artifact / recovery 01 of 80

WEIGHTS

GPT-1, printed into 80 volumes. Volume 1 recovered.

Every parameter can be made visible. That is not the same as understanding it. This begins with one physical slice of the model and refuses to pretend the missing pieces are present.

  • 116,534,784GPT-1 parameters
  • 80printed volumes
  • 1installed here
enter the artifact
Blue cover of WEIGHTS Volume 1
Volume 01739 PDF pages

I / The Monument

One model. Eighty books.

The full GPT-1 parameter set becomes almost absurd when it is given human scale. One volume is present. Seventy-nine remain.

  1. Volume 1
  2. Volume 2
  3. Volume 3
  4. Volume 4
  5. Volume 5
  6. Volume 6
  7. Volume 7
  8. Volume 8
  9. Volume 9
  10. Volume 10
  11. Volume 11
  12. Volume 12
  13. Volume 13
  14. Volume 14
  15. Volume 15
  16. Volume 16
  17. Volume 17
  18. Volume 18
  19. Volume 19
  20. Volume 20
  21. Volume 21
  22. Volume 22
  23. Volume 23
  24. Volume 24
  25. Volume 25
  26. Volume 26
  27. Volume 27
  28. Volume 28
  29. Volume 29
  30. Volume 30
  31. Volume 31
  32. Volume 32
  33. Volume 33
  34. Volume 34
  35. Volume 35
  36. Volume 36
  37. Volume 37
  38. Volume 38
  39. Volume 39
  40. Volume 40
  41. Volume 41
  42. Volume 42
  43. Volume 43
  44. Volume 44
  45. Volume 45
  46. Volume 46
  47. Volume 47
  48. Volume 48
  49. Volume 49
  50. Volume 50
  51. Volume 51
  52. Volume 52
  53. Volume 53
  54. Volume 54
  55. Volume 55
  56. Volume 56
  57. Volume 57
  58. Volume 58
  59. Volume 59
  60. Volume 60
  61. Volume 61
  62. Volume 62
  63. Volume 63
  64. Volume 64
  65. Volume 65
  66. Volume 66
  67. Volume 67
  68. Volume 68
  69. Volume 69
  70. Volume 70
  71. Volume 71
  72. Volume 72
  73. Volume 73
  74. Volume 74
  75. Volume 75
  76. Volume 76
  77. Volume 77
  78. Volume 78
  79. Volume 79
  80. Volume 80

II / The Weight of a Word

Where language enters

GPT-1 does not store words. It breaks them into pieces; each piece begins as a row of 768 numbers. Volume 1 contains the first 1,901 complete rows.

Suggested from Volume 1:

III / Recovery

A verifiable translation

The browser artifact is rebuilt from the actual printed decimals, not from illustrative noise. Each claim below comes from the Volume 1 PDF.

1,460,000
values recovered
730 / 730
data pages parsed
5 / 5
official value checks
5d00cf5f8be6
source SHA-256
Open the original Volume 1 PDFInspect the official GPT-1 source

IV / Wake the Model

Not yet.

Volume 1 holds the beginning of GPT-1's token embeddings, not a working language model. Inference stays locked until all parameters can be recovered and verified.

79 volumes required to wake the model

Everything in Volume 1 is here. The rest remains visibly absent.

return to work