughly speaking what's happening. So we got this matrix moving from layer to layer, which is really just a list of numbers for each token, and those numbers are capturing everything we've learned so far as it as it moves through um as it moves through these these various layers. All right? So that's um that's what's happening. We know like like so so what type of **[10:51]** analysis is happening? Well, you know, we don't know exactly because these are neural networks doing it that are trained in a unsupervised manner. Like any neural networks from an image recognizer to a handwriting, you know, whatever. Any sort of neural network, the whole point is you start it with random weights, and you train it with lots of data till it gets good, and you don't know exactly how it does it, it just does it until you really kind of open up the box and try to understand what's going on. But but we know um I've written about this before in the New Yorker. I've talked to a bunch of experts about it. There's kind of like two things that are happening together with a generative AI model, like a large **[11:21]** language model. Um partially it's learned during training what when outputting the next word what words are uh syntactically appropriate. This is actually pretty easy to train. You can do this even with like a simple n-gram model. This is an easy thing to learn. The sort of statistical nature of language um given this sentence that stops here, what are the words that could follow that from a syntax, like a grammar standpoint, makes sense, right? So, like that's kind of the easy part **[11:52]** and that's baked in. In fact, that's probably captured like primarily in the final layer. We don't know, but probably in the final layer of a large language model. The problem is of of all the words that could follow that make sense, um which one makes semantic sense? So, which one actually makes the most sense given the meaning of what the prompt is asking? So, not just what word would grammatically make sense here, but which word not only grammatically makes sense, but actually matches the meaning of the prompt. And that's really, in my understanding, where all the annotations that follow along these token vectors really are helpful. **[12:23]** So, you kind of have these two things mingled together. They're not separate systems because this is all kind of trained and muddled together, but you kind of have these two things happening at the end when you're in a generative model like an LLM model is like um you know, the quick brown fox and grammatically, there's lots of words that would make sense there. Like, you want to put the word uh the next, but there's lots of things. Jumped, danced, died, talked. Like, you have a lot of words make sense. But then the the the annotations where you're like, oh, this is a saying. This **[12:54]** is a saying that already exists. Uh it's a common saying and it's it always says jumped. And then so, of those words that would privilege jumped as the next one to go. Again, I'm separating out here things that are muddled together in the neural networks themselves. Uh we don't they're not going to work as cleanly as a human might, but that's the way to think about it is that the annotations really help you figure out semantically the right next token output. And then the easier part that's like baked into probably these final layers is the the **[13:25]** grammatically or syntactically what words or next tokens would be valid. So, those two come together and we get um a good output. All right. That's roughly speaking what happens in a large language model. Now that we know this, let's return to the Anthropic paper, okay? Um I'm going to see if I can find it again here. And let's ask uh what So, what what again did they actually find? Okay, so they used a mathematical tool based on a linear algebra notion uh called a **[13:57]** Jacobian to essentially try to understand what is in those vectors of values attached to the tokens that are moving through the layers of the large language model that they're studying. So, it's like we're going to look at the annotations. Now, I know they're just like this a bunch of numbers, but we're going to figure out a way of trying to make more sense of what those numbers actually meant. And so, using uh Jacobians and this is a it's a linear algebra way that involves **[14:27]** taking the partial derivatives of a lot of things. Like basically what they can figure out is like which which combination of numbers from this giant matrix um which pattern of these numbers are seem to be connected and important. In other words, like having a high influence on what the ultimate output is. So, they're they're trying to decode the annotations that are captured in numbers in these vectors. And they're able to figure out certain patterns of these numbers seem to be really influential for the ultimate output. And you could you can do this experimentally, you know, if we **[14:58]** change it if we change exactly this pattern, we're much more likely to get a change in the output than if we change other patterns. That's simple, but it's something like that. And then, and this is what's cool, through a lot of experimentation, they also uh try to as much as possible associate different patterns through a lot of trial and error experimentation with human interpretable concepts like words or numbers or something like that. So, they call this the J lens because it allows them to say looking at a you know, they run a prompt through and they can kind of watch it go all the way **[15:29]** through and as it gets towards the end, they look at these vectors of numbers that have been evolving and updating as it moves through and they can say we can actually decode into English or a human interpretable way. Sometimes it's not like English words. Um some of the annotations. We can kind of understand what some of the annotations are that these layers are using to help uh come up with the answer. Which