[{"data":1,"prerenderedAt":495},["ShallowReactive",2],{"project-47131225":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":29,"sentiment":30,"stance_mix":31,"scored_comments":35,"aspects":36,"strip":45,"post_text":47,"what":48,"said":73,"maker_in_thread":186,"shape":198,"grid":203,"attempts":457,"same_week":458,"maker_shows":33,"maker_hidden":487,"reader_version":488,"card":489},"47131225","Steerling-8B","A language model that can explain any token it generates","title","Show HN: Steerling-8B, a language model that can explain any token it generates","https:\u002F\u002Fwww.guidelabs.ai\u002Fpost\u002Fsteerling-8b-base-model-release\u002F","guidelabs.ai","adebayoj",1771893482,2026,328,91,118,4366,3,"legend","ai","interpretable language model","other",[24,25,26,27,28],"interpretability","language model","attribution","explainable ai","concept supervision","debated",0.147,{"positive":32,"neutral":32,"mixed":33,"critical":34},16,1,12,45,{"idea":37,"execution":40,"docs":41,"design":43,"performance":44},{"positive":38,"critical":39},13,11,{"positive":33,"critical":33},{"positive":33,"critical":42},0,{"positive":33,"critical":42},{"positive":42,"critical":42},[42,42,18,46,42,46,33,33,33,33,18,46,46,18,42,42,42,42,33,18,18,46,33,46,33,18,18,46,18,46,33,18,46,18,46,33,46,33,46,46],5,null,{"problem":49,"audience":47,"pricing":47,"motivation":59,"time_to_build":47,"stack":63,"openness":70,"stage":71,"team_size":72},{"value":50,"source":51,"evidence":52},"LLM outputs cannot be explained, so users cannot audit or debug why a token was generated","maker",[53,56],{"id":54,"by":11,"text":55},"47134180","op here, I mostly agree with your comment! However, our model does more than this. For any chunk the model generates, it can answer: which concept, in the model's…",{"id":57,"by":11,"text":58},"47134228","op here. Important point, but I disagree. We see explainability\u002Finterpretability as a CORE need for AI safety. We believe you can't align\u002Faudit\u002Fdebug\u002Ffix a system that you don't…",{"value":60,"source":51,"evidence":61},"interpretability as a core need for AI safety, auditing and debugging",[62],{"id":57,"by":11,"text":58},[64],{"value":65,"source":51,"evidence":66},"sparse autoencoder style concept layer",[67],{"id":68,"by":11,"text":69},"47134594","You are missing a few things, but you got some things right. 1) The is not an SAE in the way you think. It is a combination of a supervised + unsupervised layer that is…","unknown","launched","company",{"praise":74,"objections":99,"requests":125,"questions":151,"alternatives":173},[75,91],{"value":76,"source":77,"evidence":78},"Interpretability by design is novel and promising","commenter",[79,83,87],{"id":80,"by":81,"text":82},"47136032","deepdarkforest","Just wanted to say i think most interpretability research it's just a smoke show nowadays but this is actually the first one that i think has a very serious potential. I love…",{"id":84,"by":85,"text":86},"47132471","pbmango","This is very interesting. I don't see much discussion of interpretability in day to the day discourse of AI builders. I wonder if everyone assumes it to either be solved, or to…",{"id":88,"by":89,"text":90},"47134354","7777777phil","If this decomposition actually holds, it's the first model where you could show a regulator why it produced a given output.",{"value":92,"source":77,"evidence":93},"Attribution could help debugging and regulated domains",[94,98],{"id":95,"by":96,"text":97},"47133178","umairnadeem123","the practical value here is for regulated domains. in healthcare and finance you often cant deploy a model at all unless you can explain why it made a specific decision.…",{"id":88,"by":89,"text":90},[100,111,118],{"value":101,"source":77,"evidence":102},"Attribution to Arxiv or Wikipedia says little about truth or veracity",[103,107],{"id":104,"by":105,"text":106},"47139729","killerstorm","This seems to be too coarse-grained to be useful: all sciency content will be \"analytical\" and associate with sources like ArXiv. But there might be bad, malicious articles on…",{"id":108,"by":109,"text":110},"47132635","great_psy","Maybe I’m not creative enough to see the potential, but what value does this bring ? Given the example I saw about CRISPR, what does this model give over a different, non…",{"value":112,"source":77,"evidence":113},"Concept supervision may be a form of Goodhart-style gaming without causal effect",[114],{"id":115,"by":116,"text":117},"47133560","in-silico","Either I'm missing something or this is way overstated. Steerling appears to be just a discrete diffusion model where