[{"data":1,"prerenderedAt":447},["ShallowReactive",2],{"project-42208383":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":34,"aspects":35,"strip":42,"post_text":43,"what":44,"said":70,"maker_in_thread":138,"shape":149,"grid":153,"attempts":408,"same_week":409,"maker_shows":39,"maker_hidden":439,"reader_version":440,"card":441},"42208383","llama3_interpretability_sae","Llama 3.2 Interpretability with Sparse Autoencoders","title","Show HN: Llama 3.2 Interpretability with Sparse Autoencoders","https:\u002F\u002Fgithub.com\u002FPaulPauls\u002Fllama3_interpretability_sae","github.com","PaulPauls",1732221476,2024,579,99,47,3515,2,"legend","ai","mechanistic interpretability","library",[24,21,25,26,27],"sparse autoencoders","llama 3.2","dead latents","llm interpretability","loved",0.472,{"positive":31,"neutral":32,"mixed":33,"critical":18},14,12,0,28,{"idea":36,"execution":38,"docs":40},{"positive":37,"critical":18},5,{"positive":39,"critical":33},1,{"positive":41,"critical":33},3,[33,33,33,33,33,33,33,33,33,33,33,33,33,39,33,39,37,33,33,33,33,33,33,33,33,37,37,37,41,37,37,37,37,33,41,41,41,37,37,37],"I spent a lot of time and money on this rather big side project of mine that attempts to replicate the mechanistic interpretability research on proprietary LLMs that was quite popular this year and produced great research papers by Anthropic [1], OpenAI [2] and Deepmind [3].\n\nI am quite proud of this project and since I consider myself the target audience for HackerNews did I think that maybe some of you would appreciate this open research replication as well. Happy to answer any questions or face any feedback.\n\nCheers\n\n[1] https:\u002F\u002Ftransformer-circuits.pub\u002F2024\u002Fscaling-monosemanticity\u002Findex.ht",{"problem":45,"audience":51,"pricing":55,"motivation":56,"time_to_build":55,"stack":60,"openness":67,"stage":68,"team_size":69},{"value":46,"source":47,"evidence":48},"Replicates mechanistic interpretability research on proprietary LLMs using open models","maker",[49],{"id":50},"post",{"value":52,"source":47,"evidence":53},"ML researchers and HN readers interested in interpretability",[54],{"id":50},null,{"value":57,"source":47,"evidence":58},"Wanted to replicate popular interpretability research on open models",[59],{"id":50},[61],{"value":62,"source":47,"evidence":63},"PyTorch",[64],{"id":65,"by":11,"text":66},"42210325","I'm very happy you appreciate it - particularly the documentation. Writing the documentation was much harder for me than writing the code so I'm happy it is appreciated. I…","open_source","launched","solo",{"praise":71,"objections":99,"requests":107,"questions":122,"alternatives":137},[72,88],{"value":73,"source":74,"evidence":75},"Well-documented work","commenter",[76,80,84],{"id":77,"by":78,"text":79},"42208694","jaykr_","This is awesome! I really appreciate the time you took to document everything!",{"id":81,"by":82,"text":83},"42211885","enterthedragon","This is amazing, the documentation is very well organized",{"id":85,"by":86,"text":87},"42211413","Carrentt","Fantastic work! I absolutely love all the documentation.",{"value":89,"source":74,"evidence":90},"Impressive engineering effort",[91,95],{"id":92,"by":93,"text":94},"42211357","samstevens","I’m really excited to see some more open SAE work! The engineering effort is non trivial and I’m going to check out your dataloading code tomorrow. You might be interested in an…",{"id":96,"by":97,"text":98},"42212349","S-Kaenel","Amazing research!!",[100],{"value":101,"source":74,"evidence":102},"Claims about SAEs explaining model reasoning are overstated",[103],{"id":104,"by":105,"text":106},"42209839","foundry27","For anyone who hasn’t seen this before, mechanistic