<\/p>\n
One niche that Anthropic spends more time and money on than other AI companies is called mechanistic interpretability, which means looking inside the complex math of an AI model to learn why it comes up with one particular output and not another. It\u2019s complicated stuff; there are millions of data points that might contribute to any result, and wading through them can look more like word salad than anything useful. It\u2019s also controversial. Describing AI models with terms borrowed from psychology and neuroscience can make their behavior seem more sophisticated than we might otherwise judge it to be.<\/p>\n
That\u2019s why, when Anthropic announced last week that it had found a new window into its models\u2019 \u201cinternal thoughts\u201d as they reason through answers, there was one colleague I had to talk to. Senior editor Will Douglas Heaven, aside from having a PhD in computer science, has spent a lot of time digging into what we can say about how AI models work. I spoke with him about what we should take from Anthropic\u2019s new (and predictably quirky) research.<\/p>\n
What did Anthropic learn here, exactly?<\/strong><\/p>\n Anthropic has been trying to understand how large language models (LLMs) work for a few years now. Anthropic isn\u2019t the only one looking at this, but I think the company has made it part of its core mission more than most. Anthropic\u2019s CEO, Dario Amodei, has said we won\u2019t be able to control LLMs fully unless we learn more about how they work.\u00a0<\/p>\n So this new research is very much in that context. It goes deeper into the weird mechanisms inside LLMs than ever before. What Anthropic learned was that LLMs have a space inside them\u2014which Anthropic calls the J-space\u2014filled with words that don\u2019t appear in their output but that seem to influence the way they puzzle through problems. All this was hidden until Anthropic developed a new technique to probe its model Claude, so it\u2019s a genuine discovery.\u00a0<\/p>\n Sometimes these words keep track of where the LLM has got to in a particular task, sometimes they look more like flashes of recognition (for example, \u201cprotein\u201d might pop up when you give an LLM only the letters of a protein sequence), and sometimes they represent a kind of internal commentary on the model\u2019s decision-making. In my favorite example, Claude decided to cheat on a coding test when the word \u201cpanic\u201d appeared.<\/p>\n Anthropic also found that LLMs are able to describe and manipulate the words in this space. So somehow they seem to be making use of it.\u00a0<\/p>\n Let\u2019s step back for a second. I don\u2019t think of large language models as <\/strong>simple<\/em><\/strong>, but they\u2019re also not magic. There\u2019s a bunch of math that learns relationships between words, right? So why is it so hard to \u201cpeer\u201d into an LLM to know what\u2019s going on?<\/strong><\/p>\n<\/p><\/div>\n #Anthropics #latest #discovery #doesand #doesntshow<\/p>\n","protected":false},"excerpt":{"rendered":" One niche that Anthropic spends more time and money on than other AI companies is…<\/p>\n","protected":false},"author":1,"featured_media":17078,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[792,5476,11710,11711,696],"class_list":["post-17077","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-stories","tag-anthropics","tag-discovery","tag-doesand","tag-doesntshow","tag-latest"],"featured_image_urls":{"full":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2.jpg",1200,600,false],"thumbnail":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-150x150.jpg",150,150,true],"medium":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-300x150.jpg",300,150,true],"medium_large":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-768x384.jpg",640,320,true],"large":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-1024x512.jpg",640,320,true],"1536x1536":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2.jpg",1200,600,false],"2048x2048":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2.jpg",1200,600,false],"covernews-featured":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-1024x512.jpg",1024,512,true],"covernews-medium":["https:\/\/news678.top\/wp-content\/uploads\/2026\/07\/thinking2-540x340.jpg",540,340,true]},"author_info":{"display_name":"admin","author_link":"https:\/\/news678.top\/?author=1"},"category_info":"Stories<\/a>","tag_info":"Stories","comment_count":"0","_links":{"self":[{"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/posts\/17077","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/news678.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=17077"}],"version-history":[{"count":0,"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/posts\/17077\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/news678.top\/index.php?rest_route=\/wp\/v2\/media\/17078"}],"wp:attachment":[{"href":"https:\/\/news678.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=17077"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/news678.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=17077"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/news678.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=17077"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}