the final hidden states are passed through a sparse…",{"value":119,"source":77,"evidence":120},"Explainability does not address the main AI safety issues",[121],{"id":122,"by":123,"text":124},"47133129","ottah","It's a neat party trick, but explainability it's not solution to any AI safety issue I care about. It's a distraction from real problems, which is everything else around the…",[126,133,140,144],{"value":127,"source":77,"evidence":128},"Release of GGUF or llama.cpp\u002Follama support",[129],{"id":130,"by":131,"text":132},"47138023","kamranjon","I'm really interested in using this but wonder if the unique architecture means that it will not be able to be converted to a GGUF and used by ollama or llama.cpp? I certainly…",{"value":134,"source":77,"evidence":135},"Published paper or algorithm details",[136],{"id":137,"by":138,"text":139},"47134987","potato-peeler","Looks very interesting. Is there a published paper\u002Farticle on your algorithm? Would like to take a dab at implementing this on my own. I could find this [0], but not sure if that…",{"value":141,"source":77,"evidence":142},"Benchmark comparison against Llama 8B",[143],{"id":95,"by":96,"text":97},{"value":145,"source":77,"evidence":146},"Concept IDs for steering",[147],{"id":148,"by":149,"text":150},"47268293","alrmrphc-atmtn","Very interesting. I was trying to play around with it, but could not find the known concept IDs to steer the model, are they yet to be released?",[152,159,166],{"value":153,"source":77,"evidence":154},"How does training data attribution work in practice",[155],{"id":156,"by":157,"text":158},"47136238","pu_pe","Looks neat and original, congrats! I don't quite grasp how to interpret the training data attribution process. For example, it seems to say that for a given sentence like \"They…",{"value":160,"source":77,"evidence":161},"Can it reduce hallucinations",[162],{"id":163,"by":164,"text":165},"47134493","michaelmrose","Can you use this to decrease hallucinations?",{"value":167,"source":77,"evidence":168},"How does it compare to SHAP",[169],{"id":170,"by":171,"text":172},"47132478","brendanashworth","Is there a reason people don't use SHAP [1] to interpret language models more often? The in-context attribution of outputs seems very similar. [1]…",[174,178],{"value":175,"source":77,"evidence":176},"SHAP",[177],{"id":170,"by":171,"text":172},{"value":179,"source":77,"evidence":180},"Sparse autoencoders and dictionary learning",[181,185],{"id":182,"by":183,"text":184},"47135060","andy12_","This seems really interesting. While Anthropic tried to use dictionary learning over an existing model to try to extract concepts, this almost feels like training the model…",{"id":115,"by":116,"text":117},{"comments":187,"answered":34,"key":188},23,[189,192,195],{"id":190,"text":191},"47134321","Completely agree with all your points! Here is what this model does: it `rewrites` the model's activations (during pre-training) into supervised + unsupervised concepts that are…",{"id":193,"text":194},"47134366","Most interpretability techniques haven't yet to be shown to be useful for everyday model pipelines. However, the field is working hard to change this.",{"id":196,"text":197},"47136360","Great questions. We weren't quite explicit about the training data attribution process. We'll discuss this in more detail in future work. We can track down which parts of the…",{"commenters":199,"max_depth":200,"maker_replies":187,"first_comment_mins":201,"lifespan_hours":202},61,8,166,238,[204,208,211,215,218,220,223,225,227,230,231,234,235,237,241,243,247,249,253,255,258,261,263,265,268,272,274,278,280,282,286,289,293,296,298,302,305,308,311,315,318,321,324,325,328,330,333,335,338,340,342,344,346,348,350,351,353,354,355,358,362,365,367,370,372,376,380,382,384,387,390,393,398,400,404,408,412,413,415,419,421,425,427,429,430,433,435,438,440,443,446,450,453],[205,42,42,206,207,42],"47137896","other\u002Fneutral\u002Fidea","msteffen",[209,42,42,206,210,42],"47138502","idiotsecant",[212,42,42,213,214,42],"47138609","other\u002Fneutral\u002Fnone","jzig",[216,42,42,213,217,42],"47141532","mitchbob",[170,33,18,219,171,42],"project\u002Fneutral\u002Fidea",[221,33,42,206,222,42],"47132939","dwohnitmok",[190,33,42,224,11,33],"project\u002Fneutral\u002Fdesign",[226,33,42,206,222,42],"47138547",[84,228,46,229,85,42],2,"project\u002Fpositive\u002Fidea",[193,228,42,206,11,33],[232,228,42,206,233,42],"47169870","yogurt-male",[156,18,46,229,157,42],[196,18,42,236,11,33],"maker\u002Fneutral\u002Fdocs",[238,18,33,239,240,42],"47136914","project\u002Fcritical\u002Fexecution","xmcqdpt2",