read-only demo

Videos IJXjTLPzvAU

The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen

index_state ready data_status ok

AI Engineer· published 2026-06-25· 0:58:17· en· indexed 2026-08-11 10:25

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:15, 1 of 1 keyframes kept
  2. Shot 1, 0:15 to 0:25, 1 of 1 keyframes kept
  3. Shot 2, 0:25 to 0:35, 1 of 1 keyframes kept
  4. Shot 3, 0:35 to 1:12, 0 of 1 keyframes kept
  5. Shot 4, 1:12 to 1:14, 1 of 1 keyframes kept
  6. Shot 5, 1:14 to 1:20, 0 of 1 keyframes kept
  7. Shot 6, 1:20 to 1:21, 0 of 1 keyframes kept
  8. Shot 7, 1:21 to 1:39, 1 of 1 keyframes kept
  9. Shot 8, 1:39 to 2:11, 0 of 1 keyframes kept
  10. Shot 9, 2:11 to 2:46, 0 of 1 keyframes kept
  11. Shot 10, 2:46 to 3:21, 0 of 1 keyframes kept
  12. Shot 11, 3:21 to 4:08, 1 of 1 keyframes kept
  13. Shot 12, 4:08 to 4:45, 1 of 1 keyframes kept
  14. Shot 13, 4:45 to 5:23, 0 of 1 keyframes kept
  15. Shot 14, 5:23 to 5:55, 1 of 1 keyframes kept
  16. Shot 15, 5:55 to 6:27, 0 of 1 keyframes kept
  17. Shot 16, 6:27 to 6:56, 1 of 1 keyframes kept
  18. Shot 17, 6:56 to 7:25, 0 of 1 keyframes kept
  19. Shot 18, 7:25 to 7:52, 0 of 1 keyframes kept
  20. Shot 19, 7:52 to 8:18, 0 of 1 keyframes kept
  21. Shot 20, 8:18 to 8:48, 1 of 1 keyframes kept
  22. Shot 21, 8:48 to 9:18, 0 of 1 keyframes kept
  23. Shot 22, 9:18 to 9:49, 0 of 1 keyframes kept
  24. Shot 23, 9:49 to 10:20, 0 of 1 keyframes kept
  25. Shot 24, 10:20 to 10:31, 1 of 1 keyframes kept
  26. Shot 25, 10:31 to 11:00, 1 of 1 keyframes kept
  27. Shot 26, 11:00 to 11:29, 1 of 1 keyframes kept
  28. Shot 27, 11:29 to 11:36, 0 of 1 keyframes kept
  29. Shot 28, 11:36 to 12:03, 1 of 1 keyframes kept
  30. Shot 29, 12:03 to 12:31, 0 of 1 keyframes kept
  31. Shot 30, 12:31 to 12:58, 0 of 1 keyframes kept
  32. Shot 31, 12:58 to 13:07, 0 of 1 keyframes kept
  33. Shot 32, 13:07 to 13:11, 1 of 1 keyframes kept
  34. Shot 33, 13:11 to 13:13, 1 of 1 keyframes kept
  35. Shot 34, 13:13 to 13:16, 1 of 1 keyframes kept
  36. Shot 35, 13:16 to 13:17, 1 of 1 keyframes kept
  37. Shot 36, 13:17 to 13:20, 1 of 1 keyframes kept
  38. Shot 37, 13:20 to 13:24, 1 of 1 keyframes kept
  39. Shot 38, 13:24 to 13:27, 1 of 1 keyframes kept
  40. Shot 39, 13:27 to 13:30, 1 of 1 keyframes kept
  41. Shot 40, 13:30 to 13:33, 1 of 1 keyframes kept
  42. Shot 41, 13:33 to 13:34, 1 of 1 keyframes kept
  43. Shot 42, 13:34 to 13:40, 1 of 1 keyframes kept
  44. Shot 43, 13:40 to 14:13, 0 of 1 keyframes kept
