[{"data":1,"prerenderedAt":294},["ShallowReactive",2],{"project-49010782":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":43,"post_text":46,"what":47,"said":56,"maker_in_thread":130,"shape":140,"grid":145,"attempts":255,"same_week":256,"maker_shows":44,"maker_hidden":286,"reader_version":287,"card":288},"49010782","Cactus Hybrid","We taught Gemma 4 to know when it's wrong","title","Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong","https:\u002F\u002Fgithub.com\u002Fcactus-compute\u002Fcactus-hybrid","github.com","HenryNdubuaku",1784742989,2026,191,44,311,4366,8,"notable","ai","on-device language model confidence","library",[24,25,26,27],"gemma 4","on-device ai","confidence score","llm calibration","curious",0.152,{"positive":31,"neutral":32,"mixed":33,"critical":34},6,10,0,2,18,{"docs":37,"idea":39,"execution":40,"performance":41,"design":42},{"positive":33,"critical":38},1,{"positive":34,"critical":38},{"positive":34,"critical":33},{"positive":33,"critical":33},{"positive":33,"critical":33},[38,33,33,33,33,33,33,33,33,33,33,33,33,44,45,45,45,33,33,44,33,45,33,45,33,44,33,44,33,44,33,38,33,33,33,44,45,33,45,45],5,3,null,{"problem":46,"audience":46,"pricing":46,"motivation":46,"time_to_build":46,"stack":48,"openness":55,"stage":55,"team_size":55},[49],{"value":50,"source":51,"evidence":52},"Gemma 4","maker",[53],{"id":54},"post","unknown",{"praise":57,"objections":70,"requests":91,"questions":103,"alternatives":122},[58],{"value":59,"source":60,"evidence":61},"Interesting idea and useful on-device approach","commenter",[62,66],{"id":63,"by":64,"text":65},"49019286","dmrivers","Cool idea, some feedback: 1. Consider using conformal prediction to calibrate the cutoff. Conformal prediction provides a distribution-free guarantee under exchangeability. This…",{"id":67,"by":68,"text":69},"49027579","asar","Love what you're doing at cactus, keep it up!!",[71,78,85],{"value":72,"source":60,"evidence":73},"Calling it knowing when wrong overstates what a confidence score can show",[74],{"id":75,"by":76,"text":77},"49015694","BugsJustFindMe","> \"post-trained to know when it's wrong\" Is it also post-trained to know when it's wrong about whether it's wrong? > \"Every response comes with a confidence score between 0 and…",{"value":79,"source":60,"evidence":80},"Probe may be unreliable; rerunning with a new seed seems simpler",[81],{"id":82,"by":83,"text":84},"49019984","maxgashkov","I don't really get why you need handoff if your score is accurate. If it is, and it is low for a given response, just let the harness re-run the prompt with a different seed…",{"value":86,"source":51,"evidence":87},"Rerunning the model for consistency drains battery on edge devices",[88],{"id":89,"by":11,"text":90},"49020163","1. Thanks for this! 2. We combine both on Cactus; rolling entropy method helps with per-token signal for early exit, while the probe runs on complete output. 3. Users hate this…",[92,99],{"value":93,"source":60,"evidence":94},"Benchmark on coding tasks with a larger local model fallback",[95],{"id":96,"by":97,"text":98},"49015274","zdw","Have you benched this for coding tasks, with a fallback to a larger local model, for example Qwen-3.6-27B? Or using it for sub-tasks, where a framework with a larger primary…",{"value":100,"source":60,"evidence":101},"Compare against conformal prediction and consistency-based methods (AUROC)",[102],{"id":63,"by":64,"text":65},[104,115],{"value":105,"source":60,"evidence":106},"Does model quality degrade with the added weights?",[107,111],{"id":108,"by":109,"text":110},"49015777","robrenaud","Does the model quality become degraded in other ways?",{"id":112,"by":113,"text":114},"49019352","zamadatix","It's impossible to add weights which degrade the quality?",{"value":116,"source":60,"evidence":117},"How the probe is trained on hidden states to predict