[{"data":1,"prerenderedAt":3278},["ShallowReactive",2],{"content-\u002Fposts\u002Ftowards-monosemanticity":3,"blog-posts":3006},{"id":4,"title":5,"body":6,"categories":2992,"date":2993,"description":2994,"draft":2995,"extension":2996,"meta":2997,"navigation":815,"path":2998,"seo":2999,"stem":3000,"tags":3001,"__hash__":3005},"posts\u002Fposts\u002Ftowards-monosemanticity.md","LLM 可解释性（一）：SAE与单语义",{"type":7,"value":8,"toc":2985},"minimark",[9,22,27,30,41,49,545,648,684,782,1047,1050,1061,1068,1078,1081,1087,1091,1550,1999,2006,2524,2527,2800,2804,2811,2890,2897,2908,2912,2915,2922,2925,2950,2957,2961,2964,2971,2974,2981],[10,11,12,13,17,18,21],"p",{},"这就是 ",[14,15,16],"strong",{},"polysemanticity(多义性)","。Anthropic 这篇的工作是:既然神经元本身不好直接解释,那我们就用稀疏自编码器(SAE)把激活\"解混\"成一组",[14,19,20],{},"单语义特征(monosemantic features)",",并给出证据说这些特征才是模型真正在用的计算单元。",[23,24,26],"h2",{"id":25},"一可解释性的前提假设","一、可解释性的前提假设",[10,28,29],{},"机械可解释性(mechanistic interpretability)最早的想法很朴素:打开模型,逐个神经元看它对什么输入有反应,像生物学家找\"祖母细胞\"那样。但这条路撞墙了,因为绝大多数神经元是多义的,一个神经元可能同时对 DNA 序列、阿拉伯文、人名首字母大写都有反应。",[10,31,32,33,40],{},"为什么? Anthropic 在更早的 ",[34,35,39],"a",{"href":36,"rel":37},"https:\u002F\u002Ftransformer-circuits.pub\u002F2022\u002Ftoy_model\u002Findex.html#strategic-ways-out",[38],"nofollow","Toy Models of Superposition"," 里给了一个很干净的假设:",[42,43,44],"blockquote",{},[10,45,46],{},[14,47,48],{},"模型要表达的概念(features)比它拥有的维度(neurons)多得多。",[10,50,51,52,55,56,59,60,106,107,135,136,141,142,173,174,283,284,326,327,330,331,359,360,487,488,544],{},"于是模型把一堆概念\"挤\"进一个低维空间里——这就是 ",[14,53,54],{},"superposition(叠加)","。直觉上,如果每个概念是高维空间里一个",[14,57,58],{},"几乎正交","的方向,那么 ",[61,62,65,87],"span",{"className":63},[64],"katex",[61,66,69],{"className":67},[68],"katex-mathml",[70,71,73],"math",{"xmlns":72},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML",[74,75,76,83],"semantics",{},[77,78,79],"mrow",{},[80,81,82],"mi",{},"d",[84,85,82],"annotation",{"encoding":86},"application\u002Fx-tex",[61,88,92],{"className":89,"ariaHidden":91},[90],"katex-html","true",[61,93,96,101],{"className":94},[95],"base",[61,97],{"className":98,"style":100},[99],"strut","height:0.6944em;",[61,102,82],{"className":103},[104,105],"mord","mathnormal"," 维空间可以塞下远多于 ",[61,108,110,123],{"className":109},[64],[61,111,113],{"className":112},[68],[70,114,115],{"xmlns":72},[74,116,117,121],{},[77,118,119],{},[80,120,82],{},[84,122,82],{"encoding":86},[61,124,126],{"className":125,"ariaHidden":91},[90],[61,127,129,132],{"className":128},[95],[61,130],{"className":131,"style":100},[99],[61,133,82],{"className":134},[104,105]," 个几乎正交的方向(只要允许它们之间有一点干扰),代价是每个方向不再对齐到任何一根坐标轴(也就是不再等于某个神经元)。这件事不是拍脑袋--苏剑林在",[34,137,140],{"href":138,"rel":139},"https:\u002F\u002Fkexue.fm\u002Farchives\u002F7076",[38],"《n 维空间下两个随机向量的夹角分布》","里算过:两个随机单位向量的夹角 ",[61,143,145,160],{"className":144},[64],[61,146,148],{"className":147},[68],[70,149,150],{"xmlns":72},[74,151,152,157],{},[77,153,154],{},[80,155,156],{},"θ",[84,158,159],{"encoding":86},"\\theta",[61,161,163],{"className":162,"ariaHidden":91},[90],[61,164,166,169],{"className":165},[95],[61,167],{"className":168,"style":100},[99],[61,170,156],{"className":171,"style":172},[104,105],"margin-right:0.0278em;"," 的密度正比于 ",[61,175,177,214],{"className":176},[64],[61,178,180],{"className":179},[68],[70,181,182],{"xmlns":72},[74,183,184,211],{},[77,185,186,209],{},[187,188,189,198],"msup",{},[77,190,191,194],{},[80,192,193],{},"sin",[195,196,197],"mo",{},"⁡",[77,199,200,202,205],{},[80,201,82],{},[195,203,204],{},"−",[206,207,208],"mn",{},"2",[80,210,156],{},[84,212,213],{"encoding":86},"\\sin^{d-2}\\theta",[61,215,217],{"className":216,"ariaHidden":91},[90],[61,218,220,224,275,280],{"className":219},[95],[61,221],{"className":222,"style":223},[99],"height:0.9069em;",[61,225,228,231],{"className":226},[227],"mop",[61,229,193],{"className":230},[227],[61,232,235],{"className":233},[234],"msupsub",[61,236,239],{"className":237},[238],"vlist-t",[61,240,243],{"className":241},[242],"vlist-r",[61,244,247],{"className":245,"style":223},[246],"vlist",[61,248,250,255],{"style":249},"top:-3.1208em;margin-right:0.05em;",[61,251],{"className":252,"style":254},[253],"pstrut","height:2.7em;",[61,256,262],{"className":257},[258,259,260,261],"sizing","reset-size6","size3","mtight",[61,263,265,268,272],{"className":264},[104,261],[61,266,82],{"className":267},[104,105,261],[61,269,204],{"className":270},[271,261],"mbin",[61,273,208],{"className":274},[104,261],[61,276],{"className":277,"style":279},[278],"mspace","margin-right:0.1667em;",[61,281,156],{"className":282,"style":172},[104,105],",关于 ",[61,285,287,308],{"className":286},[64],[61,288,290],{"className":289},[68],[70,291,292],{"xmlns":72},[74,293,294,305],{},[77,295,296,299,303],{},[80,297,298],{},"π",[80,300,302],{"mathvariant":301},"normal","\u002F",[206,304,208],{},[84,306,307],{"encoding":86},"\\pi\u002F2",[61,309,311],{"className":310,"ariaHidden":91},[90],[61,312,314,318,322],{"className":313},[95],[61,315],{"className":316,"style":317},[99],"height:1em;vertical-align:-0.25em;",[61,319,298],{"className":320,"style":321},[104,105],"margin-right:0.0359em;",[61,323,325],{"className":324},[104],"\u002F2"," 对称,均值就是 90°。但光有均值不够--二维、三维的均值也是 90°。真正让结论成立的是",[14,328,329],{},"方差",":文章证明方差随维度 ",[61,332,334,347],{"className":333},[64],[61,335,337],{"className":336},[68],[70,338,339],{"xmlns":72},[74,340,341,345],{},[77,342,343],{},[80,344,82],{},[84,346,82],{"encoding":86},[61,348,350],{"className":349,"ariaHidden":91},[90],[61,351,353,356],{"className":352},[95],[61,354],{"className":355,"style":100},[99],[61,357,82],{"className":358},[104,105]," 趋于 0,近似为 ",[61,361,363,389],{"className":362},[64],[61,364,366],{"className":365},[68],[70,367,368],{"xmlns":72},[74,369,370,386],{},[77,371,372],{},[373,374,375,378],"mfrac",{},[206,376,377],{},"1",[77,379,380,382,384],{},[80,381,82],{},[195,383,204],{},[206,385,208],{},[84,387,388],{"encoding":86},"\\frac{1}{d-2}",[61,390,392],{"className":391,"ariaHidden":91},[90],[61,393,395,399],{"className":394},[95],[61,396],{"className":397,"style":398},[99],"height:1.2484em;vertical-align:-0.4033em;",[61,400,402,407,483],{"className":401},[104],[61,403],{"className":404},[405,406],"mopen","nulldelimiter",[61,408,410],{"className":409},[373],[61,411,414,474],{"className":412},[238,413],"vlist-t2",[61,415,417,469],{"className":416},[242],[61,418,421,443,454],{"className":419,"style":420},[246],"height:0.8451em;",[61,422,424,428],{"style":423},"top:-2.655em;",[61,425],{"className":426,"style":427},[253],"height:3em;",[61,429,431],{"className":430},[258,259,260,261],[61,432,434,437,440],{"className":433},[104,261],[61,435,82],{"className":436},[104,105,261],[61,438,204],{"className":439},[271,261],[61,441,208],{"className":442},[104,261],[61,444,446,449],{"style":445},"top:-3.23em;",[61,447],{"className":448,"style":427},[253],[61,450],{"className":451,"style":453},[452],"frac-line","border-bottom-width:0.04em;",[61,455,457,460],{"style":456},"top:-3.394em;",[61,458],{"className":459,"style":427},[253],[61,461,463],{"className":462},[258,259,260,261],[61,464,466],{"className":465},[104,261],[61,467,377],{"className":468},[104,261],[61,470,473],{"className":471},[472],"vlist-s","​",[61,475,477],{"className":476},[242],[61,478,481],{"className":479,"style":480},[246],"height:0.4033em;",[61,482],{},[61,484],{"className":485},[486,406],"mclose","(弧度²),",[61,489,491,511],{"className":490},[64],[61,492,494],{"className":493},[68],[70,495,496],{"xmlns":72},[74,497,498,508],{},[77,499,500,502,505],{},[80,501,82],{},[195,503,504],{},"=",[206,506,507],{},"1000",[84,509,510],{"encoding":86},"d=1000",[61,512,514,534],{"className":513,"ariaHidden":91},[90],[61,515,517,520,523,527,531],{"className":516},[95],[61,518],{"className":519,"style":100},[99],[61,521,82],{"className":522},[104,105],[61,524],{"className":525,"style":526},[278],"margin-right:0.2778em;",[61,528,504],{"className":529},[530],"mrel",[61,532],{"className":533,"style":526},[278],[61,535,537,541],{"className":536},[95],[61,538],{"className":539,"style":540},[99],"height:0.6444em;",[61,542,507],{"className":543},[104]," 