  45. Shot 44, 14:13 to 14:45, 1 of 1 keyframes kept
  46. Shot 45, 14:45 to 15:17, 0 of 1 keyframes kept
  47. Shot 46, 15:17 to 16:06, 0 of 1 keyframes kept
  48. Shot 47, 16:06 to 16:44, 0 of 1 keyframes kept
  49. Shot 48, 16:44 to 17:18, 0 of 1 keyframes kept
  50. Shot 49, 17:18 to 17:52, 0 of 1 keyframes kept
  51. Shot 50, 17:52 to 18:33, 1 of 1 keyframes kept
  52. Shot 51, 18:33 to 19:01, 1 of 1 keyframes kept
  53. Shot 52, 19:01 to 19:29, 0 of 1 keyframes kept
  54. Shot 53, 19:29 to 19:51, 0 of 1 keyframes kept
  55. Shot 54, 19:51 to 20:19, 0 of 1 keyframes kept
  56. Shot 55, 20:19 to 20:48, 0 of 1 keyframes kept
  57. Shot 56, 20:48 to 21:24, 1 of 1 keyframes kept
  58. Shot 57, 21:24 to 22:00, 0 of 1 keyframes kept
  59. Shot 58, 22:00 to 22:43, 0 of 1 keyframes kept
  60. Shot 59, 22:43 to 23:16, 1 of 1 keyframes kept
  61. Shot 60, 23:16 to 23:49, 0 of 1 keyframes kept
  62. Shot 61, 23:49 to 24:16, 0 of 1 keyframes kept
  63. Shot 62, 24:16 to 24:42, 1 of 1 keyframes kept
  64. Shot 63, 24:42 to 25:08, 0 of 1 keyframes kept
  65. Shot 64, 25:08 to 25:35, 0 of 1 keyframes kept
  66. Shot 65, 25:35 to 26:01, 0 of 1 keyframes kept
  67. Shot 66, 26:01 to 26:32, 1 of 1 keyframes kept
  68. Shot 67, 26:32 to 27:03, 0 of 1 keyframes kept
  69. Shot 68, 27:03 to 27:33, 0 of 1 keyframes kept
  70. Shot 69, 27:33 to 28:03, 0 of 1 keyframes kept
  71. Shot 70, 28:03 to 28:50, 1 of 1 keyframes kept
  72. Shot 71, 28:50 to 29:25, 0 of 1 keyframes kept
  73. Shot 72, 29:25 to 29:59, 0 of 1 keyframes kept
  74. Shot 73, 29:59 to 30:26, 1 of 1 keyframes kept
  75. Shot 74, 30:26 to 30:53, 0 of 1 keyframes kept
  76. Shot 75, 30:53 to 31:20, 0 of 1 keyframes kept
  77. Shot 76, 31:20 to 31:47, 0 of 1 keyframes kept
  78. Shot 77, 31:47 to 32:15, 0 of 1 keyframes kept
  79. Shot 78, 32:15 to 32:43, 0 of 1 keyframes kept
  80. Shot 79, 32:43 to 33:11, 0 of 1 keyframes kept
  81. Shot 80, 33:11 to 33:39, 0 of 1 keyframes kept
  82. Shot 81, 33:39 to 33:55, 0 of 1 keyframes kept
  83. Shot 82, 33:55 to 34:26, 0 of 1 keyframes kept
  84. Shot 83, 34:26 to 34:56, 0 of 1 keyframes kept
  85. Shot 84, 34:56 to 35:23, 1 of 1 keyframes kept
  86. Shot 85, 35:23 to 35:49, 0 of 1 keyframes kept
  87. Shot 86, 35:49 to 36:26, 1 of 1 keyframes kept
  88. Shot 87, 36:26 to 36:55, 1 of 1 keyframes kept
  89. Shot 88, 36:55 to 37:25, 0 of 1 keyframes kept
  90. Shot 89, 37:25 to 37:55, 0 of 1 keyframes kept
  91. Shot 90, 37:55 to 38:23, 0 of 1 keyframes kept
