vidtheque.

read-only demo

Videos GgLQ02aO-hs

The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest & Isaac Miller

index_state ready data_status ok

AI Engineer· published 2026-07-23· 0:17:11· en-US· indexed 2026-08-10 19:40

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:03, 1 of 1 keyframes kept
  2. Shot 1, 0:03 to 0:05, 1 of 1 keyframes kept
  3. Shot 2, 0:05 to 0:12, 1 of 1 keyframes kept
  4. Shot 3, 0:12 to 0:30, 1 of 1 keyframes kept
  5. Shot 4, 0:30 to 0:35, 1 of 1 keyframes kept
  6. Shot 5, 0:35 to 0:41, 1 of 1 keyframes kept
  7. Shot 6, 0:41 to 0:42, 1 of 1 keyframes kept
  8. Shot 7, 0:42 to 0:45, 1 of 1 keyframes kept
  9. Shot 8, 0:45 to 0:46, 1 of 1 keyframes kept
  10. Shot 9, 0:46 to 1:17, 1 of 1 keyframes kept
  11. Shot 10, 1:17 to 1:49, 0 of 1 keyframes kept
  12. Shot 11, 1:49 to 2:21, 1 of 1 keyframes kept
  13. Shot 12, 2:21 to 2:47, 1 of 1 keyframes kept
  14. Shot 13, 2:47 to 3:13, 0 of 1 keyframes kept
  15. Shot 14, 3:13 to 3:38, 1 of 1 keyframes kept
  16. Shot 15, 3:38 to 4:04, 1 of 1 keyframes kept
  17. Shot 16, 4:04 to 4:09, 1 of 1 keyframes kept
  18. Shot 17, 4:09 to 4:41, 1 of 1 keyframes kept
  19. Shot 18, 4:41 to 5:09, 1 of 1 keyframes kept
  20. Shot 19, 5:09 to 5:37, 1 of 1 keyframes kept
  21. Shot 20, 5:37 to 6:06, 1 of 1 keyframes kept
  22. Shot 21, 6:06 to 6:42, 1 of 1 keyframes kept
  23. Shot 22, 6:42 to 7:18, 0 of 1 keyframes kept
  24. Shot 23, 7:18 to 7:46, 1 of 1 keyframes kept
  25. Shot 24, 7:46 to 8:14, 0 of 1 keyframes kept
  26. Shot 25, 8:14 to 8:43, 0 of 1 keyframes kept
  27. Shot 26, 8:43 to 9:11, 0 of 1 keyframes kept
  28. Shot 27, 9:11 to 9:45, 1 of 1 keyframes kept
  29. Shot 28, 9:45 to 10:20, 1 of 1 keyframes kept
  30. Shot 29, 10:20 to 10:49, 1 of 1 keyframes kept
  31. Shot 30, 10:49 to 11:19, 1 of 1 keyframes kept
  32. Shot 31, 11:19 to 11:48, 0 of 1 keyframes kept
  33. Shot 32, 11:48 to 12:24, 1 of 1 keyframes kept
  34. Shot 33, 12:24 to 12:53, 1 of 1 keyframes kept
  35. Shot 34, 12:53 to 13:23, 0 of 1 keyframes kept
  36. Shot 35, 13:23 to 13:52, 0 of 1 keyframes kept
  37. Shot 36, 13:52 to 14:18, 1 of 1 keyframes kept
  38. Shot 37, 14:18 to 14:44, 0 of 1 keyframes kept
  39. Shot 38, 14:44 to 15:10, 0 of 1 keyframes kept
  40. Shot 39, 15:10 to 15:47, 1 of 1 keyframes kept
  41. Shot 40, 15:47 to 16:24, 0 of 1 keyframes kept
  42. Shot 41, 16:24 to 16:54, 1 of 1 keyframes kept
  43. Shot 42, 16:54 to 17:09, 0 of 1 keyframes kept
  44. Shot 43, 17:09 to 17:10, 1 of 1 keyframes kept