时标准差只有约 1.8°,夹角几乎被钉死在 90°。",[10,546,547,548,579,580,647],{},"另外有 JL 引理：塞下 ",[61,549,551,565],{"className":550},[64],[61,552,554],{"className":553},[68],[70,555,556],{"xmlns":72},[74,557,558,563],{},[77,559,560],{},[80,561,562],{},"N",[84,564,562],{"encoding":86},[61,566,568],{"className":567,"ariaHidden":91},[90],[61,569,571,575],{"className":570},[95],[61,572],{"className":573,"style":574},[99],"height:0.6833em;",[61,576,562],{"className":577,"style":578},[104,105],"margin-right:0.109em;"," 个向量只需要 ",[61,581,583,616],{"className":582},[64],[61,584,586],{"className":585},[68],[70,587,588],{"xmlns":72},[74,589,590,613],{},[77,591,592,595,599,602,605,608,610],{},[80,593,594],{},"O",[195,596,598],{"stretchy":597},"false","(",[80,600,601],{},"l",[80,603,604],{},"o",[80,606,607],{},"g",[80,609,562],{},[195,611,612],{"stretchy":597},")",[84,614,615],{"encoding":86},"O(logN)",[61,617,619],{"className":618,"ariaHidden":91},[90],[61,620,622,625,628,631,635,638,641,644],{"className":621},[95],[61,623],{"className":624,"style":317},[99],[61,626,594],{"className":627,"style":172},[104,105],[61,629,598],{"className":630},[405],[61,632,601],{"className":633,"style":634},[104,105],"margin-right:0.0197em;",[61,636,604],{"className":637},[104,105],[61,639,607],{"className":640,"style":321},[104,105],[61,642,562],{"className":643,"style":578},[104,105],[61,645,612],{"className":646},[486]," 维空间。",[42,649,650],{},[10,651,652],{},[14,653,654,655,683],{},"高维空间里随便抓两个方向就几乎正交,所以能塞下远多于 ",[61,656,658,671],{"className":657},[64],[61,659,661],{"className":660},[68],[70,662,663],{"xmlns":72},[74,664,665,669],{},[77,666,667],{},[80,668,82],{},[84,670,82],{"encoding":86},[61,672,674],{"className":673,"ariaHidden":91},[90],[61,675,677,680],{"className":676},[95],[61,678],{"className":679,"style":100},[99],[61,681,82],{"className":682},[104,105]," 个几乎正交的方向。",[10,685,686,687,715,716,781],{},"随手模拟一下:在不同维度 ",[61,688,690,703],{"className":689},[64],[61,691,693],{"className":692},[68],[70,694,695],{"xmlns":72},[74,696,697,701],{},[77,698,699],{},[80,700,82],{},[84,702,82],{"encoding":86},[61,704,706],{"className":705,"ariaHidden":91},[90],[61,707,709,712],{"className":708},[95],[61,710],{"className":711,"style":100},[99],[61,713,82],{"className":714},[104,105]," 里各取一万个随机向量,看随机一对的夹角落在 ",[61,717,719,751],{"className":718},[64],[61,720,722],{"className":721},[68],[70,723,724],{"xmlns":72},[74,725,726,748],{},[77,727,728,731,734,737,740,743,745],{},[195,729,730],{"stretchy":597},"[",[206,732,733],{},"80",[80,735,736],{"mathvariant":301},"°",[195,738,739],{"separator":91},",",[206,741,742],{},"100",[80,744,736],{"mathvariant":301},[195,746,747],{"stretchy":597},"]",[84,749,750],{"encoding":86},"[80°,100°]",[61,752,754],{"className":753,"ariaHidden":91},[90],[61,755,757,760,763,767,771,774,778],{"className":756},[95],[61,758],{"className":759,"style":317},[99],[61,761,730],{"className":762},[405],[61,764,766],{"className":765},[104],"80°",[61,768,739],{"className":769},[770],"mpunct",[61,772],{"className":773,"style":279},[278],[61,775,777],{"className":776},[104],"100°",[61,779,747],{"className":780},[486]," 的比例:",[783,784,789],"pre",{"className":785,"code":786,"language":787,"meta":788,"style":788},"language-python shiki shiki-themes github-light github-dark","import numpy as np\n\nnp.random.seed(0)\nfor d in [2, 3, 10, 100, 1000]:\n    n = 10000\n    x = np.random.randn(n, d)\n    x \u002F= np.linalg.norm(x, axis=1, keepdims=True)\n    i, j = np.random.randint(n, size=100000), np.random.randint(n, size=100000)\n    ang = np.degrees(np.arccos(np.clip((x[i] * x[j]).sum(1), -1, 1)))\n    print(f\"d={d:5d}: 夹角落在 [80°, 100°] 的比例 = {(np.abs(ang - 90) \u003C= 10).mean():.4f}\")\n","python","",[790,791,792,810,817,830,869,880,891,922,952,986],"code",{"__ignoreMap":788},[61,793,796,800,804,807],{"class":794,"line":795},"line",1,[61,797,799],{"class":798},"szBVR","import",[61,801,803],{"class":802},"sVt8B"," numpy ",[61,805,806],{"class":798},"as",[61,808,809],{"class":802}," np\n",[61,811,813],{"class":794,"line":812},2,[61,814,816],{"emptyLinePlaceholder":815},true,"\n",[61,818,820,823,827],{"class":794,"line":819},3,[61,821,822],{"class":802},"np.random.seed(",[61,824,826],{"class":825},"sj4cs","0",[61,828,829],{"class":802},")\n",[61,831,833,836,839,842,845,847,850,853,855,858,860,862,864,866],{"class":794,"line":832},4,[61,834,835],{"class":798},"for",[61,837,838],{"class":802}," d ",[61,840,841],{"class":798},"in",[61,843,844],{"class":802}," [",[61,846,208],{"class":825},[61,848,849],{"class":802},", ",[61,851,852],{"class":825},"3",[61,854,849],{"class":802},[61,856,857],{"class":825},"10",[61,859,849],{"class":802},[61,861,742],{"class":825},[61,863,849],{"class":802},[61,865,507],{"class":825},[61,867,868],{"class":802},"]:\n",[61,870,872,875,877],{"class":794,"line":871},5,[61,873,874],{"class":802},"    n ",[61,876,504],{"class":798},[61,878,879],{"class":825}," 10000\n",[61,881,883,886,888],{"class":794,"line":882},6,[61,884,885],{"class":802},"    x ",[61,887,504],{"class":798},[61,889,890],{"class":802}," np.random.randn(n, d)\n",[61,892,894,896,899,902,906,908,910,912,915,917,920],{"class":794,"line":893},7,[61,895,885],{"class":802},[61,897,898],{"class":798},"\u002F=",[61,900,901],{"class":802}," np.linalg.norm(x, ",[61,903,905],{"class":904},"s4XuR","axis",[61,907,504],{"class":798},[61,909,377],{"class":825},[61,911,849],{"class":802},[61,913,914],{"class":904},"keepdims",[61,916,504],{"class":798},[61,918,919],{"class":825},"True",[61,921,829],{"class":802},[61,923,925,928,930,933,936,938,941,944,946,948,950],{"class":794,"line":924},8,[61,926,927],{"class":802},"    i, j ",[61,929,504],{"class":798},[61,931,932],{"class":802}," np.random.randint(n, ",[61,934,935],{"class":904},"size",[61,937,504],{"class":798},[61,939,940],{"class":825},"100000",[61,942,943],{"class":802},"), np.random.randint(n, ",[61,945,935],{"class":904},[61,947,504],{"class":798},[61,949,940],{"class":825},[61,951,829],{"class":802},[61,953,955,958,960,963,966,969,971,974,977,979,981,983],{"class":794,"line":954},9,[61,956,957],{"class":802},"    ang ",[61,959,504],{"class":798},[61,961,962],{"class":802}," np.degrees(np.arccos(np.clip((x[i] ",[61,964,965],{"class":798},"*",[61,967,968],{"class":802}," x[j]).sum(",[61,970,377],{"class":825},[61,972,973],{"class":802},"), ",[61,975,976],{"class":798},"-",[61,978,377],{"class":825},[61,980,849],{"class":802},[61,982,377],{"class":825},[61,984,985],{"class":802},")))\n",[61,987,989,992,994,997,1001,1004,1006,1009,1012,1015,1017,1020,1022,1025,1028,1031,1034,1037,1040,1042,1045],{"class":794,"line":988},10,[61,990,991],{"class":825},"    print",[61,993,598],{"class":802},[61,995,996],{"class":798},"f",[61,998,1000],{"class":999},"sZZnC","\"d=",[61,1002,1003],{"class":825},"{",[61,1005,82],{"class":802},[61,1007,1008],{"class":798},":5d",[61,1010,1011],{"class":825},"}",[61,1013,1014],{"class":999},": 夹角落在 [80°, 100°] 的比例 = ",[61,1016,1003],{"class":825},[61,1018,1019],{"class":802},"(np.abs(ang ",[61,1021,976],{"class":798},[61,1023,1024],{"class":825}," 90",[61,1026,1027],{"class":802},") ",[61,1029,1030],{"class":798},"\u003C=",[61,1032,1033],{"class":825}," 10",[61,1035,1036],{"class":802},").mean()",[61,1038,1039],{"class":798},":.4f",[61,1041,1011],{"class":825},[61,1043,1044],{"class":999},"\"",[61,1046,829],{"class":802},[10,1048,1049],{},"输出大致是 d=2: 0.11、d=3: 0.17、d=10: 0.39、d=100: 0.92、d=1000: 1.00--维度一上来,几乎每一对都贴着 90°。",[10,1051,1052,1053,1056,1057,1060],{},"方向挤在一起就不再严格正交,激活一个特征会顺带蹭到别的特征一点--这是叠加付出的",[14,1054,1055],{},"干扰","。几乎正交的条件不够,真正让 superposition 成立的是",[14,1058,1059],{},"稀疏",":同时激活的特征很少,每对方向本来就接近 90°。",[10,1062,1063,1064,1067],{},"所以 polysemanticity 是模型在有限容量下",[14,1065,1066],{},"主动选择","的压缩策略:",[1069,1070,1071,1075],"ul",{},[1072,1073,1074],"li",{},"神经元 = 坐标轴方向 → 因为叠加,一个轴上叠了好几个概念 → 多义。",[1072,1076,1077],{},"特征 = 叠加后那个几乎正交的方向 → 才是\"一个概念\"的载体 → 单义。",[10,1079,1080],{},"另外，LLM 其实是对超大规模数据的压缩，superposition 可以说是压缩过程中自然的行为；对应的，我们追求可解释性就需要进行解压缩。",[10,1082,1083,1084],{},"结论很清楚:",[14,1085,1086],{},"要看懂模型,得先把叠加解开,把坐标轴换成特征方向。",[23,1088,1090],{"id":1089},"二稀疏自编码器把混合的激活解混","二、稀疏自编码器:把混合的激活\"解混\"",[10,1092,1093,1094,1157,1158,1263,1264,1317,1318,1346,1347,1456,1457,1460,1461,1464,1465,1549],{},"上面说模型把概念稀疏叠加进激活,写成式子就是 ",[61,1095,1097,1120],{"className":1096},[64],[61,1098,1100],{"className":1099},[68],[70,1101,1102],{"xmlns":72},[74,1103,1104,1117],{},[77,1105,1106,1109,1112,1115],{},[80,1107,1108],{},"x",[195,1110,1111],{},"≈",[80,1113,1114],{},"W",[80,1116,996],{},[84,1118,1119],{"encoding":86},"x \\approx Wf",[61,1121,1123,1142],{"className":1122,"ariaHidden":91},[90],[61,1124,1126,1130,1133,1136,1139],{"className":1125},[95],[61,1127],{"className":1128,"style":1129},[99],"height:0.4831em;",[61,1131,1108],{"className":1132},[104,105],[61,1134],{"className":1135,"style":526},[278],[61,1137,1111],{"className":1138},[530],[61,1140],{"className":1141,"style":526},[278],[61,1143,1145,1149,1153],{"className":1144},[95],[61,1146],{"className":1147,"style":1148},[99],"height:0.8889em;vertical-align:-0.1944em;",[61,1150,1114],{"className":1151,"style":1152},[104,105],"margin-right:0.1389em;",[61,1154,996],{"className":1155,"style":1156},[104,105],"margin-right:0.1076em;",":字典 ",[61,1159,1161,1194],{"className":1160},[64],[61,1162,1164],{"className":1163},[68],[70,1165,1166],{"xmlns":72},[74,1167,1168,1191],{},[77,1169,1170,1172,1175],{},[80,1171,1114],{},[195,1173,1174],{},"∈",[187,1176,1177,1181],{},[80,1178,1180],{"mathvariant":1179},"double-struck","R",[77,1182,1183,1185,1188],{},[80,1184,82],{},[195,1186,1187],{},"×",[80,1189,1190],{},"m",[84,1192,1193],{"encoding":86},"W \\in \\mathbb{R}^{d\\times m}",[61,1195,1197,1216],{"className":1196,"ariaHidden":91},[90],[61,1198,1200,1204,1207,1210,1213],{"className":1199},[95],[61,1201],{"className":1202,"style":1203},[99],"height:0.7224em;vertical-align:-0.0391em;",[61,1205,1114],{"className":1206,"style":1152},[104,105],[61,1208],{"className":1209,"style":526},[278],[61,1211,1174],{"className":1212},[530],[61,1214],{"className":1215,"style":526},[278],[61,1217,1219,1223],{"className":1218},[95],[61,1220],{"className":1221,"style":1222},[99],"height:0.8491em;",[61,1224,1226,1230],{"className":1225},[104],[61,1227,1180],{"className":1228},[104,1229],"mathbb",[61,1231,1233],{"className":1232},[234],[61,1234,1236],{"className":1235},[238],[61,1237,1239],{"className":1238},[242],[61,1240,1242],{"className":1241,"style":1222},[246],[61,1243,1245,1248],{"style":1244},"top:-3.063em;margin-right:0.05em;",[61,1246],{"className":1247,"style":254},[253],[61,1249,1251],{"className":1250},[258,259,260,261],[61,1252,1254,1257,1260],{"className":1253},[104,261],[61,1255,82],{"className":1256},[104,105,261],[61,1258,1187],{"className":1259},[271,261],[61,1261,1190],{"className":1262},[104,105,261]," 过完备(",[61,1265,1267,1286],{"className":1266},[64],[61,1268,1270],{"className":1269},[68],[70,1271,1272],{"xmlns":72},[74,1273,1274,1283],{},[77,1275,1276,1278,1281],{},[80,1277,1190],{},[195,1279,1280],{},"≫",[80,1282,82],{},[84,1284,1285],{"encoding":86},"m \\gg d",[61,1287,1289,1308],{"className":1288,"ariaHidden":91},[90],[61,1290,1292,1296,1299,1302,1305],{"className":1291},[95],[61,1293],{"className":1294,"style":1295},[99],"height:0.5782em;vertical-align:-0.0391em;",[61,1297,1190],{"className":1298},[104,105],[61,1300],{"className":1301,"style":526},[278],[61,1303,1280],{"className":1304},[530],[61,1306],{"className":1307,"style":526},[278],[61,1309,1311,1314],{"className":1310},[95],[61,1312],{"className":1313,"style":100},[99],[61,1315,82],{"className":1316},[104,105],"),编码 ",[61,1319,1321,1334],{"className":1320},[64],[61,1322,1324],{"className":1323},[68],[70,1325,1326],{"xmlns":72},[74,1327,1328,1332],{},[77,1329,1330],{},[80,1331,996],{},[84,1333,996],{"encoding":86},[61,1335,1337],{"className":1336,"ariaHidden":91},[90],[61,1338,1340,1343],{"className":1339},[95],[61,1341],{"className":1342,"style":1148},[99],[61,1344,996],{"className":1345,"style":1156},[104,105]," 稀疏(",[61,1348,1350,1379],{"className":1349},[64],[61,1351,1353],{"className":1352},[68],[70,1354,1355],{"xmlns":72},[74,1356,1357,1376],{},[77,1358,1359,1362,1364,1371,1374],{},[80,1360,1361],{"mathvariant":301},"∥",[80,1363,996],{},[1365,1366,1367,1369],"msub",{},[80,1368,1361],{"mathvariant":301},[206,1370,826],{},[195,1372,1373],{},"≪",[80,1375,1190],{},[84,1377,1378],{"encoding":86},"\\|f\\|_0 \\ll m",[61,1380,1382,1446],{"className":1381,"ariaHidden":91},[90],[61,1383,1385,1388,1391,1394,1437,1440,1443],{"className":1384},[95],[61,1386],{"className":1387,"style":317},[99],[61,1389,1361],{"className":1390},[104],[61,1392,996],{"className":1393,"style":1156},[104,105],[61,1395,1397,1400],{"className":1396},[104],[61,1398,1361],{"className":1399},[104],[61,1401,1403],{"className":1402},[234],[61,1404,1406,1428],{"className":1405},[238,413],[61,1407,1409,1425],{"className":1408},[242],[61,1410,1413],{"className":1411,"style":1412},[246],"height:0.3011em;",[61,1414,1416,1419],{"style":1415},"top:-2.55em;margin-left:0em;margin-right:0.05em;",[61,1417],{"className":1418,"style":254},[253],[61,1420,1422],{"className":1421},[258,259,260,261],[61,1423,826],{"className":1424},[104,261],[61,1426,473],{"className":1427},[472],[61,1429,1431],{"className":1430},[242],[61,1432,1435],{"className":1433,"style":1434},[246],"height:0.15em;",[61,1436],{},[61,1438],{"className":1439,"style":526},[278],[61,1441,1373],{"className":1442},[530],[61,1444],{"className":1445,"style":526},[278],[61,1447,1449,1453],{"className":1448},[95],[61,1450],{"className":1451,"style":1452},[99],"height:0.4306em;",[61,1454,1190],{"className":1455},[104,105],")。