[{"data":1,"prerenderedAt":140},["ShallowReactive",2],{"project-46518608":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":32,"aspects":33,"strip":35,"post_text":37,"what":38,"said":68,"maker_in_thread":86,"shape":91,"grid":94,"attempts":98,"same_week":103,"maker_shows":132,"maker_hidden":133,"reader_version":134,"card":135},"46518608","Symbolic Circuit Distillation","prove program to LLM circuit equivalence","title","Show HN: Symbolic Circuit Distillation: prove program to LLM circuit equivalence","https:\u002F\u002Fgithub.com\u002Fneelsomani\u002Fsymbolic-circuit-distillation","github.com","nsomani",1767733026,2026,16,2,2282,4366,53,"shown","learning_science","mechanistic interpretability","library",[21,24,25,26,27],"circuit distillation","smt equivalence checking","formal verification","transformer circuits","quiet",null,{"positive":31,"neutral":32,"mixed":31,"critical":31},0,1,{"docs":34},{"positive":31,"critical":31},[36,31],3,"Hi HN, I've been exploring various applications of formal methods to ML\u002Finterpretability and I've been hoping to get more eyes on the approach.\n\nI have been working on a small interpretability project I call Symbolic Circuit Distillation. The goal is to take a tiny neuron-level circuit (like the ones in OpenAI's \"Sparse Circuits\" work) and automatically recover a concise Python program that implements the same algorithm, along with a bounded formal proof that the two are equivalent on a finite token domain.\n\nRoughly, the pipeline is:\n\n1. Start from a pruned circuit graph for a specific behavio",{"problem":39,"audience":45,"pricing":29,"motivation":49,"time_to_build":29,"stack":53,"openness":66,"stage":67,"team_size":66},{"value":40,"source":41,"evidence":42},"Turning sparse neuron-level circuits into clean, human-readable algorithms with bounded equivalence proofs","maker",[43],{"id":44},"post",{"value":46,"source":41,"evidence":47},"People working on mechanistic interpretability or formal methods",[48],{"id":44},{"value":50,"source":41,"evidence":51},"Automate the manual step of turning extracted circuits into verified algorithms",[52],{"id":44},[54,58,62],{"value":55,"source":41,"evidence":56},"Python",[57],{"id":44},{"value":59,"source":41,"evidence":60},"ReLU neural network surrogate",[61],{"id":44},{"value":63,"source":41,"evidence":64},"SMT solver for bounded equivalence checking",[65],{"id":44},"unknown","prototype",{"praise":69,"objections":70,"requests":71,"questions":76,"alternatives":85},[],[],[72],{"value":73,"source":41,"evidence":74},"Suggestions for next benchmark circuits or behaviors to distill",[75],{"id":44},[77],{"value":78,"source":79,"evidence":80},"Whether the readme includes examples","commenter",[81],{"id":82,"by":83,"text":84},"46520005","aappleby","No examples in the readme?",[],{"comments":32,"answered":32,"key":87},[88],{"id":89,"text":90},"46521991","There are two examples provided - quote matching and bracket closing.",{"commenters":15,"max_depth":15,"maker_replies":32,"first_comment_mins":92,"lifespan_hours":93},117,6,[95,97],[82,31,36,96,83,31],"project\u002Fneutral\u002Fdocs",[89,31,31,96,11,32],[99,102],{"id":100,"t":101,"points":32,"comments":32},"46197485",1765227028,{"id":4,"t":12,"points":14,"comments":15},[104,109,114,118,123,127],{"id":105,"name":106,"points":107,"category":108},"46454693","OpenWorkers",500,"infrastructure",{"id":110,"name":111,"points":112,"category":113},"46491749","taws",390,"dev_tools",{"id":115,"name":116,"points":117,"category":113},"46511469","Prism.Tools",380,{"id":119,"name":120,"points":121,"category":122},"46499646","DoNotNotify",347,"productivity",{"id":124,"name":125,"points":126,"category":20},"46488654","how browsers work",302,{"id":128,"name":129,"points":130,"category":131},"46501137","Tailsnitch",280,"security_privacy",5,false,"v2",{"id":136,"name":5,"line":6,"ai":133,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":29,"top":18,"kw":137,"maker":11,"domain":10,"tier":19,"attempts":15,"strip":138,"url":9,"mix":139,"praise":-1,"objection":-1,"sub":21},46518608,[21,24,25,26,27],"30",[31,32,31,31],1794036689712]