44 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
187
whisperx 187
chunks
30
from 187 cues
keyframes
31
kept of 44 captured
frames with text
30
903 lines read
chapters
0
from the source metadata
keyframe bytes
6.2 MB
word timings on 187 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-09 23:01 0s
stt done — 2026-08-09 13:36 17s
chunk done — 2026-08-09 13:37 0s
text_embed done — 2026-08-10 19:40 0s
keyframe done — 2026-08-09 13:37 2m 38s
ocr done — 2026-08-09 13:39 16s
frame_embed done — 2026-08-10 19:40 5s

Frames, and what the machine read

  • 0:02 #0 done2 line(s)

    shot 0·sharpness 456.7

    1. AlEngineer0.96
    2. World's Fair0.97

  • 0:03 #1 done2 line(s)

    shot 1·sharpness 663.3

    1. AlEngineer0.95
    2. World's Fair0.99

  • 0:10 #2 done25 line(s)

    shot 2·sharpness 2742.9

    1. LAB & PLATINUM SPONSORS0.99
    2. Amazon AGI Lab0.97
    3. ANTHROP\C1.00
    4. Google DeepMind1.00
    5. MINIMAX0.98
    6. OpenAl0.89
    7. Akamai1.00
    8. arize1.00
    9. aws1.00
    10. Braintrust1.00
    11. bright data0.99
    12. B1.00
    13. Browserbase1.00
    14. docker1.00
    15. :neo4j0.90
    16. ORACLE1.00
    17. PayPal1.00
    18. qodo1.00
    19. reducto1.00
    20. Sonar1.00
    21. Makers of1.00
    22. togetherai1.00
    23. Unblocked1.00
    24. WorkOS1.00
    25. SonarQube1.00

  • 0:21 #3 done71 line(s)

    shot 3·sharpness 2854.2

    1. Modal1.00
    2. Ref.1.00
    3. :neo4]0.86
    4. World's Fair0.99
    5. World's Fair0.95
    6. World's Fair0.97
    7. World's Fair0.95
    8. World's Fair0.98
    9. neer1.00
    10. THE VELOCITY ROOM0.95
    11. mezmo*0.94
    12. bright data0.72
    13. ©fiddler0.91
    14. World's Fair0.99
    15. Cace0.57
    16. O INNGEST0.90
    17. PRIOR1.00
    18. SnorkeJ0.97
    19. Wort0.94
    20. World's Fair0.98
    21. BAND1.00
    22. Keycard0.99
    23. Ateirpx0.58
    24. World's Fair0.99
    25. World's Fair0.98
    26. :neo4j0.98
    27. World's Fair0.97
    28. Cleric1.00
    29. Meliculous0.96
    30. World's Fair0.97
    31. Google DeepMind1.00
    32. World's Fair0.97
    33. qodo0.97
    34. World's Fair0.95
    35. PayPal1.00
    36. À ATLASSIAN0.95
    37. FACTORY1.00
    38. Braintrust1.00
    39. World's Fair0.98
    40. World's Fair0.92
    41. Microsoft1.00
    42. World's Fair0.97
    43. arize1.00
    44. ORACLE1.00
    45. Microsoht0.89
    46. ANTHROPIC1.00
    47. ddruid0.61
    48. Crusce0.86
    49. World's Fair0.99
    50. MINIMAX1.00
    51. orld's Fair0.94
    52. World's Fair0.94
    53. World's Fai0.95
    54. ANTHROPIC1.00
    55. World's Fair0.94
    56. Resolve.ai0.96
    57. Workd's Fair0.91
    58. co0.72
    59. Optiver0.99
    60. World's Fair0.99
    61. ouilder0.92
    62. World's Fair0.96
    63. Ravenna1.00
    64. World's Fai0.96
    65. CopilotKit1.00
    66. Id's Fair0.95
    67. /orld'sFa0.99
    68. World's Fair0.98
    69. BeNCORD0.91
    70. fsFair0.97
    71. tgus0.93