这正是经典",[14,1458,1459],{},"字典学习","(dictionary learning)的形式,而解混它就是把这个方程反过来解--顺手的工具是 ",[14,1462,1463],{},"sparse autoencoder(SAE)","。对模型某一层(这篇是小型 SoLU transformer 的 MLP 层)的激活向量 ",[61,1466,1468,1490],{"className":1467},[64],[61,1469,1471],{"className":1470},[68],[70,1472,1473],{"xmlns":72},[74,1474,1475,1487],{},[77,1476,1477,1479,1481],{},[80,1478,1108],{},[195,1480,1174],{},[187,1482,1483,1485],{},[80,1484,1180],{"mathvariant":1179},[80,1486,82],{},[84,1488,1489],{"encoding":86},"x \\in \\mathbb{R}^{d}",[61,1491,1493,1511],{"className":1492,"ariaHidden":91},[90],[61,1494,1496,1499,1502,1505,1508],{"className":1495},[95],[61,1497],{"className":1498,"style":1295},[99],[61,1500,1108],{"className":1501},[104,105],[61,1503],{"className":1504,"style":526},[278],[61,1506,1174],{"className":1507},[530],[61,1509],{"className":1510,"style":526},[278],[61,1512,1514,1517],{"className":1513},[95],[61,1515],{"className":1516,"style":1222},[99],[61,1518,1520,1523],{"className":1519},[104],[61,1521,1180],{"className":1522},[104,1229],[61,1524,1526],{"className":1525},[234],[61,1527,1529],{"className":1528},[238],[61,1530,1532],{"className":1531},[242],[61,1533,1535],{"className":1534,"style":1222},[246],[61,1536,1537,1540],{"style":1244},[61,1538],{"className":1539,"style":254},[253],[61,1541,1543],{"className":1542},[258,259,260,261],[61,1544,1546],{"className":1545},[104,261],[61,1547,82],{"className":1548},[104,105,261],", 训练一个 autoencoder:",[61,1551,1554],{"className":1552},[1553],"katex-display",[61,1555,1557,1654],{"className":1556},[64],[61,1558,1560],{"className":1559},[68],[70,1561,1563],{"xmlns":72,"display":1562},"block",[74,1564,1565,1651],{},[77,1566,1567,1571,1573,1575,1588,1590,1597,1600,1602,1605,1612,1614,1616,1619,1622,1630,1632,1639,1641,1643,1645],{},[1568,1569,1570],"mtext",{},"encode: ",[80,1572,996],{},[195,1574,504],{},[77,1576,1577,1579,1582,1585],{},[80,1578,1180],{"mathvariant":301},[80,1580,1581],{"mathvariant":301},"e",[80,1583,1584],{"mathvariant":301},"L",[80,1586,1587],{"mathvariant":301},"U",[195,1589,598],{"stretchy":597},[1365,1591,1592,1594],{},[80,1593,1114],{},[1568,1595,1596],{},"enc",[1568,1598,1599],{}," ",[80,1601,1108],{},[195,1603,1604],{},"+",[1365,1606,1607,1610],{},[80,1608,1609],{},"b",[1568,1611,1596],{},[195,1613,612],{"stretchy":597},[195,1615,739],{"separator":91},[278,1617],{"width":1618},"2em",[1568,1620,1621],{},"decode: ",[1623,1624,1625,1627],"mover",{"accent":91},[80,1626,1108],{},[195,1628,1629],{},"^",[195,1631,504],{},[1365,1633,1634,1636],{},[80,1635,1114],{},[1568,1637,1638],{},"dec",[1568,1640,1599],{},[80,1642,996],{},[195,1644,1604],{},[1365,1646,1647,1649],{},[80,1648,1609],{},[1568,1650,1638],{},[84,1652,1653],{"encoding":86},"\\text{encode: } f = \\mathrm{ReLU}(W_{\\text{enc}}\\, x + b_{\\text{enc}}), \\qquad \\text{decode: } \\hat{x} = W_{\\text{dec}}\\, f + b_{\\text{dec}}",[61,1655,1657,1682,1763,1878,1946],{"className":1656,"ariaHidden":91},[90],[61,1658,1660,1663,1670,1673,1676,1679],{"className":1659},[95],[61,1661],{"className":1662,"style":1148},[99],[61,1664,1667],{"className":1665},[104,1666],"text",[61,1668,1570],{"className":1669},[104],[61,1671,996],{"className":1672,"style":1156},[104,105],[61,1674],{"className":1675,"style":526},[278],[61,1677,504],{"className":1678},[530],[61,1680],{"className":1681,"style":526},[278],[61,1683,1685,1688,1696,1699,1747,1750,1753,1757,1760],{"className":1684},[95],[61,1686],{"className":1687,"style":317},[99],[61,1689,1691],{"className":1690},[104],[61,1692,1695],{"className":1693},[104,1694],"mathrm","ReLU",[61,1697,598],{"className":1698},[405],[61,1700,1702,1705],{"className":1701},[104],[61,1703,1114],{"className":1704,"style":1152},[104,105],[61,1706,1708],{"className":1707},[234],[61,1709,1711,1739],{"className":1710},[238,413],[61,1712,1714,1736],{"className":1713},[242],[61,1715,1718],{"className":1716,"style":1717},[246],"height:0.1514em;",[61,1719,1721,1724],{"style":1720},"top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;",[61,1722],{"className":1723,"style":254},[253],[61,1725,1727],{"className":1726},[258,259,260,261],[61,1728,1730],{"className":1729},[104,261],[61,1731,1733],{"className":1732},[104,1666,261],[61,1734,1596],{"className":1735},[104,261],[61,1737,473],{"className":1738},[472],[61,1740,1742],{"className":1741},[242],[61,1743,1745],{"className":1744,"style":1434},[246],[61,1746],{},[61,1748],{"className":1749,"style":279},[278],[61,1751,1108],{"className":1752},[104,105],[61,1754],{"className":1755,"style":1756},[278],"margin-right:0.2222em;",[61,1758,1604],{"className":1759},[271],[61,1761],{"className":1762,"style":1756},[278],[61,1764,1766,1769,1815,1818,1821,1825,1828,1834,1869,1872,1875],{"className":1765},[95],[61,1767],{"className":1768,"style":317},[99],[61,1770,1772,1775],{"className":1771},[104],[61,1773,1609],{"className":1774},[104,105],[61,1776,1778],{"className":1777},[234],[61,1779,1781,1807],{"className":1780},[238,413],[61,1782,1784,1804],{"className":1783},[242],[61,1785,1787],{"className":1786,"style":1717},[246],[61,1788,1789,1792],{"style":1415},[61,1790],{"className":1791,"style":254},[253],[61,1793,1795],{"className":1794},[258,259,260,261],[61,1796,1798],{"className":1797},[104,261],[61,1799,1801],{"className":1800},[104,1666,261],[61,1802,1596],{"className":1803},[104,261],[61,1805,473],{"className":1806},[472],[61,1808,1810],{"className":1809},[242],[61,1811,1813],{"className":1812,"style":1434},[246],[61,1814],{},[61,1816,612],{"className":1817},[486],[61,1819,739],{"className":1820},[770],[61,1822],{"className":1823,"style":1824},[278],"margin-right:2em;",[61,1826],{"className":1827,"style":279},[278],[61,1829,1831],{"className":1830},[104,1666],[61,1832,1621],{"className":1833},[104],[61,1835,1838],{"className":1836},[104,1837],"accent",[61,1839,1841],{"className":1840},[238],[61,1842,1844],{"className":1843},[242],[61,1845,1847,1856],{"className":1846,"style":100},[246],[61,1848,1850,1853],{"style":1849},"top:-3em;",[61,1851],{"className":1852,"style":427},[253],[61,1854,1108],{"className":1855},[104,105],[61,1857,1858,1861],{"style":1849},[61,1859],{"className":1860,"style":427},[253],[61,1862,1866],{"className":1863,"style":1865},[1864],"accent-body","left:-0.2222em;",[61,1867,1629],{"className":1868},[104],[61,1870],{"className":1871,"style":526},[278],[61,1873,504],{"className":1874},[530],[61,1876],{"className":1877,"style":526},[278],[61,1879,1881,1884,1931,1934,1937,1940,1943],{"className":1880},[95],[61,1882],{"className":1883,"style":1148},[99],[61,1885,1887,1890],{"className":1886},[104],[61,1888,1114],{"className":1889,"style":1152},[104,105],[61,1891,1893],{"className":1892},[234],[61,1894,1896,1923],{"className":1895},[238,413],[61,1897,1899,1920],{"className":1898},[242],[61,1900,1903],{