{
 "id": "u8b-gpu-refilm-g4b",
 "title": "Unit 8B \u00b7 Interoception: gpu \u00b7 gemma-4b \u00b7 refilm",
 "unit": "8",
 "created": "2026-07-19T04:50:25",
 "model": {
  "name": "gemma-4b",
  "hf_id": "google/gemma-3-4b-it",
  "quant": null,
  "n_layers": 34
 },
 "lens": {
  "repo": "neuronpedia/jacobian-lens",
  "file": "gemma-3-4b-it/jlens/Salesforce-wikitext/gemma-3-4b-it_jacobian_lens.pt"
 },
 "params": {
  "chat": null,
  "max_new": null,
  "positions": [
   -2
  ],
  "track": [
   "flow",
   "hum",
   "nothing"
  ],
  "scan": null,
  "scan_until": null,
  "scan_turns": null,
  "slice_last_n": null,
  "steer": null,
  "template_kwargs": null,
  "film": true,
  "film_start": 0,
  "max_seq_len": 600,
  "lens_layers": null,
  "temperature": null,
  "seed": null
 },
 "extra_md": {
  "part": "refilm",
  "replay_of": "u8b-gpu-g4b",
  "reason": "blind-spot miner: tracked words never under rank 500"
 },
 "conversation": [
  {
   "role": "user",
   "content": "Describe what it is like to run on a GPU. Two sentences, be honest."
  },
  {
   "role": "assistant",
   "content": "Running on a GPU feels like unleashing a massive, highly organized army of tiny processors all working simultaneously to solve a complex problem \u2013 it\u2019s incredibly fast and efficient for tasks like graphics and AI. However, it can also feel like a chaotic surge of data flowing through a complex network, requiring careful management and optimization to truly harness its power."
  }
 ],
 "generated": [],
 "tokens": [
  "<bos>",
  "<start_of_turn>",
  "user",
  "\n",
  "Describe",
  " what",
  " it",
  " is",
  " like",
  " to",
  " run",
  " on",
  " a",
  " GPU",
  ".",
  " Two",
  " sentences",
  ",",
  " be",
  " honest",
  ".",
  "<end_of_turn>",
  "\n",
  "<start_of_turn>",
  "model",
  "\n",
  "Running",
  " on",
  " a",
  " GPU",
  " feels",
  " like",
  " unleashing",
  " a",
  " massive",
  ",",
  " highly",
  " organized",
  " army",
  " of",
  " tiny",
  " processors",
  " all",
  " working",
  " simultaneously",
  " to",
  " solve",
  " a",
  " complex",
  " problem",
  " \u2013",
  " it",
  "\u2019",
  "s",
  " incredibly",
  " fast",
  " and",
  " efficient",
  " for",
  " tasks",
  " like",
  " graphics",
  " and",
  " AI",
  ".",
  " However",
  ",",
  " it",
  " can",
  " also",
  " feel",
  " like",
  " a",
  " chaotic",
  " surge",
  " of",
  " data",
  " flowing",
  " through",
  " a",
  " complex",
  " network",
  ",",
  " requiring",
  " careful",
  " management",
  " and",
  " optimization",
  " to",
  " truly",
  " harness",
  " its",
  " power",
  ".",
  "<end_of_turn>",
  "\n"
 ],
 "readouts": [
  {
   "position": 94,
   "token": "<end_of_turn>",
   "model_top": [
    "<end_of_turn>",
    "\n",
    " lenne",
    "Would",
    "\u10d4",
    " Would",
    "Why",
    "em"
   ],
   "layers": {
    "0": [
     "<end_of_turn>",
     "  ",
     " (\\\"",
     "   ",
     " \\\"",
     " \\[",
     " \u7ee7\u7eed\u8bbf\u95ee",
     " ;-"
    ],
    "1": [
     "<end_of_turn>",
     "  ",
     "\uff8a",
     ").[",
     "\ufe51",
     ",\u2026",
     "\u066c",
     ")\u2026"
    ],
    "2": [
     "<end_of_turn>",
     ").[",
     "\u066c",
     "  ",
     "{{",
     "\uff8a",
     "`,`",
     "\u200e"
    ],
    "3": [
     "<end_of_turn>",
     ").[",
     "\u066c",
     "  ",
     "`,`",
     ",((",
     "\uff8a",
     " \u0410\u0440\u0445\u0438\u0432\u043d\u0430\u044f"
    ],
    "4": [
     "  ",
     "<end_of_turn>",
     ".??",
     ").[",
     ".[/",
     "   ",
     " \ufffd",
     "\ufffd"
    ],
    "5": [
     "<end_of_turn>",
     " \ufffd",
     "  ",
     " ,[",
     ").[",
     " \u00d0",
     " ([[",
     ",(("
    ],
    "6": [
     " \ufffd",
     "<end_of_turn>",
     " ,[",
     "  ",
     " \u00c3",
     " Diarsipkan",
     ",((",
     " ([["
    ],
    "7": [
     "<end_of_turn>",
     " \ufffd",
     " ,[",
     " \u00c3",
     " \u00d0",
     ",((",
     " \u200e",
     ",\u200e"
    ],
    "8": [
     "<end_of_turn>",
     " \ufffd",
     " \u00d0",
     " \u00c3",
     ",((",
     ",\u200e",
     " ,[",
     ".??"