  • 0:34 #4 done144 line(s)

    shot 4·sharpness 3996.8

    1. World'sFair0.94
    2. RELAI0.89
    3. World's Fair0.99
    4. World's Fair0.96
    5. Modal0.99
    6. World's Fair0.99
    7. promptqt0.99
    8. World'sFair0.95
    9. SurrealDB0.97
    10. World's Fair0.97
    11. ZERO0.94
    12. World's Fair0.97
    13. AlEngineer0.95
    14. comet1.00
    15. Wo0.98
    16. Norld's Fair0.96
    17. SOIO.IO0.91
    18. World's Fair0.94
    19. granica1.00
    20. World's Fair0.99
    21. World's Fair1.00
    22. DigitalOcean0.97
    23. World's Fair0.99
    24. Venice0.90
    25. World's Fair0.99
    26. POSTMAN1.00
    27. Wo1.00
    28. Norld's Fair0.95
    29. Modular1.00
    30. World's Fair0.96
    31. MERGE1.00
    32. World's Fair0.98
    33. AUT1.00
    34. yugabyteDB1.00
    35. World's Fair0.96
    36. Zed0.99
    37. World's Fair0.99
    38. Keycard1.00
    39. Wo0.99
    40. World's Fair0.95
    41. Daytona0.94
    42. World's Fair0.95
    43. #twilio0.92
    44. World's Fair0.94
    45. GG0.99
    46. Temporal1.00
    47. World's Fair0.99
    48. I Llamalndex0.96
    49. World's Fair0.95
    50. O INNGEST0.89
    51. World's Fair0.97
    52. BAND1.00
    53. World's Fair0.93
    54. Cleric1.00
    55. World's Fair0.99
    56. À ATLASSIAN0.95
    57. Wo0.99
    58. World's Fair0.97
    59. baseten1.00
    60. World's Fair0.95
    61. Snorkel0.99
    62. World's Fair0.97
    63. Z.AI0.98
    64. World's Fair0.96
    65. qodo0.95
    66. World's Fair0.99
    67. PayPal1.00
    68. World's Fair0.98
    69. G1.00
    70. azon AGI Lab1.00
    71. World's Fair0.97
    72. Browserbase1.00
    73. World's Fair0.97
    74. :neo4j0.98
    75. World's Fair0.98
    76. Google DeepMind1.00
    77. World's Fair0.99
    78. arize1.00
    79. World's Fair0.98
    80. reducto1.00
    81. Wo0.99
    82. World's Fair0.95
    83. OpenAl0.96
    84. World's Fair0.98
    85. WorkOs0.98
    86. World's Fair0.97
    87. Amazon AG:0.99
    88. World's Fair0.96
    89. Microsoft1.00
    90. World's Fair0.95
    91. ORACLE0.98
    92. World's Fair0.96
    93. bri1.00
    94. Microsoft1.00
    95. World's Fair0.97
    96. docker1.00
    97. World'sF1.00
    98. Braintrust1.00
    99. World's Fair0.94
    100. OpenAl0.97
    101. World's Fair0.99
    102. MINIMAX1.00
    103. World's Fair0.96
    104. Z.AI0.98
    105. Wo0.99
    106. World's Fair0.96
    107. snyk1.00
    108. World's Fair0.99
    109. Id's Fair0.89
    110. togetheral0.96
    111. orld's Fair0.97
    112. Akamai1.00
    113. World's Fair0.98
    114. Unblocked1.00
    115. Worldia0.81
    116. Lightrun1.00
    117. ain0.99
    118. World'sFa1.00
    119. LOUDF0.99
    120. World'0.95
    121. DATADOG1.00
    122. World's Fair0.99
    123. Resolve.ai1.00
    124. World's Fair0.98
    125. Wo0.96
    126. Norld's Fair0.95
    127. World's Fair0.94
    128. LanceL0.97
    129. orld's Fair0.97
    130. builder.io0.91
    131. Norld's Fair0.92
    132. Ravenna1.00
    133. World's Fair0.91
    134. CopilotKit1.00
    135. descope0.98
    136. Microsoft0.98
    137. t1.00
    138. World's Fai0.93
    139. TOPK1.00
    140. World's Fair0.97
    141. cognee1.00
    142. World's Fair0.96
    143. eNCORD0.97
    144. World's Fair0.97