"className":1901,"style":1902},[246],"height:0.3361em;",[61,1904,1905,1908],{"style":1720},[61,1906],{"className":1907,"style":254},[253],[61,1909,1911],{"className":1910},[258,259,260,261],[61,1912,1914],{"className":1913},[104,261],[61,1915,1917],{"className":1916},[104,1666,261],[61,1918,1638],{"className":1919},[104,261],[61,1921,473],{"className":1922},[472],[61,1924,1926],{"className":1925},[242],[61,1927,1929],{"className":1928,"style":1434},[246],[61,1930],{},[61,1932],{"className":1933,"style":279},[278],[61,1935,996],{"className":1936,"style":1156},[104,105],[61,1938],{"className":1939,"style":1756},[278],[61,1941,1604],{"className":1942},[271],[61,1944],{"className":1945,"style":1756},[278],[61,1947,1949,1953],{"className":1948},[95],[61,1950],{"className":1951,"style":1952},[99],"height:0.8444em;vertical-align:-0.15em;",[61,1954,1956,1959],{"className":1955},[104],[61,1957,1609],{"className":1958},[104,105],[61,1960,1962],{"className":1961},[234],[61,1963,1965,1991],{"className":1964},[238,413],[61,1966,1968,1988],{"className":1967},[242],[61,1969,1971],{"className":1970,"style":1902},[246],[61,1972,1973,1976],{"style":1415},[61,1974],{"className":1975,"style":254},[253],[61,1977,1979],{"className":1978},[258,259,260,261],[61,1980,1982],{"className":1981},[104,261],[61,1983,1985],{"className":1984},[104,1666,261],[61,1986,1638],{"className":1987},[104,261],[61,1989,473],{"className":1990},[472],[61,1992,1994],{"className":1993},[242],[61,1995,1997],{"className":1996,"style":1434},[246],[61,1998],{},[10,2000,2001,2002,2005],{},"目标函数是",[14,2003,2004],{},"重建误差 + 稀疏惩罚",":",[61,2007,2009],{"className":2008},[1553],[61,2010,2012,2088],{"className":2011},[64],[61,2013,2015],{"className":2014},[68],[70,2016,2017],{"xmlns":72,"display":1562},[74,2018,2019,2085],{},[77,2020,2021,2024,2026,2060,2062,2065],{},[80,2022,1584],{"mathvariant":2023},"script",[195,2025,504],{},[2027,2028,2029,2057],"munder",{},[2027,2030,2031,2054],{},[77,2032,2033,2035,2037,2039,2045],{},[80,2034,1361],{"mathvariant":301},[80,2036,1108],{},[195,2038,204],{},[1623,2040,2041,2043],{"accent":91},[80,2042,1108],{},[195,2044,1629],{},[2046,2047,2048,2050,2052],"msubsup",{},[80,2049,1361],{"mathvariant":301},[206,2051,208],{},[206,2053,208],{},[195,2055,2056],{"stretchy":91},"⏟",[1568,2058,2059],{},"重建",[195,2061,1604],{},[80,2063,2064],{},"λ",[2027,2066,2067,2083],{},[2027,2068,2069,2081],{},[77,2070,2071,2073,2075],{},[80,2072,1361],{"mathvariant":301},[80,2074,996],{},[1365,2076,2077,2079],{},[80,2078,1361],{"mathvariant":301},[206,2080,377],{},[195,2082,2056],{"stretchy":91},[1568,2084,1059],{},[84,2086,2087],{"encoding":86},"\\mathcal{L} = \\underbrace{\\|x - \\hat{x}\\|_2^2}_{\\text{重建}} + \\lambda \\underbrace{\\|f\\|_1}_{\\text{稀疏}}",[61,2089,2091,2110,2358],{"className":2090,"ariaHidden":91},[90],[61,2092,2094,2097,2101,2104,2107],{"className":2093},[95],[61,2095],{"className":2096,"style":574},[99],[61,2098,1584],{"className":2099},[104,2100],"mathcal",[61,2102],{"className":2103,"style":526},[278],[61,2105,504],{"className":2106},[530],[61,2108],{"className":2109,"style":526},[278],[61,2111,2113,2117,2349,2352,2355],{"className":2112},[95],[61,2114],{"className":2115,"style":2116},[99],"height:2.4404em;vertical-align:-1.5763em;",[61,2118,2121],{"className":2119},[2120,2027],"minner",[61,2122,2124,2340],{"className":2123},[238,413],[61,2125,2127,2337],{"className":2126},[242],[61,2128,2131,2150],{"className":2129,"style":2130},[246],"height:0.8641em;",[61,2132,2134,2137],{"style":2133},"top:-1.4237em;",[61,2135],{"className":2136,"style":427},[253],[61,2138,2140],{"className":2139},[258,259,260,261],[61,2141,2143],{"className":2142},[104,261],[61,2144,2146],{"className":2145},[104,1666,261],[61,2147,2059],{"className":2148},[104,2149,261],"cjk_fallback",[61,2151,2152,2155],{"style":1849},[61,2153],{"className":2154,"style":427},[253],[61,2156,2158],{"className":2157},[2120,2027],[61,2159,2161,2328],{"className":2160},[238,413],[61,2162,2164,2325],{"className":2163},[242],[61,2165,2167,2217],{"className":2166,"style":2130},[246],[61,2168,2172,2175],{"className":2169,"style":2171},[2170],"svg-align","top:-2.102em;",[61,2173],{"className":2174,"style":427},[253],[61,2176,2180,2197,2207],{"className":2177,"style":2179},[2178],"stretchy","height:0.548em;min-width:1.6em;",[61,2181,2185],{"className":2182,"style":2184},[2183],"brace-left","height:0.548em;",[2186,2187,2193],"svg",{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2192},"http:\u002F\u002Fwww.w3.org\u002F2000\u002Fsvg","400em","0.548em","0 0 400000 548","xMinYMin slice",[2194,2195],"path",{"d":2196},"M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13\n 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688\n 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7\n-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z",[61,2198,2201],{"className":2199,"style":2184},[2200],"brace-center",[2186,2202,2204],{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2203},"xMidYMin slice",[2194,2205],{"d":2206},"M199572 214\nc100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14\n 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3\n 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0\n-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z",[61,2208,2211],{"className":2209,"style":2184},[2210],"brace-right",[2186,2212,2214],{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2213},"xMaxYMin slice",[2194,2215],{"d":2216},"M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3\n 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237\n-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z",[61,2218,2219,2222],{"style":1849},[61,2220],{"className":2221,"style":427},[253],[61,2223,2225,2228,2231,2234,2237,2240,2271],{"className":2224},[104],[61,2226,1361],{"className":2227},[104],[61,2229,1108],{"className":2230},[104,105],[61,2232],{"className":2233,"style":1756},[278],[61,2235,204],{"className":2236},[271],[61,2238],{"className":2239,"style":1756},[278],[61,2241,2243],{"className":2242},[104,1837],[61,2244,2246],{"className":2245},[238],[61,2247,2249],{"className":2248},[242],[61,2250,2252,2260],{"className":2251,"style":100},[246],[61,2253,2254,2257],{"style":1849},[61,2255],{"className":2256,"style":427},[253],[61,2258,1108],{"className":2259},[104,105],[61,2261,2262,2265],{"style":1849},[61,2263],{"className":2264,"style":427},[253],[61,2266,2268],{"className":2267,"style":1865},[1864],[61,2269,1629],{"className":2270},[104],[61,2272,2274,2277],{"className":2273},[104],[61,2275,1361],{"className":2276},[104],[61,2278,2280],{"className":2279},[234],[61,2281,2283,2316],{"className":2282},[238,413],[61,2284,2286,2313],{"className":2285},[242],[61,2287,2289,2301],{"className":2288,"style":2130},[246],[61,2290,2292,2295],{"style":2291},"top:-2.453em;margin-left:0em;margin-right:0.05em;",[61,2293],{"className":2294,"style":254},[253],[61,2296,2298],{"className":2297},[258,259,260,261],[61,2299,208],{"className":2300},[104,261],[61,2302,2304,2307],{"style":2303},"top:-3.113em;margin-right:0.05em;",[61,2305],{"className":2306,"style":254},[253],[61,2308,2310