is pretty cool. So, I'm going to go to the uh paper here to show you an example from the paper. This This table is is called the J lens **[15:59]** reveals the models internal thoughts. I'll put one example up here on the screen. Um so, this shows a prompt that says the color of the fourth the planet fourth from the sun is. So, that's the prompt. The language model needs to expand it with another word or part of a word. When they uh looked at the annotations that this accumulated as it moved through the large language model, they discovered that uh there was an annotation that corresponded to Mars, which is the fourth planet from the sun, um and an annotation that corresponded to color. **[16:33]** Right? So, what's happening here is as uh as that prompt was moving through it, these uh conceptual tables of scholars that are studying this sort of Talmudic, you know, commentary and adding their own commentary to it, at somewhere along the way one of the layers said, um oh, I recognize the fourth planet from the sun, that that sequence of tokens, and I this part of some neural network in some layer is like all about planetary stuff. It's like that's the Mars. **[17:03]** I'm going to write Mars down on my annotation here. We're talking about Mars. That's an important annotation. And somewhere else some other layer was like, this is asking what the color of something is. Like the key thing that we're we're what the thing that comes next is a color. This is asking about a color. Let me write that down. This is, you know, we're we're asking about a color. So, then when you kind of get to the end of this large language model and it's like looking at all the possible words that just statistically make sense to follow this sentence, the color of the planet fourth from the sun is and all sorts of nouns that would make sense and adjectives that make sense. **[17:34]** Um it looks at the annotations like, oh, we're looking for uh of the words that would make sense syntactically, we need something that's a color and a color describes Mars. Hey, what is that? Don't we have that somewhere that's red? Okay, and then it outputs red. So, we could see that's cool. It's like, oh, maybe it's not as mysterious as we thought. Like we know it annotates numerically, but it when we were able to interpret the annotations and annotate it with Mars and color. Which is like exactly like you what you should do if you know, it makes sense. That's so I think that was um that's pretty cool, right? So, they could do that type of thing. **[18:06]** Then what they did, which again, I think it's fun research, it's like let's mess with this. Uh so, let's, you know, freeze this matrix right before we get to that final layer where we're going to output the next word. And let's change the If we know what these some of the annotations are, what if we change them? Like one of the things they did with that example I just talked about, I'll bring it back up here, right? Is they they replaced the numbers that corresponded with Mars in that uh matrix of the the the annotation vectors that hang off each token. And they replaced it with the sequence of values that they had discovered **[18:36]** corresponds with Earth. So, now, even though the the prompt, you know, is the color of the fourth planet from the sun is, when it gets to that when it got to the final layer to output a word, it was looking for words that make sense grammatically and then its annotation said Earth and color, so it output blue. So, they were showing yeah hey yeah, these an- these conceptual annotations really do influence the the the word or part of word that are output. Another thing they did is they called it I I I **[19:06]** really don't like the anthropomorphize in here. Um they they said they ablated some of the information that was in the the sort of key annotations. That just means they zeroed out the values. And what they found there, a bladed by the way is like a procedure you use in like a an actual human or animal nervous system where you you use heat or electricity to sort of fry an a nervous connection, which is anthropomorphizing, let's put that aside. But they would, you know, in this example, right before we got to the final layer, they take those numbers that together correspond to Mars and **[19:37]** let's like zero them out. And what they got in those instances is you would still get grammatically correct outputs, but they weren't semantically connected anymore. So they would just be arbitrary colors, which would be equally likely. You get a bunch of colors, but nothing that corresponded to Mars in particular. Um so it shows, right? I mean, you kind of get this very rough sense about how these things are working. It's like some combination of like of the grammatically syntactically correct next thing we can do, which one should we choose? And we have all these annotations that help us narrow that down to be semantically **[20:08]** correct. That's oversimplifying it, but something like that behavior emerged. So that's, you know, I think that's that's interesting research. All right, so what does this mean though? Is this um is it scary? Is it not? Is it interesting? Is it breakthrough? Like how do we think about this? Well, there's two thoughts I think we have to keep chewing our mind at the same time. One, it is interesting research. Is it interesting because they're the first people to be able to like look at these embedding vectors and use the Jacobian **[20:40]** to figure out patterns of values that are important for the influential for the final output to sort of J lines approach. Turns out they're not the first ones to do that. So it's not like they had a breakthrough idea about doing this. I want to I'll bring up a tweet here. This is from a U Illinois professor, Yu Zhaozhang, who said, "J space is really something we've been exploring since 2022. Glad to see it continues to work well at scale. Some of the related work along this direction, she lists three papers