[242,18,42,236,11,33],"47137287",[244,245,42,206,246,42],"47133075",4,"gormen",[54,245,42,248,11,33],"maker\u002Fneutral\u002Fidea",[250,245,33,251,252,42],"47134312","project\u002Fcritical\u002Fidea","codeflo",[254,245,42,236,11,33],"47134445",[256,245,33,251,257,42],"47135752","Grimblewald",[259,245,42,213,260,42],"47136025","gchamonlive",[262,245,33,251,257,42],"47136275",[264,245,42,213,260,42],"47136983",[266,245,42,267,11,33],"47137209","maker\u002Fneutral\u002Fexecution",[269,245,18,270,271,42],"47137774","project\u002Fneutral\u002Fdocs","IshKebab",[130,46,46,273,131,42],"project\u002Fpositive\u002Fexecution",[275,46,18,276,277,42],"47138142","project\u002Fneutral\u002Fexecution","monocasa",[80,279,46,229,81,42],6,[281,279,42,236,11,33],"47136154",[283,279,18,284,285,42],"47138170","project\u002Fmixed\u002Fidea","rao-v",[287,279,18,224,288,42],"47172380","giang_at_glai",[290,291,42,206,292,42],"47136148",7,"crimsonnoodle58",[294,291,42,213,295,42],"47136190","theMMaI",[297,291,42,229,11,33],"47136204",[299,291,42,300,301,42],"47139004","other\u002Fneutral\u002Fprivacy","jacquesm",[303,291,42,213,304,42],"47140596","jbonatakis",[306,291,42,213,307,42],"47140680","Imustaskforhelp",[309,291,42,213,310,42],"47136918","KingOfCoders",[312,291,42,313,314,42],"47137005","commenter\u002Fneutral\u002Fnone","lelanthran",[316,291,42,313,317,42],"47138096","wasmainiac",[319,291,42,213,320,42],"47141016","ranger_danger",[322,291,42,206,323,42],"47173602","holdmycoffee",[104,200,33,251,105,42],[326,200,18,219,327,42],"47140143","Macuyiko",[108,329,33,251,109,42],9,[331,329,46,229,332,42],"47132679","voidhorse",[334,329,18,219,109,42],"47132773",[336,329,42,313,337,42],"47133269","rickydroll",[339,329,46,229,109,42],"47133643",[341,329,42,236,11,33],"47134909",[343,329,18,219,337,42],"47136378",[345,329,42,248,11,33],"47134390",[347,329,33,251,257,42],"47135784",[148,349,18,270,149,42],10,[182,39,46,229,183,42],[352,39,42,219,11,33],"47135101",[115,34,33,251,116,42],[68,34,42,219,11,33],[356,38,18,219,357,42],"47148226","audunw",[359,360,42,213,361,42],"47136607",14,"rippeltippel",[363,360,42,213,364,42],"47137160","abcd_f",[137,366,18,236,138,42],15,[368,366,42,369,11,33],"47135108","maker\u002Fpositive\u002Fdocs",[371,366,46,369,138,42],"47136116",[373,32,46,374,375,42],"47138261","maker\u002Fpositive\u002Fdesign","schopra909",[377,378,33,251,379,42],"47136595",17,"MagicMoonlight",[88,381,46,229,89,42],18,[383,381,42,248,11,33],"47134454",[385,381,46,386,89,42],"47134537","maker\u002Fpositive\u002Fidea",[388,381,42,389,11,33],"47134729","maker\u002Fneutral\u002Fnone",[391,381,33,251,392,42],"47134763","yorwba",[394,395,18,396,397,42],"47136122",19,"maker\u002Fneutral\u002Fperformance","whinvik",[399,395,42,248,11,33],"47136981",[401,402,18,219,403,42],"47137300",20,"chmod775",[405,406,46,229,407,42],"47136557",21,"ZeroAurora",[409,410,42,213,411,42],"47138963",22,"exabrial",[163,187,18,219,164,42],[414,187,42,219,11,33],"47134911",[416,417,18,270,418,42],"47136012",24,"aziis98",[420,417,42,236,11,33],"47136373",[422,423,46,229,424,42],"47132502",25,"rvz",[426,423,42,248,11,33],"47134605",[122,428,33,251,123,42],26,[57,428,42,386,11,33],[431,428,46,229,432,42],"47134750","vintagedave",[434,428,42,229,11,33],"47134897",[436,428,18,219,437,42],"47138735","0xdeadbeefbabe",[439,428,33,251,123,42],"47142983",[441,428,46,229,442,42],"47135774","adwf",[95,444,18,445,96,42],27,"project\u002Fneutral\u002Fperformance",[447,444,42,448,449,42],"47133393","commenter\u002Fpositive\u002Fperformance","snowhale",[451,444,42,452,11,33],"47134768","maker\u002Fpositive\u002Fperformance",[454,455,46,229,456,42],"47140336",28,"johntheagent",[],[459,464,469,473,478,482],{"id":460,"name":461,"points":462,"category":463},"47075124","Micasa",657,"productivity",{"id":465,"name":466,"points":467,"category":468},"47114530","CIA World Factbook Archive",496,"learning_science",{"id":470,"name":471,"points":472,"category":20},"47104667","ntransformer",395,{"id":474,"name":475,"points":476,"category":477},"47123631","PgDog",326,"infrastructure",{"id":479,"name":480,"points":481,"category":20},"47143755","Moonshine Open-Weights STT models",316,{"id":483,"name":484,"points":485,"category":486},"47132102","X86CSS",275,"languages",false,"v2",{"id":490,"name":5,"line":6,"ai":487,"cat":20,"year":13,"points":14,"comments":15,"rec":29,"s":491,"top":18,"kw":492,"maker":11,"domain":10,"tier":19,"attempts":42,"strip":493,"url":9,"mix":494,"praise":76,"objection":101,"sub":21},47131225,0.15,[24,25,26,27,28],"0035051111355300001335151335351353515155",[32,32,33,34],1794036694148]