],{"className":2309},[258,259,260,261],[61,2311,208],{"className":2312},[104,261],[61,2314,473],{"className":2315},[472],[61,2317,2319],{"className":2318},[242],[61,2320,2323],{"className":2321,"style":2322},[246],"height:0.247em;",[61,2324],{},[61,2326,473],{"className":2327},[472],[61,2329,2331],{"className":2330},[242],[61,2332,2335],{"className":2333,"style":2334},[246],"height:0.898em;",[61,2336],{},[61,2338,473],{"className":2339},[472],[61,2341,2343],{"className":2342},[242],[61,2344,2347],{"className":2345,"style":2346},[246],"height:1.5763em;",[61,2348],{},[61,2350],{"className":2351,"style":1756},[278],[61,2353,1604],{"className":2354},[271],[61,2356],{"className":2357,"style":1756},[278],[61,2359,2361,2365,2368,2371],{"className":2360},[95],[61,2362],{"className":2363,"style":2364},[99],"height:2.3263em;vertical-align:-1.5763em;",[61,2366,2064],{"className":2367},[104,105],[61,2369],{"className":2370,"style":279},[278],[61,2372,2374],{"className":2373},[2120,2027],[61,2375,2377,2516],{"className":2376},[238,413],[61,2378,2380,2513],{"className":2379},[242],[61,2381,2384,2401],{"className":2382,"style":2383},[246],"height:0.75em;",[61,2385,2386,2389],{"style":2133},[61,2387],{"className":2388,"style":427},[253],[61,2390,2392],{"className":2391},[258,259,260,261],[61,2393,2395],{"className":2394},[104,261],[61,2396,2398],{"className":2397},[104,1666,261],[61,2399,1059],{"className":2400},[104,2149,261],[61,2402,2403,2406],{"style":1849},[61,2404],{"className":2405,"style":427},[253],[61,2407,2409],{"className":2408},[2120,2027],[61,2410,2412,2505],{"className":2411},[238,413],[61,2413,2415,2502],{"className":2414},[242],[61,2416,2418,2448],{"className":2417,"style":2383},[246],[61,2419,2421,2424],{"className":2420,"style":2171},[2170],[61,2422],{"className":2423,"style":427},[253],[61,2425,2427,2434,2441],{"className":2426,"style":2179},[2178],[61,2428,2430],{"className":2429,"style":2184},[2183],[2186,2431,2432],{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2192},[2194,2433],{"d":2196},[61,2435,2437],{"className":2436,"style":2184},[2200],[2186,2438,2439],{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2203},[2194,2440],{"d":2206},[61,2442,2444],{"className":2443,"style":2184},[2210],[2186,2445,2446],{"xmlns":2188,"width":2189,"height":2190,"viewBox":2191,"preserveAspectRatio":2213},[2194,2447],{"d":2216},[61,2449,2450,2453],{"style":1849},[61,2451],{"className":2452,"style":427},[253],[61,2454,2456,2459,2462],{"className":2455},[104],[61,2457,1361],{"className":2458},[104],[61,2460,996],{"className":2461,"style":1156},[104,105],[61,2463,2465,2468],{"className":2464},[104],[61,2466,1361],{"className":2467},[104],[61,2469,2471],{"className":2470},[234],[61,2472,2474,2494],{"className":2473},[238,413],[61,2475,2477,2491],{"className":2476},[242],[61,2478,2480],{"className":2479,"style":1412},[246],[61,2481,2482,2485],{"style":1415},[61,2483],{"className":2484,"style":254},[253],[61,2486,2488],{"className":2487},[258,259,260,261],[61,2489,377],{"className":2490},[104,261],[61,2492,473],{"className":2493},[472],[61,2495,2497],{"className":2496},[242],[61,2498,2500],{"className":2499,"style":1434},[246],[61,2501],{},[61,2503,473],{"className":2504},[472],[61,2506,2508],{"className":2507},[242],[61,2509,2511],{"className":2510,"style":2334},[246],[61,2512],{},[61,2514,473],{"className":2515},[472],[61,2517,2519],{"className":2518},[242],[61,2520,2522],{"className":2521,"style":2346},[246],[61,2523],{},[10,2525,2526],{},"几个关键点:",[2528,2529,2530,2618,2761],"ol",{},[1072,2531,2532,2535,2536,2588,2589,2617],{},[14,2533,2534],{},"字典比原维度大很多。"," MLP 宽度 ",[61,2537,2539,2558],{"className":2538},[64],[61,2540,2542],{"className":2541},[68],[70,2543,2544],{"xmlns":72},[74,2545,2546,2555],{},[77,2547,2548,2550,2552],{},[80,2549,82],{},[195,2551,1111],{},[206,2553,2554],{},"2048",[84,2556,2557],{"encoding":86},"d \\approx 2048",[61,2559,2561,2579],{"className":2560,"ariaHidden":91},[90],[61,2562,2564,2567,2570,2573,2576],{"className":2563},[95],[61,2565],{"className":2566,"style":100},[99],[61,2568,82],{"className":2569},[104,105],[61,2571],{"className":2572,"style":526},[278],[61,2574,1111],{"className":2575},[530],[61,2577],{"className":2578,"style":526},[278],[61,2580,2582,2585],{"className":2581},[95],[61,2583],{"className":2584,"style":540},[99],[61,2586,2554],{"className":2587},[104],",但字典可以开到几十万甚至上百万个特征(扩张系数 expansion factor = 字典大小 \u002F ",[61,2590,2592,2605],{"className":2591},[64],[61,2593,2595],{"className":2594},[68],[70,2596,2597],{"xmlns":72},[74,2598,2599,2603],{},[77,2600,2601],{},[80,2602,82],{},[84,2604,82],{"encoding":86},[61,2606,2608],{"className":2607,"ariaHidden":91},[90],[61,2609,2611,2614],{"className":2610},[95],[61,2612],{"className":2613,"style":100},[99],[61,2615,82],{"className":2616},[104,105],",这篇扫到量级几百倍)。这正对应\"概念数 > 神经元数\"的叠加假设——故意给一个过完备(overcomplete)的字典。",[1072,2619,2620,2623,2624,2652,2653,2656,2657,2686,2687,2760],{},[14,2621,2622],{},"ReLU 强制非负 + L1 强制稀疏。"," 重建要求字典能还原 ",[61,2625,2627,2640],{"className":2626},[64],[61,2628,2630],{"className":2629},[68],[70,2631,2632],{"xmlns":72},[74,2633,2634,2638],{},[77,2635,2636],{},[80,2637,1108],{},[84,2639,1108],{"encoding":86},[61,2641,2643],{"className":2642,"ariaHidden":91},[90],[61,2644,2646,2649],{"className":2645},[95],[61,2647],{"className":2648,"style":1452},[99],[61,2650,1108],{"className":2651},[104,105],",L1 要求每个样本只激活很少几个特征。两个约束一起,逼着字典学出",[14,2654,2655],{},"稀疏组合","——每个激活向量被表示成\"少数几个特征方向\"的叠加。(",[61,2658,2660,2674],{"className":2659},[64],[61,2661,2663],{"className":2662},[68],[70,2664,2665],{"xmlns":72},[74,2666,2667,2671],{},[77,2668,2669],{},[80,2670,2064],{},[84,2672,2673],{"encoding":86},"\\lambda",[61,2675,2677],{"className":2676,"ariaHidden":91},[90],[61,2678,2680,2683],{"className":2679},[95],[61,2681],{"className":2682,"style":100},[99],[61,2684,2064],{"className":2685},[104,105]," 量级在 ",[61,2688,2690,2712],{"className":2689},[64],[61,2691,2693],{"className":2692},[68],[70,2694,2695],{"xmlns":72},[74,2696,2697,2709],{},[77,2698,2699],{},[187,2700,2701,2703],{},[206,2702,857],{},[77,2704,2705,2707],{},[195,2706,204],{},[206,2708,852],{},[84,2710,2711],{"encoding":86},"10^{-3}",[61,2713,2715],{"className":2714,"ariaHidden":91},[90],[61,2716,2718,2722,2725],{"className":2717},[95],[61,2719],{"className":2720,"style":2721},[99],"height:0.8141em;",[61,2723,377],{"className":2724},[104],[61,2726,2728,2731],{"className":2727},[104],[61,2729,826],{"className":2730},[104],[61,2732,2734],{"className":2733},[234],[61,2735,2737],{"className":2736},[238],[61,2738,2740],{"className":2739},[242],[61,2741,2743],{"className":2742,"style":2721},[246],[61,2744,2745,2748],{"style":1244},[61,2746],{"className":2747,"style":254},[253],[61,2749,2751],{"className":2750},[258,259,260,261],[61,2752,2754,2757],{"className":2753},[104,261],[61,2755,204],{"className":2756},[104,261],[61,2758,852],{"className":2759},[104,261],",具体值看模型规模。)",[1072,2762,2763,2795,2796,2799],{},[14,2764,2765,2766,2794],{},"L0 远小于 ",[61,2767,2769,2782],{"className":2768},[64],[61,2770,2772],{"className":2771},[68],[70,2773,2774],{"xmlns":72},[74,2775,2776,2780],{},[77,2777,2778],{},[80,2779,82],{},[84,2781,82],{"encoding":86},[61,2783,2785],{"className":2784,"ariaHidden":91},[90],[61,2786,2788,2791],{"className":2787},[95],[61,2789],{"className":2790,"style":100},[99],[61,2792,82],{"className":2793},[104,105],",但方差解释率(FVE)很高。"