  • 0:37 #5 done11 line(s)

    shot 5·sharpness 622.2

    1. AlEngineer0.95
    2. WorkOS1.00
    3. World's Fair0.96
    4. AlEngineer0.97
    5. World'sF1.00
    6. Braintrust1.00
    7. AlEngineer0.99
    8. aw1.00
    9. rld'sFair1.00
    10. AlEngineer0.97
    11. WMooVoF0.60

  • 0:41 #6 done11 line(s)

    shot 6·sharpness 466.6

    1. AlEngineer0.99
    2. Amazon AGI Lab1.00
    3. World'sFair1.00
    4. AlEngineer0.96
    5. OpenAl0.92
    6. World'sFair1.00
    7. together.0.99
    8. husFair0.95
    9. AlEngineer0.95
    10. DATADOG0.98
    11. WooF0.52

  • 0:44 #7 done12 line(s)

    shot 7·sharpness 472.3

    1. AlEngineer0.96
    2. Microsoft1.00
    3. World's Fair0.97
    4. AlEngineer0.96
    5. n MINIMAX0.86
    6. World's Fa0.95
    7. AlEngineer0.97
    8. Akar0.99
    9. Id'sFair0.99
    10. AlEnginee0.92
    11. Pesolve ai0.94
    12. WAooe0.52

  • 0:46 #8 done22 line(s)

    shot 8·sharpness 1180.5

    1. AlEngineer1.00
    2. World's Fair0.99
    3. Repeated logic becomes1.00
    4. afunction.1.00
    5. DUPLICATION1.00
    6. A NAMED FUNCTION1.00
    7. total = price_a + price_a * 0.150.99
    8. def with_tax(price):1.00
    9. total = price_b + price_b * 0.150.98
    10. return price * 1.150.99
    11. total = price_c + price_c * 0.150.99
    12. total = with_tax(price_a)0.99
    13. AIEr0.89
    14. Micros1.00
    15. Work0.97
    16. - AlEngineer0.93
    17. Mn0.62
    18. rld's'0.85
    19. AIEr0.85
    20. Ak1.00
    21. lorke0.73
    22. - AIE0.88

  • 1:11 #9 done22 line(s)

    shot 9·sharpness 2561.0

    1. AlEngineer0.99
    2. World's Fair0.98
    3. Al programs should be0.99
    4. functions.1.00
    5. PRESENTED BY0.99
    6. Reusable.1.00
    7. Microsoft1.00
    8. Composable.1.00
    9. Testable.1.00
    10. Optimizable.1.00
    11. OS0.69
    12. World's Fair0.95
    13. air1.00
    14. Braintrus0.98
    15. Engineer1.00
    16. The Unreasonable Effectiveness of Separating the Task from the Model1.00
    17. 'd's Fair0.98
    18. Maxime Rivest0.99
    19. Isaac Miller1.00
    20. DSPy1.00
    21. Core Contributor1.00
    22. Lead Maintainer1.00

  • 1:42 #10 skipped

    shot 10·duplicate of #9

  • 2:02 #11 done36 line(s)