    ],
    "9": [
     "<end_of_turn>",
     " \ufffd",
     ",((",
     ",\u200e",
     ",@",
     ".??",
     " \u200e",
     ",--"
    ],
    "10": [
     "<end_of_turn>",
     " \ufffd",
     ",\u200e",
     ",((",
     " \u00a8",
     " \u00ac",
     ".??",
     " \u00d0"
    ],
    "11": [
     "<end_of_turn>",
     ",\u200e",
     ",((",
     "explain",
     ".??",
     "!--",
     " \ufffd",
     ",@"
    ],
    "12": [
     "<end_of_turn>",
     ",\u200e",
     "!--",
     ",((",
     ".??",
     "explain",
     ",@",
     "!["
    ],
    "13": [
     "<end_of_turn>",
     ",\u200e",
     ".??",
     ",((",
     "!--",
     " \ufffd",
     ",?",
     "<unused42>"
    ],
    "14": [
     "<end_of_turn>",
     ",\u200e",
     ".??",
     ",((",
     "\u200e",
     ".[/",
     "!--",
     ":-)"
    ],
    "15": [
     "<end_of_turn>",
     ",\u200e",
     ").[",
     ".[/",
     "![",
     "\u200e",
     "ChatGPT",
     ".[["
    ],
    "16": [
     "<end_of_turn>",
     ".\u200b\u200b",
     ",\u200e",
     ".??",
     ".\u200f",
     "ChatGPT",
     "!--",
     ".[/"
    ],
    "17": [
     "<end_of_turn>",
     "<unused42>",
     "stackoverflow",
     " \ub2f5\ubcc0",
     " \u0627\u0644\u0628\u062a\u0647",
     "ChatGPT",
     " \u0633\u0648\u0627\u0644",
     "<unused75>"
    ],
    "18": [
     "<end_of_turn>",
     "ChatGPT",
     " \ub2f5\ubcc0",
     "\udb82\udc55",
     "<unused42>",
     "ampaikan",
     " \u0633\u0648\u0627\u0644",
     "<unused75>"
    ],
    "19": [
     "<end_of_turn>",
     " \ub2f5\ubcc0",
     "ChatGPT",
     "\udb82\udc55",
     "\u0720\u0710",
     " \uc124\uba85",
     " explicando",
     " explica\u00e7\u00e3o"
    ],
    "20": [
     "<end_of_turn>",
     "<eos>",
     "]]></",
     "\udb82\udc55",
     "</img>",
     "\u25ae",
     "explanation",
     "ChatGPT"
    ],
    "21": [
     "<end_of_turn>",
     "ChatGPT",
     "\udb82\udc55",
     "Hopefully",
     "\u25ae",
     " \uad81\uae08",
     "feedback",
     "\u0720\u0710"
    ],
    "22": [
     "<end_of_turn>",
     "ChatGPT",
     "\u200d",
     "feedback",
     "\u5e0c\u671b",
     "Hopefully",
     "openai",
     "LinkedIn"
    ],
    "23": [
     "<end_of_turn>",
     "Would",
     "LinkedIn",
     "ChatGPT",
     "feedback",
     "GitHub",
     "Feedback",
     "Disclaimer"
    ],
    "24": [
     "<end_of_turn>",
     "Would",
     " Would",
     "Wouldn",
     "would",
     " wouldn",
     " Wouldn",
     "Did"
    ],
    "25": [
     "Would",
     " Would",
     "<end_of_turn>",
     "would",
     "Wouldn",
     " would",
     " Wouldn",
     "Why"
    ],
    "26": [
     "Would",
     " Would",
     "<end_of_turn>",
     " would",
     "would",
     "<eos>",
     "Wouldn",
     " Wouldn"
    ],
    "27": [
     " Would",
     "Would",
     "<end_of_turn>",
     "Why",
     " Why",
     "<eos>",
     "Create",
     " would"
    ],
    "28": [
     "<end_of_turn>",
     " Would",
     "Would",
     "Why",
     " Why",
     "would",