interpretability solves a very common problem with LLMs: when you ask a model to explain itself, you’re playing a game of…",[108,115],{"value":109,"source":74,"evidence":110},"Integration with SAELens",[111],{"id":112,"by":113,"text":114},"42209081","JackYoustra","Very cool work! Any plans to integrate it with SAELens?",{"value":116,"source":74,"evidence":117},"Golden Gate Llama demo",[118],{"id":119,"by":120,"text":121},"42212667","moconnor","Find a latent for the Golden Gate bridge and put a Golden Gate Llama 3.2 on HuggingFace. This will get even more attention and love, more so if you include link to a space to…",[123,130],{"value":124,"source":74,"evidence":125},"How to evaluate SAE quality",[126],{"id":127,"by":128,"text":129},"42209072","curious_cat_163","Hey - Thanks for sharing! Will take a closer look later but if you are hanging around now, it might be worth asking this now. I read this blog post recently:…",{"value":131,"source":74,"evidence":132},"Total cost of the project",[133],{"id":134,"by":135,"text":136},"42211493","vivekkalyan","This is great work! Mechanistic interpretability has tons of use cases, it's great to see open research in that field. You mentioned you spent your own time and money on it,…",[],{"comments":139,"answered":140,"key":141},7,4,[142,143,146],{"id":65,"text":66},{"id":144,"text":145},"42210293","So evaluating SAEs - determining which SAE is better at creating the most unique features while being as sparse as possible at the same time - is a very complex topic that is…",{"id":147,"text":148},"42210178","Not sure yet to be honest. I'll definitely consider it but I'll reorient myself and what I plan to do next in the coming week. I also planned on maybe starting a simpler project…",{"commenters":150,"max_depth":32,"maker_replies":139,"first_comment_mins":151,"lifespan_hours":152},53,33,441,[154,156,160,163,166,168,170,173,175,178,181,184,187,190,192,194,197,199,202,204,206,208,211,213,215,217,220,222,224,226,228,230,233,236,240,243,245,247,251,254,258,261,263,266,269,271,274,276,279,282,284,286,288,290,292,294,296,298,300,303,305,308,310,313,317,319,322,324,326,329,333,336,337,338,342,345,347,350,352,354,356,359,363,365,367,369,372,375,377,380,383,385,389,392,394,396,398,402,404],[104,33,33,155,105,33],"other\u002Fneutral\u002Fnone",[157,33,33,158,159,33],"42209976","commenter\u002Fneutral\u002Fnone","stavros",[161,33,33,158,162,33],"42211372","sinuhe69",[164,33,33,158,165,33],"42211546","TeMPOraL",[167,33,33,158,162,33],"42212253",[169,33,33,158,165,33],"42212620",[171,33,33,158,172,33],"42212963","worldsayshi",[174,33,33,158,165,33],"42213323",[176,33,33,158,177,33],"42212427","mdp2021",[179,33,33,158,180,33],"42217404","legel",[182,33,33,158,183,33],"42228640","HeavyStorm",[185,33,33,155,186,33],"42211403","LoganDark",[188,33,33,155,189,33],"42213336","btbuildem",[191,33,33,158,177,33],"42212158",[193,33,33,158,159,33],"42212162",[195,33,33,158,196,33],"42212340","DSingularity",[198,33,33,158,159,33],"42212366",[200,33,33,158,201,33],"42212536","f_devd",[203,33,33,158,159,33],"42212586",[205,33,33,158,201,33],"42213145",[207,33,33,158,177,33],"42212350",[209,33,33,158,210,33],"42218115","famouswaffles",[212,33,33,158,177,33],"42219971",[214,33,33,158,210,33],"42223864",[216,33,33,158,177,33],"42224709",[218,33,33,158,219,33],"42215646","catskul2",[221,33,33,158,177,33],"42215861",[223,33,33,158,210,33],"42218141",[225,33,33,158,177,33],"42220041",[227,33,33,158,210,33],"42223735",[229,33,33,158,177,33],"42224616",[231,33,33,232,177,33],"42216841","commenter\u002Fcritical\u002Fnone",[234,33,33,155,235,33],"42210215","fsndz",[237,33,39,238,239,33],"42210349","project\u002Fcritical\u002Fidea","benchmarkist",[241,33,33,158,242,33