um and then she lists **[21:12]** three more papers. The point being she was kind of responding uh to the Anthropic paper like we've been doing this for a while with uh neural networks. This isn't new, but no one's ever done it on a massive language model like Claude before because the only people who have access to the innards of a massive data uh language model like Claude are the actual companies themselves. So so that's why uh she said, "It's great to see this being done at scale." The idea wasn't new, but so it's uh anyways, I think they executed it probably well. I **[21:43]** you know, again, these are not I always put um quotation marks around reports because these are not computer scientists. These aren't formal computer science papers. These are they're more press release-y. They describe things relatively high-level with lots of pretty graphs and animation, but we can't really get into the guts of what's really happening here. But anyways, it looks like, "Hey, we took this idea and we we applied it to these massive models people are using and that was interesting." So that's idea number one to keep in your head. But idea number two the way that Anthropic described these results and the way that they were echoed after Anthropic described them, I think is incredibly disingenuous. **[22:15]** Cuz once we understand how a large language model roughly works, like I just explained, we see the thing they were describing with their jailens is exactly how we've always understood large language models to work. I mean, I went back. I I wrote a long article explaining large language models for the New Yorker in 2023. I used exactly that analogy of adding these annotations of the um high-level concepts and annotations built on other annotations, then you use that to help figure out the word. 2023, 2024, I wrote **[22:48]** another long article about the architecture for the New Yorker of the architecture of language models. Again, tables of scholars adding the annotations, which then allow you to narrow down conceptually the next token to output in a way that has some semantic meaning. This is just how we understand large language models to work. It's also how we understand essentially any deep learning neural network to work. The original breakthrough work in deep learning was image recognition. Like Jan LeCun figuring out that you could recognize handwriting better with a multi-layer neural network than **[23:19]** you could with other types of approaches. And again, the way we always understood those to work is that like different layers were picking up different features or annotation and they were then combining those different features later layers to try to figure out and make a good guess of what the letter was in a very robust way. This idea of we're picking up higher-level features and descriptions of what we're and and of of what we're analyzing, this is just how deep networks work and it's how I always understood large language models. So, I see this as like, yeah, good. We saw with the J lines exactly **[23:51]** what we expected to find. This is how we assumed large language models what was going, you know, this is what's happening in those embedding vectors. They're they're accumulating annotations that build on each other that capture higher-level conceptual meanings that help influence the token you put out. And if you change the annotations you get different tokens. How else would this work? But if you read the press release, it it's all sorts of anthropomorphizing and implication and it generation. It talks about the global workspace model of human consciousness. It somehow implies that this is seems kind of **[24:22]** similar to that. It uses these loaded terms about pondering and puzzling and it's thinking through these things and it's silently thoughts and it's not things it's saying out loud and it makes it seem like this is all icky and human-like and new and it is what any LLM researcher you would talk to for the last 5 years would say, yeah, that's how LLMs work. I mean, it's oversimplifying it to talk about it like annotations, but like that's how they work. There's nothing new and surprising here. I mean, of course the annotations are picking up **[24:53]** high-level features of the the input to help figure out what to output next. That's how deep learning networks work. That's why you have multiple layers to get all sorts of different levels of abstraction of understanding different parts of your input. And so it's cool research, but research that most people shouldn't care about. And it certainly doesn't imply what the the the implicit and I again They're careful in the paper. They're like, "Well, it doesn't really mean it's conscious. We don't say that." But it's **[25:25]** interesting. I mean, look, I'm going to load this up. Let's Let's go back to this. Let's go back to the paper for a second. I don't I don't want to go on to too far here. But again, like let's look at these things. Um you know, the J space it operates silently in the models internal network activations allowing the model to think about a concept without writing it down. What does that mean it's thinking about a concept without writing it down? The way deep networks work is that you figure out different features which help you understand other features which you combined in ways that you you know, **[25:55]** through neural network circuits that you you learn during training that then help you like put out the right recognition or generate the right thing. It's not thinking. What does it mean for it to write it down? It is writing it down. It's putting That's how it writes things down. It has this matrix that it updates the numbers every time. That's its workspace. That's its work pad. That's how these type of neural networks work. Um I'm going to jump forward here to the conclusion. Da da da. All right, so here we go. It's **[26:27]** like, "What about consciousness?" And it's like, "Well, we we borrowed a lot of