    shot 11·sharpness 5080.5

    1. prompting, prompt engineering, chain-of-thought, few-shot, zero-shot, one-shot, system1.00
    2. prompt, developer message, instruction hierarchy, routing, context engineering, harness1.00
    3. AlEngineer1.00
    4. World's Fair0.98
    5. engineering, in-context learning, exemplars, demonstrations, self-critique, self-reflection,1.00
    6. self-consistency, ensemble, best-of-n, RAG, structured outputs, mixture of experts, guardrails,0.99
    7. alignment, prompt injection, prompt leaking, instruction override, agentic, agents, agent loop,1.00
    8. ReAct, tool calling, function calling, tool schema, computer use, agent swarm, verifier, judge,0.99
    9. memory, MCP, embeddings, vector store, semantic search, chunking, chunk size, sliding1.00
    10. PRESENTED BY0.98
    11. window, reranking, context window, long context, compaction, pre-training, post-training,0.99
    12. Microsoft0.99
    13. fine-tuning, SFT, reinforcement learning, RLHF, reward model, preference data, evals, DPO,1.00
    14. ORPO, KTO, GRPO, distillation, teacher model, student model, multi-agent system, LoRA,1.00
    15. QLoRA, adapters, PEFT, quantization, checkpoints, open weights, frontier model, base model,1.00
    16. instruct model, inference, decode, sampling, temperature, top-p, top-k, beam search,1.00
    17. logprobs, KV cache, paged attention, flash attention, speculative decoding, test-time compute,0.99
    18. reasoning models, loop engineering, subagents, orchestrator, planner, scratchpad, retrieval,1.00
    19. orld's Fair0.95
    20. Amazor1.00
    21. grounding, hallucination, citations, batching, streaming, caching, prompt caching, tracing,0.99
    22. observability, guard models, red-teaming, jailbreaks, curriculum, synthetic data, self-play,1.00
    23. AIEr0.79
    24. verifier loops, reflection, tree search, MCTS, graph-of-thought, skeleton-of-thought,0.99
    25. Braintru0.99
    26. Work0.93
    27. constitutionaLAL RI AIE nrocess reward outcome reward tokenizer multimodal VI M STT0.95
    28. The Unreasonable Effectiveness of Separating the Task from the Model0.98
    29. orld's0.98
    30. Maxime Rivest1.00
    31. Isaac Miller1.00
    32. DSPy1.00
    33. Wod.0.61
    34. AIEr0.85
    35. Core Contributor1.00
    36. Lead Maintainer1.00

  • 2:44 #12 done23 line(s)

    shot 12·sharpness 2578.4

    1. AlEngineer0.99
    2. World's Fair0.99
    3. Every repeated Al task deserves0.99
    4. aclearinput/outputinterface.1.00
    5. PRESENTED BY0.98
    6. invoice tax extraction0.96
    7. Microsoft1.00
    8. rld's Fair1.00
    9. Amazon0.99
    10. a signature: well-defined input, well-defined output, repeatable objective1.00
    11. AIEn0.93
    12. Braint0.99
    13. Work0.91
    14. - AlEngi0.90
    15. The Unreasonable Effectiveness of Separating the Task from the Model0.99
    16. orld'0.90
    17. tog1.00
    18. Maxime Rivest1.00
    19. Isaac Miller1.00
    20. DSPy1.00
    21. AlEn0.93
    22. Core Contributor0.98
    23. Lead Maintainer1.00

  • 3:10 #13 skipped

    shot 13·duplicate of #12

  • 3:30 #14 done27 line(s)

    shot 14·sharpness 3119.2

    1. AlEngineer1.00
    2. World's Fair0.98
    3. Every repeated Al task deserves0.99
    4. aclearinput/outputinterface.1.00
    5. PRESENTED BY0.96
    6. invoice tax extraction0.95
    7. text→ grammar correction0.96
    8. Microsoft1.00
    9. text rewrite for clarity0.97
    10. articleproduct classification1.00
    11. entire inbox + new email drafted reply0.98
    12. 'orld's Fr0.93
    13. Amazo1.00
    14. a signature: well-defined input, well-defined output, repeatable objective1.00
    15. Al0.79
    16. Braintr1.00
    17. Wor1.00
    18. AlEnginee0.98
    19. The Unreasonable Effectiveness of Separating the Task from the Model0.99
    20. 'orld's0.94
    21. to1.00
    22. Maxime Rivest1.00
    23. Isaac Miller1.00
    24. DSPy1.00
    25. Core Contributor1.00
    26. Lead Maintainer1.00
    27. LAL0.57