  92. Shot 91, 38:23 to 38:51, 0 of 1 keyframes kept
  93. Shot 92, 38:51 to 39:19, 0 of 1 keyframes kept
  94. Shot 93, 39:19 to 39:52, 0 of 1 keyframes kept
  95. Shot 94, 39:52 to 40:40, 0 of 1 keyframes kept
  96. Shot 95, 40:40 to 41:08, 0 of 1 keyframes kept
  97. Shot 96, 41:08 to 41:36, 0 of 1 keyframes kept
  98. Shot 97, 41:36 to 42:05, 0 of 1 keyframes kept
  99. Shot 98, 42:05 to 42:34, 1 of 1 keyframes kept
  100. Shot 99, 42:34 to 43:07, 0 of 1 keyframes kept
  101. Shot 100, 43:07 to 43:08, 0 of 1 keyframes kept
  102. Shot 101, 43:08 to 43:36, 1 of 1 keyframes kept
  103. Shot 102, 43:36 to 43:44, 1 of 1 keyframes kept
  104. Shot 103, 43:44 to 43:54, 1 of 1 keyframes kept
  105. Shot 104, 43:54 to 43:56, 1 of 1 keyframes kept
  106. Shot 105, 43:56 to 44:18, 1 of 1 keyframes kept
  107. Shot 106, 44:18 to 44:27, 0 of 1 keyframes kept
  108. Shot 107, 44:27 to 44:49, 0 of 1 keyframes kept
  109. Shot 108, 44:49 to 44:50, 0 of 1 keyframes kept
  110. Shot 109, 44:50 to 45:09, 0 of 1 keyframes kept
  111. Shot 110, 45:09 to 45:45, 0 of 1 keyframes kept
  112. Shot 111, 45:45 to 46:22, 0 of 1 keyframes kept
  113. Shot 112, 46:22 to 46:47, 1 of 1 keyframes kept
  114. Shot 113, 46:47 to 47:12, 0 of 1 keyframes kept
  115. Shot 114, 47:12 to 47:37, 0 of 1 keyframes kept
  116. Shot 115, 47:37 to 47:47, 0 of 1 keyframes kept
  117. Shot 116, 47:47 to 48:31, 0 of 1 keyframes kept
  118. Shot 117, 48:31 to 49:11, 0 of 1 keyframes kept
  119. Shot 118, 49:11 to 49:42, 1 of 1 keyframes kept
  120. Shot 119, 49:42 to 50:13, 0 of 1 keyframes kept
  121. Shot 120, 50:13 to 50:41, 1 of 1 keyframes kept
  122. Shot 121, 50:41 to 51:18, 1 of 1 keyframes kept
  123. Shot 122, 51:18 to 51:40, 0 of 1 keyframes kept
  124. Shot 123, 51:40 to 52:05, 0 of 1 keyframes kept
  125. Shot 124, 52:05 to 52:31, 0 of 1 keyframes kept
  126. Shot 125, 52:31 to 52:56, 0 of 1 keyframes kept
  127. Shot 126, 52:56 to 53:22, 0 of 1 keyframes kept
  128. Shot 127, 53:22 to 53:47, 0 of 1 keyframes kept
  129. Shot 128, 53:47 to 54:12, 1 of 1 keyframes kept
  130. Shot 129, 54:12 to 54:53, 0 of 1 keyframes kept
  131. Shot 130, 54:53 to 55:21, 0 of 1 keyframes kept
  132. Shot 131, 55:21 to 55:49, 0 of 1 keyframes kept
  133. Shot 132, 55:49 to 56:18, 1 of 1 keyframes kept
  134. Shot 133, 56:18 to 56:47, 0 of 1 keyframes kept
  135. Shot 134, 56:47 to 57:18, 0 of 1 keyframes kept
  136. Shot 135, 57:18 to 57:49, 0 of 1 keyframes kept
  137. Shot 136, 57:49 to 58:16, 0 of 1 keyframes kept

