[{"data":1,"prerenderedAt":209},["ShallowReactive",2],{"project-40623883":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":35,"aspects":36,"strip":40,"post_text":43,"what":44,"said":69,"maker_in_thread":122,"shape":129,"grid":132,"attempts":170,"same_week":171,"maker_shows":34,"maker_hidden":201,"reader_version":202,"card":203},"40623883","Inspectus","We've open-sourced our LLM attention visualization library","title","Show HN: We've open-sourced our LLM attention visualization library","https:\u002F\u002Fgithub.com\u002Flabmlai\u002Finspectus","github.com","lakshith-403",1717934753,2024,197,15,400,3515,12,"notable","ai","model interpretability visualization","library",[24,25,26,27],"attention visualization","transformer interpretability","jupyter notebook","llm debugging","curious",0.105,{"positive":31,"neutral":32,"mixed":33,"critical":34},2,7,0,1,10,{"design":37,"idea":38,"docs":39},{"positive":34,"critical":33},{"positive":33,"critical":33},{"positive":33,"critical":34},[41,41,41,42,33,33,41,41,41,33,33,41,33,34,42],3,5,"Inspectus allows you to create interactive visualizations of attention matrices with just a few lines of Python code. It’s designed to run smoothly in Jupyter notebooks through an easy-to-use Python API. Inspectus provides multiple views to help you understand language model behaviors. If you have any questions, feel free to ask!",{"problem":45,"audience":54,"pricing":54,"motivation":55,"time_to_build":54,"stack":61,"openness":66,"stage":67,"team_size":68},{"value":46,"source":47,"evidence":48},"Understanding how language models attend to different parts of their input","maker",[49,51],{"id":50},"post",{"id":52,"by":11,"text":53},"40626567","Interesting. I think OpenAI here uses sparse autoencoders to map out sparse activation patterns in networks. Comparing them to how a real person reasons about a situations.…",null,{"value":56,"source":47,"evidence":57},"Open-sourcing a cleaned-up existing codebase",[58],{"id":59,"by":11,"text":60},"40626382","I'm not a primary user. Just cleaned up the existing codebase to make it open source. But you could use this to visualise attentions and debug the model. For an example if you're…",[62],{"value":63,"source":47,"evidence":64},"Python",[65],{"id":50},"open_source","launched","unknown",{"praise":70,"objections":80,"requests":92,"questions":101,"alternatives":106},[71],{"value":72,"source":73,"evidence":74},"Useful for understanding model behavior and debugging","commenter",[75,79],{"id":76,"by":77,"text":78},"40628448","JackYoustra","Hey! This is pretty neat, it reminds me of the graphs made by transformer_lens. Cool to see all of these visualization libraries popping up!",{"id":59,"by":11,"text":60},[81],{"value":82,"source":73,"evidence":83},"Documentation and usage guidance are lacking",[84,88],{"id":85,"by":86,"text":87},"40631871","3abiton","The issue with the ambiguity of usage plague lots of OSS projects. Guides\u002FTutorials will always help drive usage much more, just look at the usage of GPT-3 vs ChatGPT (which is…",{"id":89,"by":90,"text":91},"40625250","benf76","This looks cool but can you explain how to make it useful?",[93],{"value":94,"source":73,"evidence":95},"Walkthrough tutorial or guide for non-engineers",[96,100],{"id":97,"by":98,"text":99},"40625255","ravjo","Sounds great. Non-engineer, but curious. Is there a walkthrough blog post or video that can help someone appreciate\u002Funderstand this easily?",{"id":85,"by":86,"text":87},[102],{"value":103,"source":73,"evidence":104},"How to make it useful in practice",[105],{"id":89,"by":90,"text":91},[107,111],{"value":108,"source":73,"evidence":109},"transformer_lens visualizations",[110],{"id":76,"by":77,"text":78},{"value":112,"source":73,"evidence":113},"OpenAI and Anthropic sparse autoencoder concept extraction research",[114,118],{"id":115,"by":116,"text":117},"40625839","SushiHippie","This seems to be what Anthropic and OpenAI did in their research Golden Gate Claude - https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=40459543 - (60 comments, 16 days ago) Extracting…",{"id":119,"by":120,"text":121},"40629090","dimatura","That OpenAI work is more elaborate. It trains an additional network in such a way that it encodes what GPT is doing in terms of activations, but in a more interpretable way…",{"comments":41,"answered":41,"key":123},[124,125,128],{"id":52,"text":53},{"id":126,"text":127},"40626569","Thank you",{"id":59,"text":60},{"commenters":18,"max_depth":41,"maker_replies":41,"first_comment_mins":130,"lifespan_hours":131},224,22,[133,137,141,143,147,149,151,152,154,157,159,163,164,165,167],[134,33,41,135,136,33],"40630178","project\u002Fneutral\u002Fnone","xcodevn",[138,33,41,139,140,33],"40630465","project\u002Fneutral\u002Fdesign","vpj",[142,33,41,139,136,33],"40630564",[144,33,42,145,146,33],"40630677","project\u002Fpositive\u002Fnone","skulk",[115,34,33,148,116,33],"other\u002Fneutral\u002Fnone",[52,34,33,150,11,34],"project\u002Fneutral\u002Fidea",[119,34,41,150,120,33],[97,31,41,153,98,33],"project\u002Fneutral\u002Fdocs",[155,31,41,153,156,33],"40626487","swifthesitation",[126,31,33,158,11,34],"maker\u002Fpositive\u002Fdocs",[160,31,33,161,162,33],"40631999","other\u002Fneutral\u002Fdocs","blackbear_",[89,41,41,153,90,33],[59,41,33,153,11,34],[85,41,34,166,86,33],"project\u002Fcritical\u002Fdocs",[76,168,42,169,77,33],4,"project\u002Fpositive\u002Fdesign",[],[172,177,182,187,192,197],{"id":173,"name":174,"points":175,"category":176},"40632773","at.hn",459,"consumer_social",{"id":178,"name":179,"points":180,"category":181},"40646741","Revideo",298,"dev_tools",{"id":183,"name":184,"points":185,"category":186},"40635397","Probabilistic Tic-Tac-Toe",293,"games",{"id":188,"name":189,"points":190,"category":191},"40645743","Unforget",274,"productivity",{"id":193,"name":194,"points":195,"category":196},"40607889","sol-mate-eink",229,"hardware",{"id":198,"name":199,"points":200,"category":181},"40632533","mpa-archive",223,false,"v2",{"id":204,"name":5,"line":6,"ai":201,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":205,"top":18,"kw":206,"maker":11,"domain":10,"tier":19,"attempts":33,"strip":207,"url":9,"mix":208,"praise":72,"objection":82,"sub":21},40623883,0.11,[24,25,26,27],"333500333003015",[31,32,33,34],1793991093398]