  • 3:56 #15 done26 line(s)

    shot 15·sharpness 3276.7

    1. AlEngineer0.98
    2. World's Fair0.97
    3. Every repeated Al task deserves0.99
    4. aclearinput/outputinterface.1.00
    5. PRESENTED BY1.00
    6. invoice tax extraction0.97
    7. text grammar correction0.97
    8. Microsoft1.00
    9. text rewrite for clarity0.98
    10. article product classification0.98
    11. entire inbox + new email drafted reply0.99
    12. design spec + repository pull request0.97
    13. WorkOs0.92
    14. Wc0.92
    15. a signature: well-defined input, well-defined output, repeatable objective1.00
    16. AlEngineer0.99
    17. World'1.00
    18. #E0.74
    19. The Unreasonable Effectiveness of Separating the Task from the Model0.99
    20. Maxime Rivest1.00
    21. Isaac Miller1.00
    22. DSPy1.00
    23. AlEngine0.96
    24. Core Contributor1.00
    25. Lead Maintainer1.00
    26. A_J0.56

  • 4:07 #16 done22 line(s)

    shot 16·sharpness 2525.4

    1. AlEngineer0.98
    2. World's Fair0.99
    3. The signature defines1.00
    4. the taskboundary.0.97
    5. PRESENTED BY0.96
    6. Microsoft1.00
    7. def extract_taxes(invoice) -> taxes: # the what - fixed0.99
    8. # the how - completely free to evolve0.98
    9. 'orkOS0.95
    10. World's1.00
    11. ngineer1.00
    12. d's Fair0.84
    13. Braint1.00
    14. IEnginee0.93
    15. The Unreasonable Effectiveness of Separating the Task from the Model1.00
    16. WS0.91
    17. Maxime Rivest1.00
    18. Isaac Miller1.00
    19. DSPy1.00
    20. gineer1.00
    21. Core Contributor1.00
    22. Lead Maintainer1.00

  • 4:31 #17 done28 line(s)

    shot 17·sharpness 2698.2

    1. AlEngineer0.97
    2. World'sFair1.00
    3. Everything behind the boundary1.00
    4. isreplaceable.1.00
    5. PRESENTED BY0.96
    6. Microsoft1.00
    7. def extract_taxes(invoice) -> taxes:0.99
    8. prompt1.00
    9. agent1.00
    10. tools1.00
    11. loops0.99
    12. different models1.00
    13. optimizers1.00
    14. rkOs0.95
    15. World's F0.99
    16. Everything else remains unchanged.1.00
    17. eer1.00
    18. sF2:0.66
    19. Braintri0.94
    20. AlEngineer0.98
    21. The Unreasonable Effectiveness of Separating the Task from the Model1.00
    22. UIT0.59
    23. Id'sFi0.89
    24. Maxime Rivest0.99
    25. IsaacMiller0.96
    26. DSPy1.00
    27. Core Contributor1.00
    28. Lead Maintainer1.00

  • 5:00 #18 done22 line(s)

    shot 18·sharpness 2459.1

    1. AlEngineer0.99
    2. World's Fair0.99
    3. PRESENTED BY0.97
    4. A signature alone doesn't1.00
    5. Microsoft1.00
    6. completely specify the task.0.98
    7. d'sFair1.00
    8. mazon A1.00
    9. AlEnginer0.92
    10. aintru0.99
    11. World's1.00
    12. Engineer0.99
    13. Yoget0.87
    14. The Unreasonable Effectiveness of Separating the Task from the Model1.00
    15. d'sF0.86
    16. Maxime Rivest0.99
    17. Isaac Miller1.00
    18. DSPy0.99
    19. AlEnginer0.97
    20. World'e0.88
    21. Core Contributor1.00
    22. Lead Maintainer0.97