p(wrong)",[118],{"id":119,"by":120,"text":121},"49011476","cacio-e-pepe","> So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various…",[123],{"value":124,"source":60,"evidence":125},"Goodfire's RLFR work",[126],{"id":127,"by":128,"text":129},"49014892","astrobiased","Is this in any way similar to Goodfire's work? https:\u002F\u002Fwww.goodfire.ai\u002Fresearch\u002Frlfr#",{"comments":131,"answered":31,"key":132},9,[133,134,137],{"id":89,"text":90},{"id":135,"text":136},"49018345","YOu are thinking along the right direction, we are going deeper into the signals.",{"id":138,"text":139},"49014942","Thats an interesting outlook, loosely similar.",{"commenters":141,"max_depth":142,"maker_replies":131,"first_comment_mins":143,"lifespan_hours":144},26,11,47,51,[146,148,152,154,156,159,162,165,167,169,171,173,175,178,181,185,189,192,196,198,202,205,207,209,212,214,215,216,219,221,224,226,229,230,232,235,237,240,242,243,244,247,249,252],[75,33,38,147,76,33],"project\u002Fcritical\u002Fdocs",[149,33,33,150,151,33],"49015744","commenter\u002Fneutral\u002Fnone","mistrial9",[153,33,33,150,76,33],"49016135",[155,33,33,150,151,33],"49016519",[157,33,33,150,158,33],"49017634","jurgenburgen",[160,33,33,150,161,33],"49018286","somenameforme",[163,33,33,150,164,33],"49018588","naasking",[166,33,33,150,76,33],"49023879",[168,33,33,150,164,33],"49039145",[170,33,33,150,76,33],"49040702",[172,33,33,150,164,33],"49041309",[174,33,33,150,76,33],"49041763",[176,33,33,150,177,33],"49018526","antonvs",[179,33,33,150,180,33],"49019319","shakna",[182,33,44,183,184,33],"49018009","project\u002Fpositive\u002Fidea","baq",[186,33,45,187,188,33],"49018206","project\u002Fneutral\u002Fidea","NitpickLawyer",[190,33,45,187,191,33],"49019554","cmrdporcupine",[193,33,45,194,195,33],"49021474","project\u002Fneutral\u002Fnone","htrp",[197,33,33,150,76,33],"49022478",[199,33,33,200,201,33],"49019018","other\u002Fneutral\u002Fnone","pjc50",[203,33,33,200,204,33],"49019495","embedding-shape",[63,38,44,206,64,33],"project\u002Fpositive\u002Fexecution",[89,38,33,208,11,38],"maker\u002Fneutral\u002Fnone",[210,34,45,187,211,33],"49015640","mncharity",[135,34,33,213,11,38],"maker\u002Fneutral\u002Fidea",[127,45,45,187,128,33],[138,45,33,208,11,38],[119,217,44,218,120,33],4,"maker\u002Fpositive\u002Fidea",[220,217,33,208,11,38],"49013096",[222,217,44,223,120,33],"49013906","maker\u002Fpositive\u002Fnone",[225,217,33,208,11,38],"49014691",[227,44,44,206,228,33],"49015646","olafura",[96,31,45,194,97,33],[231,31,33,208,11,38],"49018327",[82,233,38,234,83,33],7,"project\u002Fcritical\u002Fidea",[236,233,33,187,11,38],"49020110",[238,18,33,200,239,33],"49018377","jstummbillig",[241,18,33,208,11,38],"49020167",[67,131,44,223,68,33],[108,32,45,194,109,33],[245,32,33,246,11,38],"49018333","project\u002Fneutral\u002Fexecution",[112,32,45,248,113,33],"project\u002Fneutral\u002Fperformance",[250,32,45,248,251,33],"49019534","jstanley",[253,32,45,254,113,33],"49026010","project\u002Fneutral\u002Fdesign",[],[257,262,267,271,276,281],{"id":258,"name":259,"points":260,"category":261},"48968606","I replaced a $120k bowling center system with $1,600 in ESP32s",2935,"hardware",{"id":263,"name":264,"points":265,"category":266},"49008211","Bento",1033,"creative",{"id":268,"name":269,"points":270,"category":266},"48984254","Grace Cathedral Gaussian Splat tour",263,{"id":272,"name":273,"points":274,"category":275},"49008406","HN Hall of Fame",219,"consumer_social",{"id":277,"name":278,"points":279,"category":280},"48946738","Justif",198,"languages",{"id":282,"name":283,"points":284,"category":285},"48947548","honeypotlive",181,"security_privacy",false,"v2",{"id":289,"name":5,"line":6,"ai":286,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":290,"top":18,"kw":291,"maker":11,"domain":10,"tier":19,"attempts":33,"strip":292,"url":9,"mix":293,"praise":59,"objection":72,"sub":21},49010782,0.15,[24,25,26,27],"1000000000000533300503030505050100053033",[31,32,33,34],1794036705737]