," 这是最漂亮的结果:每个 token 的 MLP 激活,只需要",[14,2797,2798],{},"几十到几百个特征","稀疏激活就能重建绝大部分方差。如果神经元真是计算单元,你不可能用这么少的方向重建得这么好——这反过来支持\"特征才是真实单元\"。",[23,2801,2803],{"id":2802},"三sae-与语义的映射","三、SAE 与语义的映射",[10,2805,2806,2807,2810],{},"光重建好不够,得证明每个特征",[14,2808,2809],{},"确实对应一个可解释的概念","。他们做了大量 case study,挑几个最能说明问题的:",[2812,2813,2814,2830],"table",{},[2815,2816,2817],"thead",{},[2818,2819,2820,2824,2827],"tr",{},[2821,2822,2823],"th",{},"特征",[2821,2825,2826],{},"激活条件",[2821,2828,2829],{},"为什么重要",[2831,2832,2833,2851,2864,2877],"tbody",{},[2818,2834,2835,2841,2848],{},[2836,2837,2838],"td",{},[14,2839,2840],{},"DNA 特征",[2836,2842,2843,2844,2847],{},"输入是 ",[790,2845,2846],{},"ATGC..."," 这样的 DNA 序列",[2836,2849,2850],{},"跨 token、跨上下文都稳定,甚至能拿来给文本\"是否含 DNA\"打标",[2818,2852,2853,2858,2861],{},[2836,2854,2855],{},[14,2856,2857],{},"阿拉伯文特征",[2836,2859,2860],{},"文本里出现阿拉伯字符",[2836,2862,2863],{},"与主题无关,纯粹认\"书写系统\"这个抽象",[2818,2865,2866,2871,2874],{},[2836,2867,2868],{},[14,2869,2870],{},"专有名词 \u002F 缩写特征",[2836,2872,2873],{},"首字母大写的人名、全大写缩写",[2836,2875,2876],{},"语法-词性层面的抽象",[2818,2878,2879,2884,2887],{},[2836,2880,2881],{},[14,2882,2883],{},"无用 token 特征",[2836,2885,2886],{},"模型词表里那些训练时几乎从不出现的\"垃圾\" token",[2836,2888,2889],{},"关键对照:这种 token 上特征的激活模式异常,说明特征不是无脑触发",[10,2891,2892,2893,2896],{},"最有说服力的是 ",[14,2894,2895],{},"无用 token(unuseful tokens)"," 实验:Anthropic 在词表里故意塞了一批训练数据里几乎不出现的 token。如果特征只是字典学习的数学产物,它对这些 token 应该也\"随便\"激活;但实际观察到的激活模式是有结构的、能和\"这个 token 从没被真正学过\"对上。这是对\"特征是真实单元、不是字典 artifact\"的一个反证辩护。",[10,2898,2899,2900,2903,2904,2907],{},"合起来,论文给出的画像:每一个 SAE 特征,在一个",[14,2901,2902],{},"很窄但很一致","的语义条件下激活——这就是 ",[14,2905,2906],{},"monosemantic(单语义)","。和原始神经元的 polysemantic 形成鲜明对比。",[23,2909,2911],{"id":2910},"四sae-真的有用","四、SAE 真的有用",[10,2913,2914],{},"这一节是我觉得最该细读的部分,也是最容易被科普文章跳过的。",[10,2916,2917,2918,2921],{},"最自然的反对意见是:",[14,2919,2920],{},"你开一个过完备字典 + L1,当然能把任何信号拆成\"稀疏的零件\",但这些零件是模型在用的东西,还是你强加的基?","(basis dependence \u002F 字典依赖)",[10,2923,2924],{},"他们的回应分两层:",[2528,2926,2927,2936],{},[1072,2928,2929,2932,2933],{},[14,2930,2931],{},"特征分裂(feature splitting)是良性的。"," 字典开得越大,一个粗特征会\"分裂\"成几个更细的特征(比如\"阿拉伯文\" → \"阿拉伯文-宗教文本\"\"阿拉伯文-代码注释\"……)。这看起来像过拟合,但分裂后每个子特征依然单义、依然可解释——只是粒度更细。",[14,2934,2935],{},"所以字典大小控制的是\"放大镜的倍率\",不是凭空造概念。",[1072,2937,2938,2941,2942,2945,2946],{},[14,2939,2940],{},"行为层面的因果验证。"," 光说\"可解释\"是旁观证据,真正站住脚的是:能不能拿特征去",[14,2943,2944],{},"干预","模型行为。把某个特征在推理时强行激活或抑制(clamp \u002F ablation),输出会按预期改变。",[2947,2948,2949],"em",{},"(注:大规模 steering 实验——包括那个有名的\"毒性特征\"案例——主要在 2024 年的续作 Scaling Monosemanticity 里,本篇规模较小,但因果验证的思路已立住。)",[10,2951,2952,2953,2956],{},"我的看法:这两点合起来基本回应了 basis dependence,但没完全消解。",[14,2954,2955],{},"特征是\"字典相关的\"这件事是事实",",只是\"字典相关\"不等于\"任意\"——好的字典会让特征和模型的真实计算对齐。这有点像量子力学里的表象问题:换基会变,但物理不变;SAE 找的是那个让特征尽量\"自然\"的基。",[23,2958,2960],{"id":2959},"五延伸思考","五、延伸思考",[10,2962,2963],{},"总的来说，SAEs 为 Transformer 的可解释性提供了可能性。我们朴素直觉(神经元=概念)被一个更准确结构(在神经元中被叠加某一方向=概念)取代。可解释性的基本单元从 neuron 让位给了 feature direction。",[10,2965,2966,2967,2970],{},"不过，SAE 只是\"放大镜\"。它给的是一个",[14,2968,2969],{},"可调粒度","的视图(字典大小=倍率)。模型的特征不是一个有确定数目的东西,而是依观察尺度展开的——这跟物理里\"尺度决定你看到什么\"是一回事。别把某一次 SAE 的特征数当成模型\"真正的概念数\"。SAE 只是把多语义的拆分成更细的特征，便于我们解释，以及在不修改参数的情况下进行操作。",[2972,2973],"hr",{},[42,2975,2976],{},[10,2977,2978],{},[14,2979,2980],{},"模型把比神经元多得多的概念稀疏地压进激活空间,所以神经元是多义的;稀疏自编码器通过\"过完备字典 + L1 稀疏\"把这个叠加解混,得到的特征在很大概率上才是模型真正在用的、单义的计算单元——而\"是不是真在用\",最终只能靠因果干预来回答。",[2982,2983,2984],"style",{},"html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":788,"searchDepth":812,"depth":819,"links":2986},[2987,2988,2989,2990,2991],{"id":25,"depth":812,"text":26},{"id":1089,"depth":812,"text":1090},{"id":2802,"depth":812,"text":2803},{"id":2910,"depth":812,"text":2911},{"id":2959,"depth":812,"text":2960},"技术","2026-07-12 22:00:00","读 Anthropic《Towards Monosemanticity》。从难以解释的神经元出发, 讲清superposition为什么让单个神经元混成一团, SAE怎么把多语义的激活分解成单语义特征。",false,"md",{},"\u002Fposts\u002Ftowards-monosemanticity",{"title":5,"description":2994},"posts\u002Ftowards-monosemanticity",[3002,3003,3004],"LLMs","可解释性","SAE","boAExu6KsW5C43yZ6qJz_kq4RtNzhmV-F-P782yKI6A",[3007,3016,3022,3029,3034,3043,3052,3058,3062,3071,3080,3086,3092,3100,3104,3108,3118,3125,3131,3135,3139,3141,3148,3156,3161,3167,3173,3178,3183,3190,3197,3204,3210,3217,3225,3231,3237,3243,3249,3254,3263,3271],{"path":3008,"title":3009,"description":3010,"date":3011,"categories":3012,"tags":3013,"draft":815},"\u002Fposts\u002Ftransposition-borrowed-chords","平移借用：大调里立刻能用的和弦","把大调整体平移，借用各调的顺阶和弦；有符号平移量 ±2、±3、±5、±7 给出七个近关系调，合计 25 个现成三和弦与 31 个现成七和弦，全部 12 个平移量给 36 与 48。","2026-11-08 16:20:00","音乐",[3014,3015],"乐理","和弦",{"path":3017,"title":3018,"description":3019,"date":3020,"categories":3012,"tags":3021,"draft":815},"\u002Fposts\u002Fdiatonic-chords-major-system","大调体系下的三和弦和七和弦","以同主音 C 为基准，逐个推出七套音阶的顺阶三和弦与七和弦，标注每个和弦的出处，去重后得到 21 个三和弦与 28 个七和弦。","2026-11-08 15:20:00",[3014,3015],{"path":3023,"title":3024,"description":788,"date":3025,"categories":3026,"tags":3027,"draft":815},"\u002Fposts\u002F作息感慨","作息感慨","2026-11-04 07:15:08","生活杂谈",[3028],"日常",{"path":3030,"title":3031,"description":788,"date":3032,"categories":788,"tags":3033,"draft":815},"\u002Fposts\u002Fgp-tab-测试","gp-tab-测试","2026-10-12 10:51:22",null,{"path":3035,"title":3036,"description":3037,"date":3038,"categories":2992,"tags":3039,"draft":2995},"\u002Fposts\u002Fnetwork-clean","怎么让网络变得干净","一个干净的网络，让人省心的使用服务。","2026-10-11 19:40:02",[3040,3041,3042],"VPN","网络","AI",{"path":3044,"title":3045,"description":3046,"date":3047,"categories":3026,"tags":3048,"draft":2995},"\u002Fposts\u002Fthe-age-of-hesitation","[搬运] 不敢投入的时代","为什么我们现在越来越不敢投入了。很多时候我们以为的选择，比如你有两个工作offer，现在选择好像是我到底应该考公还是考编。明明这两个选项都不是真正的选项，因为你还没有开始，也不知道最后能不能成。这个时候我们就已经开始犹豫了。","2026-10-10 11:44:12",[3049,3050,3051],"翻译","搬运","心理学",{"path":3053,"title":3054,"description":3055,"date":3056,"categories":3026,"tags":3057,"draft":815},"\u002Fposts\u002Fwhy-happiness-look-like-same","[搬运] 为什么大家的幸福都长一个样？","为什么大家的幸福都长一个样？