  • 5:21 #19 done32 line(s)

    shot 19·sharpness 2744.4

    1. AlEngineer1.00
    2. World'sFair1.00
    3. SPECS1.00
    4. QUESTION 10.98
    5. import dspy1.00
    6. PRESENTED BY0.98
    7. What1.00
    8. lm = dspy.LM("openai/gpt-4.1-mini")1.00
    9. dspy.configure(lm=lm)1.00
    10. Microsoft1.00
    11. should1.00
    12. class ExtractTaxes(dspy.Signature):0.99
    13. """Extract all tax lines from an invoice.0.99
    14. happen?1.00
    15. Use 0.0 for illegible amounts."""1.00
    16. invoice: str = dspy.InputField()0.97
    17. taxes: dict[str, float] = dspy.OutputField()0.99
    18. orld's Fai0.97
    19. All1.00
    20. Brows1.00
    21. Worl0.89
    22. - Al0.64
    23. The Unreasonable Effectiveness of Separating the Task from the Model0.99
    24. Crus1.00
    25. orl0.99
    26. Maxime Rivest0.99
    27. IsaacMiller1.00
    28. DSPy0.99
    29. Ido0.67
    30. All1.00
    31. Core Contributor1.00
    32. Lead Maintainer1.00

  • 6:02 #20 done32 line(s)

    shot 20·sharpness 2744.8

    1. AlEngineer0.98
    2. World'sFair1.00
    3. SPECS1.00
    4. QUESTION 10.96
    5. import dspy1.00
    6. PRESENTED BY0.97
    7. What1.00
    8. lm = dspy.LM("openai/gpt-4.1-mini")1.00
    9. dspy.configure(lm=lm)1.00
    10. Microsoft1.00
    11. should1.00
    12. class ExtractTaxes(dspy.Signature):0.99
    13. """Extract all tax lines from an invoice.0.99
    14. happen?1.00
    15. Use 0.0 for illegible amounts."""1.00
    16. invoice: str = dspy.InputField()0.97
    17. taxes: dict[str, float] = dspy.OutputField()0.99
    18. AlEng0.72
    19. MINIMAX1.00
    20. World0.96
    21. AIEngineer0.96
    22. Ope0.99
    23. rld's Fair0.99
    24. Eng1.00
    25. The Unreasonable Effectiveness of Separating the Task from the Model1.00
    26. JSOe0.73
    27. Maxime Rivest1.00
    28. IsaacMiller0.95
    29. DSPy0.99
    30. AIEngineer0.95
    31. Core Contributor1.00
    32. Lead Maintainer1.00

  • 6:31 #21 done30 line(s)

    shot 21·sharpness 2935.9

    1. AlEngineer1.00
    2. SPECS + CODE0.96
    3. World'sFair1.00
    4. import dspy1.00
    5. lm = dspy.LM("openai/gpt-4.1-mini")0.99
    6. dspy.configure(lm=lm)1.00
    7. class ExtractTaxes(dspy.Signature):1.00
    8. QUESTION 20.99
    9. """Extract all tax lines from an invoice.0.99
    10. Use 0.0 for illegible amounts."""0.96
    11. PRESENTED BY0.99
    12. What1.00
    13. invoice: str = dspy.InputField()1.00
    14. taxes: dict[str, float] = dspy.OutputField()0.98
    15. Microsoft1.00
    16. must1.00
    17. class TaxPipeline(dspy.Module):1.00
    18. def __init__(self):0.99
    19. super().__init__()0.97
    20. happen?1.00
    21. self.extract = dspy.Predict(ExtractTaxes)0.99
    22. self.recheck = dspy.ChainOfThought(ExtractTaxes)0.99
    23. def forward(self, invoice: str):0.98
    24. pred = self.extract(invoice=invoice)0.99
    25. if not pred.taxes:1.00
    26. # routing0.99
    27. pred = self.recheck(invoice=invoice)1.00
    28. soft1.00
    29. World's Fai0.99
    30. if any(a < 0 for a in pred.taxes.values()):0.99

  • 7:07 #22 skipped

    shot 22·duplicate of #21

  • 7:35 #23 done

    shot 23·sharpness 2522.9

The page's on-screen-text budget of 600 lines is spent, so the last cards in this grid list fewer lines than they hold. Narrow the page with ?frames= to read them.