],"42210754","HeatrayEnjoyer",[244,33,33,232,239,33],"42210772",[246,33,33,155,177,33],"42212381",[248,33,41,249,250,33],"42210223","project\u002Fneutral\u002Fidea","Onavo",[252,33,39,238,253,33],"42210551","benreesman",[255,33,37,256,257,33],"42210779","project\u002Fpositive\u002Fidea","txnf",[259,33,33,158,260,33],"42210829","drdeca",[262,33,33,155,253,33],"42211243",[264,33,33,155,265,33],"42211240","snthpy",[267,33,33,155,268,33],"42211371","briffid",[270,33,33,158,265,33],"42213813",[272,33,33,155,273,33],"42211658","omgwtfbyobbq",[275,33,33,158,265,33],"42213800",[277,33,33,158,278,33],"42211705","shshshshs",[280,33,33,158,281,33],"42212022","Hedepig",[283,33,33,158,177,33],"42212205",[285,33,33,158,281,33],"42216102",[287,33,33,158,177,33],"42216879",[289,33,33,158,281,33],"42219675",[291,33,33,158,177,33],"42219903",[293,33,33,158,281,33],"42220377",[295,33,33,158,177,33],"42220894",[297,33,33,158,281,33],"42221174",[299,33,33,158,265,33],"42213867",[301,33,33,158,302,33],"42213893","fragmede",[304,33,33,158,177,33],"42212141",[306,33,33,155,307,33],"42212911","jamesemmott",[309,33,33,158,265,33],"42213808",[311,33,33,155,312,33],"42224791","bubaumba",[314,39,37,315,316,33],"42210024","project\u002Fpositive\u002Fexecution","jwuphysics",[65,39,33,318,11,39],"maker\u002Fpositive\u002Fdocs",[320,18,37,256,321,33],"42210668","Eliezer",[127,41,41,323,128,33],"project\u002Fneutral\u002Fexecution",[144,41,33,325,11,39],"project\u002Fneutral\u002Fdocs",[327,41,37,328,128,33],"42211084","maker\u002Fpositive\u002Fnone",[330,140,41,331,332,33],"42275327","project\u002Fneutral\u002Fnone","OrangeMusic",[334,140,41,331,335,33],"42373953","dimitry12",[112,37,37,256,113,33],[147,37,33,249,11,39],[339,340,41,249,341,33],"42210896",6,"monkeycantype",[343,340,37,256,344,33],"42215006","Majromax",[92,139,37,346,93,33],"project\u002Fpositive\u002Fnone",[77,348,37,349,78,33],8,"project\u002Fpositive\u002Fdocs",[351,348,33,318,11,39],"42210147",[119,353,41,249,120,33],9,[134,355,37,346,135,33],10,[357,355,33,358,11,39],"42216413","project\u002Fneutral\u002Fpricing",[360,361,33,155,362,33],"42218603",11,"westurner",[364,361,33,155,362,33],"42218619",[366,361,33,155,362,33],"42218670",[368,361,33,155,362,33],"42229295",[370,32,41,325,371,33],"42212240","imranhou",[373,32,41,249,374,33],"42212302","versteegen",[376,32,41,249,371,33],"42212343",[378,32,41,249,379,33],"42226836","dontknowit",[381,32,33,382,11,39],"42216112","maker\u002Fneutral\u002Fdocs",[384,32,41,382,371,33],"42217763",[386,387,37,256,388,33],"42213777",13,"yangwang92",[390,31,37,346,391,33],"42211228","batterylake",[393,31,33,328,11,39],"42216244",[81,395,37,349,82,33],15,[85,397,37,349,86,33],16,[399,400,41,331,401,33],"42220003",17,"coolvision",[96,403,37,346,97,33],18,[405,406,33,155,407,33],"42221745",19,"jzjsj",[],[410,415,420,425,430,434],{"id":411,"name":412,"points":413,"category":414},"42254156","Intenty",817,"productivity",{"id":416,"name":417,"points":418,"category":419},"42212644","5 Million Devs",616,"games",{"id":421,"name":422,"points":423,"category":424},"42256104","TeaTime",307,"learning_science",{"id":426,"name":427,"points":428,"category":429},"42258540","Feels Like Paper",265,"creative",{"id":431,"name":432,"points":433,"category":419},"42212071","Rebels in the sky",242,{"id":435,"name":436,"points":437,"category":438},"42200099","Monty",236,"hardware",false,"v2",{"id":442,"name":5,"line":6,"ai":439,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":443,"top":18,"kw":444,"maker":11,"domain":10,"tier":19,"attempts":33,"strip":445,"url":9,"mix":446,"praise":73,"objection":101,"sub":21},42208383,0.47,[24,21,25,26,27],"0000000000000101500000000555355550333555",[31,32,33,18],1794036664708]