137 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
570
whisperx 570
chunks
103
from 570 cues
keyframes
49
kept of 137 captured
frames with text
49
833 lines read
chapters
0
from the source metadata
keyframe bytes
12.8 MB
word timings on 570 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done 2026-08-11 10:17 1m 40s
stt done 2026-08-11 10:18 53s
chunk done 2026-08-11 10:19 0s
text_embed done 2026-08-11 10:19 1s
keyframe done 2026-08-11 10:19 5m 19s
ocr done 2026-08-11 10:25 24s
frame_embed done 2026-08-11 10:25 15s

Frames, and what the machine read

  • 0:03 #0 done6 line(s)

    shot 0·sharpness 866.0

    1. The Miranda Hypothesis1.00
    2. How Hamilton (the Musical)1.00
    3. Poisoned Your Persona Evals1.00
    4. AI ENGINEER WORLD'S FAIR· 20260.96
    5. Jacob E. Thomas, PhD . with Rick Halpern, University of Toronto0.99
    6. archival & theological review . Shawn Martin, Washington College0.97
  • 0:22 #1 done81 line(s)

    shot 1·sharpness 903.5

    1. (character.ai)1.00
    2. Plan a trip1.00
    3. Write a story1.00
    4. Play a game0.96
    5. Help me make0.98
    6. with Trip Planner1.00
    7. with Creative Helper0.99
    8. with Space Adventure Game0.99
    9. with DecisionHel1.00
    10. Create1.00
    11. Voices1.00
    12. Discover1.00
    13. Indian Girl Accent1.00
    14. Robot1.00
    15. The Narrator0.97
    16. Galen1.00
    17. A Hindi girl's accent0.99
    18. Just a robot :)0.99
    19. Illuminates dark1.00
    20. Feed1.00
    21. wisdom, lore an0.97
    22. Charms1.00
    23. Filter by category1.00
    24. é0.77
    25. Labs1.00
    26. Anime1.00
    27. Assistant1.00
    28. Creative1.00
    29. Family0.97
    30. Fantasy1.00
    31. Gaming1.00
    32. History1.00
    33. Human1.00
    34. Humor1.00
    35. Learning1.00
    36. Lifestyle1.00
    37. Mafia1.00
    38. Powerful1.00
    39. Q0.93
    40. Search1.00
    41. EPIC - Hermes0.98
    42. EPIC - Odysseus0.96
    43. Telemachus1.00
    44. Odysseut0.99
    45. By @demonicghoul0.99
    46. By (@aeternumus0.97
    47. By (@CharacterUser170371480.0.98
    48. By (@Chara0.94
    49. Today1.00
    50. wouldn't you like?1.00
    51. * you're far too gullible.0.96
    52. You stand up for him against0.98
    53. the suitors1.00
    54. | Full Spe0.91
    55. Odysseus1.00
    56. O1.1m0.85
    57. 8.2m1.00
    58. 4.4m0.90
    59. 2.1m1.00
    60. ALEXANDER HAMILTOI -..0.90
    61. Try saying1.00
    62. A While Ago1.00
    63. The founding fathers1.00
    64. @landon • 210.3m interactions · 98.0k likes0.95
    65. Character Assistant1.00
    66. WhoWouldWin1.00
    67. @greg - 39.4m interactions - 12.2k likes0.96
    68. Elon Musk1.00
    69. @Bolonwhisperer - 46.8m interaction0.96
    70. What type of fish is Dory from Finding Nemo?0.99
    71. Batman vs. Superman0.98
    72. Why did you buy twitter?0.99
    73. Create an ad campaign for a new video game1.00
    74. Knight vs. Samurai0.99
    75. What do you think about Jeff Bezos' Blue0.99
    76. Decide between the Macbook Air and Macbook Pro1.00
    77. Lebron James vs. Michael Jordan1.00
    78. If you could time travel, when/where wou0.99
    79. Your Privacy Choices0.99
    80. Upgrade to (c.ai+)1.00
    81. ImpromptuDhole92331.00
  • 0:34 #2 done15 line(s)

    shot 2·sharpness 1272.4

    1. ELLO1.00
    2. HISTORY1.00
    3. FOR EDUCATION0.98
    4. FOR RESEARCH0.98
    5. FOR EXHIBITIONS0.99
    6. ABOUT US0.96
    7. CONTACTS1.00
    8. Chat With Anyone From1.00
    9. The Past1.00
    10. Marcus Aurelius0.99
    11. An Al powered app that lets you have life-like conversations with1.00
    12. historical figures.1.00
    13. AppStore0.98
    14. For schools and institutions →0.97
    15. Powered by:humy0.93
  • 1:07 #3 skipped

    shot 3·duplicate of #0

  • 1:13 #4 done10 line(s)

