--- title: Anthropic Found Something That Shouldn't Exist | daily.dev description: Anthropic researchers discovered that AI language models internally develop geometric structures analogous to biological place cells in animal brains. When... canonical: https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic Found Something That Shouldn't Exist | daily.dev og:description: Anthropic researchers discovered that AI language models internally develop geometric structures analogous to biological place cells in animal brains. When... og:url: https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd og:image: https://api.daily.dev/og/posts/FkVwCxdKD.png og:image:alt: Anthropic Found Something That Shouldn't Exist og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic Found Something That Shouldn't Exist **[Two Minute Papers](https://daily.dev/sources/twoninutepapers)** · 6 min read · 4 upvotes · 0 comments ## Summary Anthropic researchers discovered that AI language models internally develop geometric structures analogous to biological place cells in animal brains. When tasked with judging character counts and line lengths — something never explicitly taught — models spontaneously build low-dimensional curved manifolds (spirals) to represent positional information. These spiral counters space out numerical representations like radio stations to avoid interference, improving reliability. The finding suggests AI systems invent internal tools from scratch during training, and opens a window into mechanistic interpretability — understanding what's actually happening inside neural networks. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How do language models count characters or judge line length if they don't receive character count as input? Language models learn to approximate character counts by counting tokens and multiplying by roughly four characters per token, similar to estimating a shelf of books by assuming each is about an inch thick. This counting mechanism, along with a sense of how far along a line or page the model is, emerges spontaneously during training without being explicitly designed by engineers. _Track how daily.dev surfaces interpretability research for those digging into why LLMs behave the way they do._ ### What internal structure did Anthropic researchers find models use to represent numbers reliably? Researchers found models represent numbers using a spiral-shaped internal structure rather than a single linear counter, spacing out numeric values so they don't bleed into each other, much like spacing out radio stations to avoid signal interference. This spiral emerged during training without being explicitly programmed, and it appears to make numeric representation more reliable. _For engineers evaluating model reliability, daily.dev tracks emerging interpretability findings like this._ ### What are place cells and boundary cells, and why are they relevant to AI interpretability research? Place cells and boundary cells are neurons discovered in mouse brains that fire only when the animal is in a specific location or near a wall, functioning like a built-in GPS. Anthropic researchers found language models develop analogous neuron-like features that activate based on position along a line or proximity to the end of a page, suggesting parallels between biological and artificial spatial representations. _Readers comparing neuroscience and AI interpretability research can follow developments like this on daily.dev._ ## Similar posts on daily.dev - [Meet the new biologists treating LLMs like aliens](https://daily.dev/posts/meet-the-new-biologists-treating-llms-like-aliens-i8raaa1kt) · MIT Technology Review · 1 upvotes · 0 comments - [The Case That A.I. Is Thinking](https://daily.dev/posts/the-case-that-a-i-is-thinking-nvqecepwm) · Hacker News · 1 upvotes · 0 comments - [How Major Reasoning Models Converge to the Same “Brain” as They Model Reality Increasingly Better](https://daily.dev/posts/how-major-reasoning-models-converge-to-the-same-brain-as-they-model-reality-increasingly-better-rgdeaj23l) · Towards Data Science · 2 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm), [#anthropic](https://daily.dev/tags/anthropic), [#neural-networks](https://daily.dev/tags/neural-networks) [View this post on daily.dev](https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic Found Something That Shouldn't Exist","url":"https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd"},"datePublished":"2026-07-15T14:09:23.175Z","dateModified":"2026-09-13T20:53:34.317Z","description":"Anthropic researchers discovered that AI language models internally develop geometric structures analogous to biological place cells in animal brains. When...","image":"https://i.ytimg.com/vi/0CqLVnx-2UM/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/0CqLVnx-2UM/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Two Minute Papers","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Two Minute Papers","logo":"https://media.daily.dev/image/upload/s--8KXsSJ5q--/f_auto/v1711188923/logos/twoninutepapers","url":"https://daily.dev/sources/twoninutepapers"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm,anthropic,neural-networks","timeRequired":"PT6M","video":{"@type":"VideoObject","name":"Anthropic Found Something That Shouldn't Exist","description":"Anthropic researchers discovered that AI language models internally develop geometric structures analogous to biological place cells in animal brains. When...","thumbnailUrl":"https://i.ytimg.com/vi/0CqLVnx-2UM/sddefault.jpg","uploadDate":"2026-07-15T14:09:23.175Z","duration":"PT6M","url":"https://api.daily.dev/r/FkVwCxdKD","embedUrl":"https://www.youtube.com/embed/0CqLVnx-2UM"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Two Minute Papers","item":"https://daily.dev/sources/twoninutepapers"},{"@type":"ListItem","position":3,"name":"Anthropic Found Something That Shouldn't Exist"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-found-something-that-shouldn-t-exist-fkvwcxdkd#faq","mainEntity":[{"@type":"Question","name":"How do language models count characters or judge line length if they don't receive character count as input?","acceptedAnswer":{"@type":"Answer","text":"Language models learn to approximate character counts by counting tokens and multiplying by roughly four characters per token, similar to estimating a shelf of books by assuming each is about an inch thick. This counting mechanism, along with a sense of how far along a line or page the model is, emerges spontaneously during training without being explicitly designed by engineers. Track how daily.dev surfaces interpretability research for those digging into why LLMs behave the way they do."}},{"@type":"Question","name":"What internal structure did Anthropic researchers find models use to represent numbers reliably?","acceptedAnswer":{"@type":"Answer","text":"Researchers found models represent numbers using a spiral-shaped internal structure rather than a single linear counter, spacing out numeric values so they don't bleed into each other, much like spacing out radio stations to avoid signal interference. This spiral emerged during training without being explicitly programmed, and it appears to make numeric representation more reliable. For engineers evaluating model reliability, daily.dev tracks emerging interpretability findings like this."}},{"@type":"Question","name":"What are place cells and boundary cells, and why are they relevant to AI interpretability research?","acceptedAnswer":{"@type":"Answer","text":"Place cells and boundary cells are neurons discovered in mouse brains that fire only when the animal is in a specific location or near a wall, functioning like a built-in GPS. Anthropic researchers found language models develop analogous neuron-like features that activate based on position along a line or proximity to the end of a page, suggesting parallels between biological and artificial spatial representations. Readers comparing neuroscience and AI interpretability research can follow developments like this on daily.dev."}}]} ```