一种认知被广泛传播后似乎变得不断扁平，这种感觉很奇怪。","2026-10-09 13:57:15",[3049,3050,3051],{"path":3059,"title":3060,"description":788,"date":3061,"categories":3026,"tags":3033,"draft":815},"\u002Fposts\u002Feverything-is-paid-twice","任何事情都要被消费两次","2026-10-06 12:30:06",{"path":3063,"title":3064,"description":3065,"date":3066,"categories":3067,"tags":3068,"draft":2995},"\u002Fposts\u002Fllm","为什么模型上线后，体感像是“降智”了？","权重与 KV cache 的量化都能降低 serving 成本，也可能改变输出；但体感变弱不足以证明服务后来切换了精度。","2026-09-29 21:44:08","技术杂谈",[3002,3069,3070],"Serving","量化",{"path":3072,"title":3073,"description":3074,"date":3075,"categories":2992,"tags":3076,"draft":2995},"\u002Fposts\u002Fspinning-up-policy-optimization","[译] 深度强化学习入门（三）：策略优化入门","OpenAI Spinning Up 教程第三部分的全文翻译：推导最简单的策略梯度、log-derivative trick、EGLP 引理、reward-to-go 与基线，配 PyTorch 示例代码。","2026-09-28 17:49:25",[3077,3078,3079,3049],"强化学习","RL","Spinning Up",{"path":3081,"title":3082,"description":3083,"date":3084,"categories":2992,"tags":3085,"draft":2995},"\u002Fposts\u002Fspinning-up-rl-intro-2","[译] 深度强化学习入门（二）：强化学习算法的分类","OpenAI Spinning Up 第二部分的全文翻译：无模型与基于模型的 RL、策略优化与 Q-learning 的取舍，以及规划、专家迭代和经验增强等方法。","2026-09-28 10:20:21",[3077,3078,3079,3049],{"path":3087,"title":3088,"description":3089,"date":3090,"categories":2992,"tags":3091,"draft":2995},"\u002Fposts\u002Fspinning-up-rl-intro","[译] 深度强化学习入门（一）：RL 的关键概念","OpenAI Spinning Up 教程第一部分的全文翻译：智能体与环境、状态与观测、动作空间、策略、轨迹、奖励与回报、价值函数、贝尔曼方程、优势函数，附 MDP 形式化。","2026-09-28 09:29:40",[3077,3078,3079,3049],{"path":3093,"title":3094,"description":3095,"date":3096,"categories":3026,"tags":3097,"draft":815},"\u002Fposts\u002Fwhen-machines-come-alive","冰箱之后，生活还会怎样改变？","从厨房里的果蝇想到冰箱、移动互联网与 AI 的多轮调用：AGI 已经到来，只是昂贵的智能会怎样改变生活？","2026-09-25 06:36:22",[3042,3098,3099],"随笔","科技",{"path":3101,"title":3102,"description":788,"date":3103,"categories":788,"tags":3033,"draft":815},"\u002Fposts\u002Fmathematical-foundations-of-llms-1","LLM的数学基础（1）：","2026-09-19 22:59:55",{"path":3105,"title":3106,"description":788,"date":3107,"categories":788,"tags":3033,"draft":815},"\u002Fposts\u002Ftowards-maximize-serving-capatity","如何在 offline bench 里最大化 LLM serving system 的吞吐","2026-09-19 05:45:05",{"path":3109,"title":3110,"description":3111,"date":3112,"categories":3067,"tags":3113,"draft":2995},"\u002Fposts\u002Fllm-memory-compute-tradeoff","LLM 的 Memory-Compute Trade-off：从 MHA 到 MLA、Sparse 与 Linear Attention","ds4.1f 上线了。它的架构 novel 在哪，便宜和快又是从哪省出来的。","2026-09-11 12:00:00",[3002,3114,3115,3116,3117],"Attention","MLA","DeepSeek","长上下文",{"path":3119,"title":3120,"description":3121,"date":3122,"categories":3067,"tags":3123,"draft":815},"\u002Fposts\u002Fmarkdown-extensions","Markdown 扩展语法速查","本站 Markdown 支持的扩展语法演示与速查：admonition 提示块、GitHub 仓库卡片、spoiler 遮罩、ABC 乐谱、GP 乐谱、音频播放器、浏览器可运行代码块。","2026-09-01 15:30:00",[3124],"博客",{"path":3126,"title":3127,"description":3128,"date":3129,"categories":3026,"tags":3130,"draft":815},"\u002Fposts","为什么还需要手写博客？","AI这么强，还要古法写博客吗？","2026-09-01 13:17:53",[3124],{"path":3132,"title":3133,"description":788,"date":3134,"categories":788,"tags":3033,"draft":815},"\u002Fposts\u002Fllm-serving-1","LLM Serving（一）:以在H200上部署GLM-5.2为例","2026-07-21 06:15:44",{"path":3136,"title":3137,"description":788,"date":3138,"categories":3012,"tags":3033,"draft":2995},"\u002Fposts\u002Finteresting-music-rhythm","有趣的音乐节奏","2026-07-16 12:47:24",{"path":2998,"title":5,"description":2994,"date":2993,"categories":2992,"tags":3140,"draft":2995},[3002,3003,3004],{"path":3142,"title":3143,"description":3144,"date":3145,"categories":3026,"tags":3146,"draft":815},"\u002Fposts\u002Fpiketty-ai-china-k-shaped","用皮凯蒂的镜头看 AI 与中国 K 型社会","一次凌晨的对话整理。从 r>g 出发，看 AI 时代的资本分配、中国的世袭化趋势、体制内的二元劳动力市场，以及一个普通 00 后在一线打拼的真实牌面。","2026-06-24 03:00:00",[3098,3147,3042],"经济",{"path":3149,"title":3150,"description":788,"date":3151,"categories":788,"tags":3152,"draft":815},"\u002Fposts\u002Fclaude-code-compact-behavior","cluade code compact 行为","2026-06-23 20:17:42",[3153,3154,3155],"Agent","AI Agent","Claude Code",{"path":3157,"title":3158,"description":788,"date":3159,"categories":2992,"tags":3160,"draft":815},"\u002Fposts\u002Fclaude-code-settings","Claude Code 配置分享","2026-06-23 14:22:12",[3002,3154,3153,3155],{"path":3162,"title":3163,"description":3164,"date":3165,"categories":3067,"tags":3166,"draft":2995},"\u002Fposts\u002Fmy-thoughts-on-the-ai-industry","简单聊聊对 AI Agent 行业的理解","写这篇文章的初衷，是我发觉逐渐忘了前两个月的 AI\u002FLLMs 是什么样的能力和情况了。意识中似乎某些在突然加速，但理性和直觉告诉我这不对，作为行内人，我知道这是一个渐进的、累积的过程。我想自己回忆一下，Agent的发展。","2026-06-20 13:14:25",[3002,3154,3153],{"path":3168,"title":3169,"description":788,"date":3170,"categories":2992,"tags":3171,"draft":2995},"\u002Fposts\u002Frotational-positional-encoding","RoPE","2026-06-17 18:31:40",[3002,3172,3169],"Transformer",{"path":3174,"title":3175,"description":788,"date":3176,"categories":2992,"tags":3177,"draft":2995},"\u002Fposts\u002Fintroduction-to-large-language-model","大语言模型入门基础","2026-06-16 22:36:11",[3172,3002],{"path":3179,"title":3180,"description":788,"date":3181,"categories":3067,"tags":3182,"draft":815},"\u002Fposts\u002Fcoding-agent","关于 Coding Agent 使用的遐思","2026-06-03 17:54:20",[3153],{"path":3184,"title":3185,"description":3186,"date":3187,"categories":2992,"tags":3188,"draft":815},"\u002Fposts\u002Fabout-ssh-usage","关于SSH的使用——客户端篇","介绍SSH客户端的用法。","2025-06-17 02:44:57",[3189],"工具",{"path":3191,"title":3192,"description":3193,"date":3194,"categories":3067,"tags":3195,"draft":2995},"\u002Fposts\u002Fopensourcesoftwarestarttutorial","如何开始一个开源项目？","浅谈我对开源项目的认知。","2025-04-17 18:00:00",[3196],"开源",{"path":3198,"title":3199,"description":3200,"date":3201,"categories":3202,"tags":3203,"draft":815},"\u002Fposts\u002Fmotivation","渴求动力","有意识的生活并主动地行动会使得生活更美好。","2025-03-05 00:44:57","生活-情感",[3098],{"path":3205,"title":3206,"description":3207,"date":3208,"categories":3202,"tags":3209,"draft":815},"\u002Fposts\u002F2024summary","2024年终总结","啊，痛彻的时光；啊，悲怆的人！","2025-01-01 01:00:00",[3098],{"path":3211,"title":3212,"description":3213,"date":3214,"categories":2992,"tags":3215,"draft":815},"\u002Fposts\u002Fvps-crash-startup","快速搭建个人VPS","如何快速搭建个人VPS。","2024-12-31 19:00:00",[3216],"VPS",{"path":3218,"title":3219,"description":3220,"date":3221,"categories":2992,"tags":3222,"draft":815},"\u002Fposts\u002Fdata-label-software","AI数据（自动）标注软件概述","对市面上目前的AI数据（自动）标注软件进行调研。","2024-12-20 14:00:00",[3223,3224],"数据标注","深度学习",{"path":3226,"title":3227,"description":3228,"date":3229,"categories":3202,"tags":3230,"draft":2995},"\u002Fposts\u002Fself-destruction","生活的崩坏与自我毁灭","是什么在牵引着我？我的行为，我的思想？我的感受？我所处的地方？它们在何处？","2024-12-19 12:00:00",[3098],{"path":3232,"title":3233,"description":3234,"date":3235,"categories":2992,"tags":3236,"draft":815},"\u002Fposts\u002Fcs194-196_llmagents","LLM Agents 课程笔记","UCB CS294\u002F194-196 Large Language Model Agents 的课程笔记。","2024-11-30 20:00:00",[3153],{"path":3238,"title":3239,"description":3240,"date":3241,"categories":3202,"tags":3242,"draft":815},"\u002Fposts\u002Fwhatsyourworth","价值、位置","你们的生存究竟有何价值？如果毫无价值，你们究竟为何存在？","2024-11-29 02:00:00",[3098],{"path":3244,"title":3245,"description":3246,"date":3247,"categories":3202,"tags":3248,"draft":2995},"\u002Fposts\u002Fblogrelifever2","重启博客v2","忙里偷闲学习前端，还是决定结合想法手搓一个自己的博客。","2024-11-25 00:00:00",[3124],{"path":3250,"title":3251,"description":3252,"date":3253,"categories":3202,"tags":3033,"draft":815},"\u002Fposts\u002Fblogrelife","重启博客","在我的表达欲与现实所需(问题解决记录)驱动下，我决定重写博客。","2024-10-21 20:44:57",{"path":3255,"title":3256,"description":3257,"date":3258,"categories":3026,"tags":3259,"draft":2995},"\u002Fposts\u002Fsurvival-welcome-to-university","生存手册——欢迎来到大学","大学同时生产三样东西：可验证技能、认证信号和驯化资本。你付了四年的钱，真正拿到手的是哪一样，取决于你毕业时要进的那个市场在为什么付费。","2024-10-12 20:00:00",[3260,3261,3262],"大学","求职","方法论",{"path":3264,"title":3265,"description":3266,"date":3267,"categories":3202,"tags":3268,"draft":2995},"\u002Fposts\u002Frevival","240205 复活","好吧，断更了许久，我又终于回来了，我又终于回来了。","2024-02-05 20:00:00",[3098,3269,3270],"生活","返乡",{"path":3272,"title":3273,"description":3274,"date":3275,"categories":3202,"tags":3276,"draft":815},"\u002Fposts\u002Fclouds","「230717」云","我已经很久忘记了不我已经很少把自己的想法在lofter上发布的。","2023-07-17 20:00:00",[3098,3277],"散文",1795115925453]