[{"data":1,"prerenderedAt":304},["ShallowReactive",2],{"project-42495936":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":27,"sentiment":28,"stance_mix":29,"scored_comments":34,"aspects":35,"strip":43,"post_text":45,"what":46,"said":49,"maker_in_thread":114,"shape":125,"grid":130,"attempts":265,"same_week":266,"maker_shows":40,"maker_hidden":296,"reader_version":297,"card":298},"42495936","Goodfire","Llama 3.3 70B Sparse Autoencoders with API access","title","Show HN: Llama 3.3 70B Sparse Autoencoders with API access","https:\u002F\u002Fwww.goodfire.ai\u002Fpapers\u002Fmapping-latent-spaces-llama\u002F","goodfire.ai","trq_",1734974297,2024,201,51,393,3515,12,"notable","ai","sparse autoencoders","web_app",[21,24,25,26],"llama 3.3","mechanistic interpretability","llm features","curious",-0.125,{"positive":30,"neutral":31,"mixed":32,"critical":33},3,8,0,4,15,{"execution":36,"docs":38,"idea":39,"design":41,"privacy":42},{"positive":32,"critical":37},1,{"positive":32,"critical":32},{"positive":32,"critical":40},2,{"positive":37,"critical":32},{"positive":37,"critical":37},[30,30,44,32,32,32,32,32,32,32,32,32,32,30,30,30,32,32,44,37,32,37,30,37,37,44,32,32,32,32,32,32,32,32,32,32,32,32,32,32],5,null,{"problem":45,"audience":45,"pricing":45,"motivation":45,"time_to_build":45,"stack":47,"openness":48,"stage":48,"team_size":48},[],"unknown",{"praise":50,"objections":66,"requests":88,"questions":100,"alternatives":109},[51,59],{"value":52,"source":53,"evidence":54},"Zoomable UMAP visualisation is nice","commenter",[55],{"id":56,"by":57,"text":58},"42496667","swyx","nice work. enjoyed the zoomable UMAP. i wonder if there are hparams to recluster the UMAP in interesting ways. after the idea that Claude 3.5 Sonnet used SAEs to improve its…",{"value":60,"source":53,"evidence":61},"OpenAI-compatible sampling API",[62],{"id":63,"by":64,"text":65},"42596235","detente18","Should we expect the current API (https:\u002F\u002Fdocs.goodfire.ai\u002Fintroduction) to be good for prod or is it for testing only? I see that the sampling API is OpenAI-compatible (nice!).…",[67,74,81],{"value":68,"source":53,"evidence":69},"Skepticism about generic sparsification efforts",[70],{"id":71,"by":72,"text":73},"42497032","owenthejumper","I am skeptical of generic sparsification efforts. After all, companies like Neural Magic spent years trying to make it work, only to pivot to 'vLLM' engine and be sold to Red Hat",{"value":75,"source":53,"evidence":76},"Login issue: app logs out after first click",[77],{"id":78,"by":79,"text":80},"42499619","Inviz","The app keeps logging me out after first click. The tech seems to be intriguiging for me as a software engineer looking to get into custom llm stuff",{"value":82,"source":53,"evidence":83},"Concern about sending data to a foreign service",[84],{"id":85,"by":86,"text":87},"42496615","I_am_tiberius","I wonder how many people or companies choose to send their data to foreign services for analysis. Personally, I would approach this with caution and am curious to see how this…",[89,93],{"value":90,"source":53,"evidence":91},"Native LiteLLM integration with a goodfire provider route",[92],{"id":63,"by":64,"text":65},{"value":94,"source":53,"evidence":95},"Try PaCMAP for the UMAP view",[96],{"id":97,"by":98,"text":99},"42499812","bravura","I'd be really curious to see what happens if you use PaCMAP (https:\u002F\u002Fjmlr.org\u002Fpapers\u002Fvolume22\u002F20-1061\u002F20-1061.pdf) and more recent large-scale variants…",[101,105],{"value":102,"source":53,"evidence":103},"Is the API meant for production or testing only?",[104],{"id":63,"by":64,"text":65},{"value":106,"source":53,"evidence":107},"Any practical use of SAEs beyond Golden Gate Claude?",[108],{"id":56,"by":57,"text":58},[110],{"value":111,"source":53,"evidence":112},"Anthropic's SAE API",[113],{"id":56,"by":57,"text":58},{"comments":30,"answered":40,"key":115},[116,119,122],{"id":117,"text":118},"42499262","Hmm the hallucination would happen in the auto labelling, but we review