# Anthropic's New "Research" Report is Dumb. — Cal Newport ## 🗂️ Triage — annotated highlights surfaced > [!note]- Cross Reference > - **Anthropic's finding matches the pre-existing account of how deep networks work** — the load-bearing novelty rebuttal; a confirmatory result framed as a discovery. Connects to [[A Trusted Intermediary Makes a False Claim Durable]]. #xref > - **The paper's "thinking without writing it down" describes normal activation updates** — a template for reading lab communications critically. #xref ## 🧭 Bib capsule Cal Newport, "Anthropic's New 'Research' Report is Dumb," YouTube, 2026 — https://www.youtube.com/watch?v=m6ucYXJyjAk ## 🧠 Author's thesis (Adler Q1+Q2) Newport argues that Anthropic's interpretability report — decoding human-readable concepts out of Claude's activations and describing it with language like "pondering" and "on its own" — is competent but non-novel work dressed in anthropomorphic framing. The finding matches the standard, years-old account of how transformers and deep networks work; the framing is hedged just enough to avoid a false claim while manufacturing a sense of novelty and "digital ick." He ties the hype to distraction from harder questions about Anthropic's unit economics and moat. ## 🧩 Key arguments - The interpretable "annotations" that narrow output tokens are how deep networks were already understood to work; nothing here is new capability. - Loaded, consciousness-adjacent language ("thinking," "on its own," Global Workspace comparisons) manufactures novelty the result doesn't support. - Transformer inference is stateless feedforward, so the Global Workspace Theory comparison — which needs stateful integration — doesn't hold. ## 🗒️ Notes — Literature atoms ### A confirmatory finding matches the pre-existing account of deep networks > [!quote] > 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. - **Where**: [00:23:00](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1380) - **Why**: The core novelty rebuttal the whole critique builds on — a result that confirms the prior model isn't a discovery. ### "Thinking without writing it down" describes normal activation updates > [!quote] > 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. - **Where**: [00:26:08](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1568) - **Why**: A quotable rebuttal of a specific misleading phrase; a template for reading AI-lab framing critically. ### Transformer inference is feedforward, not stateful integration > [!quote] > Remember, this is feed forward, one layer after another in order, and then it's done, and the token is out. Nothing is saved. - **Where**: [00:28:31](https://www.youtube.com/watch?v=m6ucYXJyjAk&t=1711) - **Why**: Names the concrete architectural property (statelessness) the Global Workspace consciousness comparison ignores. ## 🔖 Permanent candidates - A confirmatory finding isn't a novel finding — A result that exactly matches a well-established prior model of how a system works is evidence it works as expected, not evidence of new capability, regardless of framing. (No existing Permanent covers this general epistemics-of-research point — worth promoting.) ## Links - [[AI & the Future of Knowledge Work MOC]] — home MOC; this is about how AI capability claims get communicated and understood, the labor-and-framing side rather than operational controls - [[AI Cost & Tokenomics MOC]] — Newport's closing argument ties the anthropomorphic framing to distraction from Anthropic's unit-economics and moat questions - [[A Trusted Intermediary Makes a False Claim Durable]] — the mechanism Newport worries about: a credible lab plus amplifying press makes an over-read claim stick regardless of the hedge ## ✅ Processing checklist - [x] Claim atoms carried from deep-read analysis - [ ] Permanent candidates reviewed for promotion - [ ] Version History Diff check