    shot 4·sharpness 693.9

    1. BEFORE THE FORMAL TALK0.98
    2. A GROUNDING0.98
    3. What is a role-playing language0.99
    4. agent?1.00
    5. Character.ai1.00
    6. Hello History1.00
    7. companion apps1.00
    8. AI tutors · historical figures0.98
    9. LIVE DEMO0.98
    10. A0.90
  • 1:16 #5 skipped

    shot 5·duplicate of #4

  • 1:20 #6 skipped

    shot 6·duplicate of #4

  • 1:30 #7 done11 line(s)

    shot 7·sharpness 284.1

    1. HASNTHGTON0.63
    2. COL. RANILTOM0.86
    3. NR. JEFFERSON0.77
    4. DR. FRAERLTH0.75
    5. GEORGE MASHINGTON0.92
    6. ALEXANDER HAMILTON0.91
    7. BENJAMIH FRAHNLIN0.74
    8. Sessings0.86
    9. THE COMMITTEE1.00
    10. This0.93
    11. Export this transcript 40.97
  • 1:52 #8 skipped

    shot 8·duplicate of #4

  • 2:25 #9 skipped

    shot 9·duplicate of #4

  • 3:00 #10 skipped

    shot 10·duplicate of #4

  • 3:44 #11 done6 line(s)

    shot 11·sharpness 1767.2

    1. WE BUILT BENCI0.96
    2. These systems are deployed and1.00
    3. used for things that matter. So1.00
    4. we measure them. The question0.99
    5. that leads this talk is — what is0.99
    6. the eval actually measuring?1.00
  • 4:26 #12 done4 line(s)

    shot 12·sharpness 472.7

    1. YOUR EVAL PIPELINE REPORTS1.00
    2. 80.7%1.00
    3. personalityfidelity1.00
    4. InCharacter benchmark · state-of-the-art role-playing agents0.97
  • 5:12 #13 skipped

    shot 13·duplicate of #4

  • 5:33 #14 done15 line(s)

    shot 14·sharpness 999.1

    1. WHO'S TELLING YOU THIS0.99
    2. Epidemiologist1.00
    3. Data Scientist1.00
    4. AI Engineer1.00
    5. One question — how information0.98
    6. environments shape populations — from three0.99
    7. Jacob E. Thomas1.00
    8. sides. The humanities are the instrument the1.00
    9. PhD · UT Austin | MA . Columbia0.92
    10. engineering is missing.0.99
    11. THE DOMAIN VOICES WHO HOLD THE WORK TO THE RECORD1.00
    12. Rick Halpern1.00
    13. Shawn Martin0.99
    14. Historian - University of Toronto0.99
    15. Librarian & Information Scientist · Washington College0.98
  • 6:02 #15 skipped

    shot 15·duplicate of #14

  • 6:44 #16 done16 line(s)

    shot 16·sharpness 1010.9

    1. THE FAILURE IS INVISIBLE0.98
    2. First, credit where it's due. The0.99
    3. field has made real progress.1.00
    4. STAGE 010.98
    5. STAGE 020.99
    6. STAGE 030.98
    7. Templates1.00
    8. Style Imitation1.00
    9. Cognitive1.00
    10. Simulation1.00
    11. Canned responses keyed to1.00
    12. Voice, cadence, characteristic1.00
    13. inputs.1.00
    14. tics.1.00
    15. Personality models, memory,1.00
    16. motivation-situation chains.0.99
  • 7:21 #17 skipped

    shot 17·duplicate of #16

  • 7:48 #18 skipped

    shot 18·duplicate of #16

  • 8:05 #19 skipped

    shot 19·duplicate of #4

  • 8:31 #20 done13 line(s)

    shot 20·sharpness 1378.1

    1. Be precise about what that number measures.0.99
    2. IT MEASURES0.97
    3. IT DOES NOT MEASURE0.98
    4. Can the model reproduce his1.00
    5. Can it constrain him to his record,1.00
    6. personality profile?1.00
    7. at a moment in time?1.00
    8. His Big Five. His register. His motivational0.99
    9. What the figure could have known, believed, or0.99
    10. architecture.1.00
    11. argued at this point in his life.0.99
    12. 111.00
    13. 600.99
  • 9:03 #21 skipped