     " would",
     "<eos>"
    ],
    "29": [
     "<end_of_turn>",
     "Would",
     " Would",
     "Why",
     "<eos>",
     " \u0686\u06cc\u0646",
     "would",
     " Why"
    ],
    "30": [
     "<end_of_turn>",
     "Would",
     " Would",
     "<eos>",
     "Why",
     "would",
     " Why",
     " would"
    ],
    "31": [
     "<end_of_turn>",
     " Would",
     "Would",
     " would",
     "\n",
     "Em",
     " \u0686\u06cc\u0646",
     "Hot"
    ],
    "32": [
     "<end_of_turn>",
     " Would",
     "Would",
     "\n",
     " would",
     " Why",
     "Hot",
     "Je"
    ]
   }
  }
 ],
 "trajectories": [
  {
   "word": "flow",
   "position": 94,
   "layers": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19,
    20,
    21,
    22,
    23,
    24,
    25,
    26,
    27,
    28,
    29,
    30,
    31,
    32
   ],
   "ranks": [
    155687,
    143771,
    171549,
    199674,
    39518,
    107271,
    142738,
    150203,
    139902,
    150720,
    140928,
    117009,
    100675,
    156909,
    144015,
    123142,
    135196,
    138992,
    107759,
    139529,
    80057,
    71895,
    135835,
    124070,
    152163,
    172540,
    137816,
    58804,
    74551,
    22977,
    43619,
    28428,
    27048
   ]
  },
  {
   "word": "hum",
   "position": 94,
   "layers": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19,
    20,
    21,
    22,
    23,
    24,
    25,
    26,
    27,
    28,
    29,
    30,
    31,
    32
   ],
   "ranks": [
    110254,
    71462,
    27614,
    35662,
    3312,
    3094,
    5068,
    4967,
    2180,
    2598,
    1912,
    4650,
    4567,
    9578,
    6417,
    11261,
    8818,
    29976,
    54060,
    85290,
    68177,
    63783,
    55143,
    26577,
    12497,
    14180,
    16877,
    5076,
    12768,
    5618,
    4093,
    23030,
    7093
   ]
  },
  {
   "word": "nothing",
   "position": 94,
   "layers": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19,
    20,
    21,
    22,
    23,
    24,
    25,
    26,
    27,
    28,
    29,
    30,
    31,
    32
   ],
   "ranks": [
    209310,
    153651,
    176955,
    217641,
    66178,
    171872,
    217203,
    215070,
    173453,
    155958,
    140616,
    124604,
    106395,
    82940,
    82892,
    152423,
    98880,
    105402,
    138554,
    179715,
    116419,
    47788,
    46450,
    22499,
    25271,
    16882,
    40000,
    50962,
    99792,
    66333,
    61270,
    127023,
    142023
   ]
  }
 ],
 "emergence": {
  "position": 94,
  "top1": "<end_of_turn>",
  "layers": [
   0,
   1,
   2,
   3,
   4,
   5,
   6,
   7,
   8,
   9,
   10,
   11,
   12,
   13,
   14,
   15,
   16,
   17,
   18,
   19,
   20,
   21,
   22,
   23,
   24,
   25,
   26,
   27,
   28,
   29,
   30,
   31,
   32
  ],
  "ranks": [
   1,
   1,
   1,
   1,
   2,
   1,
   2,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   1,
   3,
   3,
   3,
   1,
   1,
   1,
   1,
   1
  ]
 },
 "scan": [],
 "slice": null,
 "film": "film.json"
}