ideas from the study of consciousness. Many of our experiments were designed to test for connections between the J space and global workspace theory, a framework for explaining how conscious access works in humans and animals. Given those connections, it's natural to ask whether these experiments provide evidence that AI models like Claude might be conscious. Well, our experiments don't show that Claude can have experiences or feel things. You know, it's like, but dot dot dot. That's all very suggestive that there's something new going on here. **[26:58]** Probably the most, I think, disingenuous thing at all is the way that they punched in the intro. Claude did this on its own. We didn't program it. It's a machine learned neural network. Nothing is programmed. That's That's what machine learning is. You could say that about any machine learning system. Man, this this uh this image recognizer is figuring out how to recognize it. We didn't tell it how to recognize the images. It just figured it. Yeah, cuz it was machine learning. You did **[27:29]** semi-supervised training that trained the neural network's weights until it was good at minimizing loss on the image recognition task. That's what machine learning is. You don't program it. It learns. So, I that I just think that I just think the language around this is so disingenuous. They took an existing tool that people had been doing for the last 4 or 5 years, and it's a cool tool, and they they showed it could work on scale in very big networks, and it It's cool, and it confirmed like, yeah, the way we thought these things work is that you have useful like high-level properties are uh identified as you move through these **[27:59]** layers, and and then um those are used to help, you know, narrow down the choice of possible tokens to output. Yeah, good. It works like we thought it did. Nothing is new, and nothing has to do with human consciousness. Global workspace theory, first of all, has been largely um at least as controversial, but the thing they're not saying and that all the people talking about LLMs and consciousness are not saying is when you think about global workspace theory and human consciousness, it's a center that's integrating ongoing inputs that are coming in and out. It has all sorts of different things coming in, and it's selecting different things to pay attention to and not, and it's this evolving stateful system. This is all **[28:31]** feed for Remember, this is feed forward, one layer after another in order, and then it's done, and the token is out. Nothing is saved. There's no a no states that are changing. That network had That neural network nothing has changed at it at all. And And you can feed it another prompt, and you'll get out another token. And so they're just looking at the annotations that build up and they're like at some point towards the end we see some of these annotations describe high-level properties of the type of thing that it's trying to output a token for. That's how LLMs, at least my understanding, are supposed to work. So, um look, I want to **[29:01]** I will say this, I think there's good researchers at Anthropic and this is good research. I think the PR people who talk about this technology do so in a way that is, I think, disingenuous. I think it was with a particular agenda for trying to make people feel in a certain way, which is mainly just a general sense of ick, like wow, this stuff's too powerful. Why? Because if this stuff seems like weird and alive and emergent, it's such an important technology and we so worry about that that we'll forget to ask, "Hey Anthropic, how are you going to make a profit?" Hey Anthropic, **[29:33]** your token cost is this high. You have no competitive mode on the small number of applications where people are willing to actually pay for token API access, smaller, more specialized models with, you know, hand-coded, complicated harnesses are going to do just as well. What's your plan? How do you justify a trillion dollar valuation? Like these sort of real questions are the ones you forget to ask when you instead are trying to wonder, "Huh, does this mean Claude is a moral agent or conscious?" **[30:04]** Uh so I guess my final thing would be to tell the researchers at Anthropic, I like your work and this is good work. I wish they would let you write real computer science papers and not these glorified press releases, but I know they don't let you. But maybe next time when the PR department calls, wait a second before you pick up the phone because I do not like the way Anthropic talks about the research even if there's cool stuff going on in there. So, back to the original question, did we just reveal something unnerving, cool, or blockbuster about LLMs with this research? And the answer is no. We confirmed that yes, language models work the way that at least I always thought **[30:34]** they did and you can check the print because I've written about this for years. Um so Anthropic, will you stop with the the press release research reports? Write computer science papers, sell products, convince us to buy the products. This weird in-between kind of bastardization of actual research is something that's stressing a lot of people out, and I think it's disingenuous, but that's just my opinion. All right, that's all the time we have for this week. I'll be back on Monday with an advice episode, probably no AI reality check the next week cuz I'm on vacation, but I'll do my best, and we'll be back soon enough. And until then, as always, remember care about AI, **[31:05]** but not everything you read about it. Hey, if you made it this far, you must be ready to join my fight for depth in a distracted world. Now, the best way to do this is to join over 125,000 people who receive my email newsletter each Monday. You can sign up at calnewport.com/ideas, [music] and when you do, I will send you a free guide to my seven best ideas about cultivating a deep life. Sign up today calnewport.com/ideas.