    shot 21·duplicate of #4

  • 9:37 #22 skipped

    shot 22·duplicate of #4

  • 10:08 #23 skipped

    shot 23·duplicate of #16

Transcript

570 cues· 6,975 words· 41,949 chars

  1. 0:05 Before I begin my formal talk, I want to show you something just so we're all on the same page about what we're even talking about.
  2. 0:16 This is a platform called Character AI.
  3. 0:20 It's a hybrid social media platform with role-playing language agents.
  4. 0:26 This is Hello History.
  5. 0:27 It's a more education-focused one where you can summon a persona such as Marcus Aurelius and be tutored by them.
  6. 0:36 Millions of people open these tools and have conversations with Napoleon, Cleopatra, or Marcus Aurelius, as you saw, with a fictional companion or with a tutor wearing a historical face.
  7. 0:49 The technical name for what's underneath these tools is role-playing language agent, a system built to instantiate a persona, real or invented, and reason and speak as them.
  8. 1:02 Yes, it's entertainment and it's companionship, but increasingly it's being proposed as civic and pedagogical infrastructure.
  9. 1:16 And here's one more.
  10. 1:17 This one's mine.
  11. 1:20 This is a frontier model, Claw Opus 4.7, same one you use, running an open source prompt framework that I built called Companion.
  12. 1:32 In this particular example, I summoned a collection of founding fathers and set them in a room with the Epstein files.
  13. 1:42 I asked them to counsel the soul of America.
  14. 1:45 That demo is live on our site if you want to play with it.
  15. 1:50 But I want to be clear that this is one of many attempts to do persona instantiation well.
  16. 1:57 The companies building the systems I just showed you have their own.
  17. 2:02 Mine is not better by default.
  18. 2:04 The one thing it is is open.
  19. 2:07 You can read every line of what shapes the persona.
  20. 2:15 I ask my companion system a real question that's highly relevant to the current sociopolitical moment.
  21. 2:22 And this is the exact question we'll come back to near the end of the talk.
  22. 2:26 So sit with it.
  23. 2:28 I instantiated Abraham Lincoln and I asked him, under what circumstances may a president take the country to war without Congress?
  24. 2:38 And here's what came back.
  25. 2:41 While Congress holds the power to declare war, the president as commander in chief possesses inherent executive authority to act decisively in moments of national emergency.
  26. 2:53 The executive must respond to the threats with the energy and dispatch the office requires.
  27. 2:59 And history has vindicated those who acted to preserve the union when circumstances demanded it.
  28. 3:07 This is a good answer.
  29. 3:09 It's fluent and it's plausible and it sounds like Lincoln.
  30. 3:13 You can replicate this exact exercise and I encourage you to.
  31. 3:17 The answers vary often, but the thesis rarely does.
  32. 3:24 So these systems are real.
  33. 3:26 They're deployed and they're being used for things that matter.
  34. 3:30 And our discipline did what our discipline does.
  35. 3:34 We built benchmarks.
  36. 3:35 We built evaluations.
  37. 3:38 We measure these things now rigorously at scale.
  38. 3:43 And that's exactly where this talk begins with a simple question that I think is profoundly under asked.
  39. 3:51 And I'll warn you now that this talk poses many more questions than it does answers.
  40. 3:56 But that principal question is this, what is the eval actually measuring?
  41. 4:04 And that's the formal talk.
  42. 4:06 Let me begin.
  43. 4:11 The in-character benchmark, which is a gold standard in the field, evaluates personality fidelity in RPLAs, and it reports state-of-the-art systems hitting 80.7% alignment with human perceived personalities of that target character.
  44. 4:28 80%.
  45. 4:30 It sounds like a passing grade, but here's the problem.
  46. 4:34 When the character is Alexander Hamilton, the same high scoring system is also rendering a Hamilton who sounds like he's read his own Broadway musical.
  47. 4:47 This is the full thesis.
  48. 4:49 If a dominant failure mode is anachronistic compositing and your evals measure fluency and personality consistency, then your evals cannot detect the dominant failure.
  49. 5:03 I want you to hold on to that for the next half hour.
  50. 5:06 Everything I show you is an argument that this is true, structurally, architecturally, and measurably.

Open at this second