Transcript

187 cues· 2,727 words· 14,639 chars

  1. 0:12 Please welcome to the stage, Maxime Rivest and Isaac Miller.
  2. 0:19 Wow.
  3. 0:20 Isaac, myself, all of the DSPy community are so grateful to be here today to get to talk to you about AI programming, DSPy, and the unreasonable effectiveness of separating the task from the model, its harness, and all of the implementation details.
  4. 0:38 When you think about it,
  5. 0:40 In programming, if we want to repeat a task often, we make it a function.
  6. 0:46 We believe the same should be true for AI programs.
  7. 0:51 Functions are awesome.
  8. 0:52 Functions are reusable, composable, testable, and optimizable.
  9. 0:58 To make a function, you give it a name, you define some inputs, some outputs, and then you have some implementation logic inside of it.
  10. 1:08 You get to reuse your functions thousands of times.
  11. 1:11 You can optimize it, but you can also compose it into bigger programs.
  12. 1:17 One of the really nice things about functions is that you can also package it and distribute it, and someone else can use it, and they just need to know about the contract on top of it to use it, and they can treat it as a black box.
  13. 1:30 DSPy brings all of these properties to AR programs.
  14. 1:34 And so DSPy is an open source software in Python that lets you, like I said, bring these properties to your AI workflows and AI programs.
  15. 1:47 And it gives you all of the toolings you need to do that.
  16. 1:52 Why do you want that?
  17. 1:54 Well, we have been inventing a lot of terms in our fields in the last three years.
  18. 1:59 It's growing fast.
  19. 2:00 We have new models coming every other week.
  20. 2:04 We have new techniques, new strategies.
  21. 2:06 And if you're like me, you want to try all of them.
  22. 2:09 But will any of these new specific techniques coming out at a different time really help on your task, on your job?
  23. 2:19 Well, these are all just implementation tactics.
  24. 2:23 and you want to put them inside of clear contract.
  25. 2:27 If for your repeated AI task, you define an input interface and an output interface, you get to play in the internals.
  26. 2:35 You get a lot of agility.
  27. 2:37 Let's make it concrete for AI.
  28. 2:40 So my first AI program I made when I discovered DSPy was that I had some invoices from my farm and I wanted to extract them to do my taxes.
  29. 2:49 I wanted to extract the tax values from there.
  30. 2:52 Then, another AI program I did is that on my keyboard in my computer, I have a little command that reads my keyboard shortcuts, read my clipboard, and will correct the grammar for me.
  31. 3:06 Sometimes, I actually want it to also rewrite for clarity, so I have another program that takes text and just rewrites it for clarity, put it back in my keyboard, and that's a command, and then I can have a lot of agility and bring it different places.
  32. 3:19 Inside of that, I can change it however I want.
  33. 3:22 A new model comes up, and I can change that.
  34. 3:23 It's super easy, because my interface is fixed like that.
  35. 3:27 I'll skip that one.
  36. 3:29 But they're not restrained to very easy things and small input-output.
  37. 3:35 You can be very ambitious with AI programs.
  38. 3:38 So in this example, you could have your entire inbox and a new email coming in, and you want to compose a new drafted reply.
  39. 3:46 We can do that in DSPy with RLM, recursive language models.
  40. 3:50 This is an idea that came from around our community.
  41. 3:53 Or, more like things we probably all do, agentic engineering or vibe coding, you can give it a spec, a repository, and you get a PR.
  42. 4:02 Those are repeatable tasks.
  43. 4:06 And so, as I have been telling you, when you fix that boundary,
  44. 4:10 You can focus on the how on the top and then inside of it.
  45. 4:13 You can have a little chat with just a simple prompt.
  46. 4:17 You can iterate on that prompt.
  47. 4:18 Agents come out, you change it to be an agent.
  48. 4:20 Tools gets invented, you add tools.
  49. 4:23 And then we get into loop engineering, you put that inside of it too.
  50. 4:28 Anything on the outside of it doesn't change.

Open at this second