and test our labels and they seem correct!",{"id":120,"text":121},"42496745","If you're hacking on this and have questions, please join us on Discord: https:\u002F\u002Fdiscord.gg\u002FvhT9Chrt",{"id":123,"text":124},"42496731","We haven't yet found generalizable \"make this model smarter\" features, but there is a tradeoff of putting instructions in system prompts, e.g. if you have a chatbot that…",{"commenters":126,"max_depth":127,"maker_replies":30,"first_comment_mins":128,"lifespan_hours":129},30,7,41,288,[131,135,139,141,145,148,152,154,157,160,162,164,166,169,172,175,178,181,185,187,191,192,193,195,197,200,201,203,206,208,210,213,216,221,224,227,229,231,234,237,239,242,244,246,248,250,252,255,257,259,262],[132,32,30,133,134,32],"42496695","project\u002Fneutral\u002Fexecution","paraschopra",[136,32,30,137,138,32],"42496800","maker\u002Fneutral\u002Fexecution","tMcGrath",[63,37,44,140,64,32],"project\u002Fpositive\u002Fnone",[142,40,32,143,144,32],"42501046","other\u002Fneutral\u002Fnone","lukeramsden",[146,40,32,143,147,32],"42501195","stavros",[149,40,32,150,151,32],"42504922","commenter\u002Fneutral\u002Fnone","ffsm8",[153,40,32,150,147,32],"42504933",[155,40,32,143,156,32],"42501531","UltraSane",[158,40,32,143,159,32],"42502410","SXX",[161,40,32,150,156,32],"42506116",[163,40,32,150,159,32],"42507467",[165,40,32,150,156,32],"42508398",[167,40,32,143,168,32],"42501546","ben_w",[170,40,32,143,171,32],"42501548","cornholio",[173,40,32,143,174,32],"42503175","unshavedyak",[176,40,32,143,177,32],"42503674","dlmotol",[179,30,30,180,138,32],"42496250","maker\u002Fneutral\u002Fnone",[182,30,30,183,184,32],"42496517","maker\u002Fneutral\u002Fdocs","goldemerald",[186,30,30,180,138,32],"42496647",[188,30,30,189,190,32],"42497326","project\u002Fneutral\u002Fidea","wg0",[117,30,32,180,11,37],[120,33,32,180,11,37],[56,44,44,194,57,32],"project\u002Fpositive\u002Fdesign",[196,44,30,189,138,32],"42496725",[198,44,37,199,57,32],"42498652","project\u002Fcritical\u002Fidea",[123,44,32,189,11,37],[71,202,37,199,72,32],6,[204,202,32,150,205,32],"42497832","refulgentis",[97,127,30,207,98,32],"project\u002Fneutral\u002Fdesign",[78,31,37,209,79,32],"project\u002Fcritical\u002Fexecution",[85,211,37,212,86,32],9,"project\u002Fcritical\u002Fprivacy",[214,211,44,215,138,32],"42496666","maker\u002Fpositive\u002Fprivacy",[217,218,32,219,220,32],"42497298",10,"other\u002Fcritical\u002Fnone","ed",[222,218,32,143,223,32],"42497386","echelon",[225,218,32,143,226,32],"42497720","Steen3S",[228,218,32,143,220,32],"42497724",[230,218,32,143,223,32],"42500040",[232,218,32,143,233,32],"42498141","pennomi",[235,218,32,143,236,32],"42498833","imiric",[238,218,32,143,223,32],"42499972",[240,218,32,150,241,32],"42500165","cbg0",[243,218,32,150,223,32],"42500174",[245,218,32,143,168,32],"42502566",[247,218,32,143,236,32],"42501132",[249,218,32,143,223,32],"42502399",[251,218,32,143,236,32],"42504757",[253,218,32,143,254,32],"42499889","Rastonbury",[256,218,32,150,223,32],"42500009",[258,218,32,143,168,32],"42502595",[260,218,32,143,261,32],"42501913","weberer",[263,218,32,150,264,32],"42502890","Philpax",[],[267,272,277,282,287,292],{"id":268,"name":269,"points":270,"category":271},"42507116","PaperMatch",324,"languages",{"id":273,"name":274,"points":275,"category":276},"42506569","FixBrowser",320,"dev_tools",{"id":278,"name":279,"points":280,"category":281},"42471913","Artemis",312,"productivity",{"id":283,"name":284,"points":285,"category":286},"42497173","isle",251,"games",{"id":288,"name":289,"points":290,"category":291},"42493832","Super Snowflake Maker",228,"creative",{"id":293,"name":294,"points":295,"category":276},"42464310","Mizu.js",225,false,"v2",{"id":299,"name":5,"line":6,"ai":296,"cat":20,"year":13,"points":14,"comments":15,"rec":27,"s":300,"top":18,"kw":301,"maker":11,"domain":10,"tier":19,"attempts":32,"strip":302,"url":9,"mix":303,"praise":52,"objection":68,"sub":21},42495936,-0.12,[21,24,25,26],"3350000000000333005101311500000000000000",[30,31,32,33],1794036666231]