10 Years of Open Source: Navigating the Next AI Revolution

A lot has been happening in the field of AI and Natural Language Processing: there's endless excitement about new technologies, sobering post-hype hangovers and also uncertainty about where the field is heading next. In this talk, I'll share the most important lessons we've learned in 10 years of working on open-source software, our core philosophies that helped us adapt to an ever-changing AI landscape and why open source and interoperability still wins over black-box, proprietary APIs.

Transcript

1. Ines Montani Explosion
2. Open-source library for industrial-strength natural language processing spacy.io 255m+ downloads
3. Open-source library for industrial-strength natural language processing spacy.io ChatGPT can write spaCy code! 255m+ downloads
4. Modern scriptable annotation tool for machine learning developers prodigy.ai 900+ companies 10k+ users
5. Alex Smith Developer Kim Miller Analyst GPT-4 API Modern scriptable annotation tool for machine learning developers prodigy.ai 900+ companies 10k+ users
7. ^ first commit to spaCy
8. ^ first commit to spaCy spaCy is first released spacy.io
9. OUR DEVELOPMENT PHILOSOPHY OUR DEVELOPMENT PHILOSOPHY “Let T h e m W rite Code” spacy.fyi/ltwc
10. OUR DEVELOPMENT PHILOSOPHY OUR DEVELOPMENT PHILOSOPHY “Let T h e m W rite Code” spacy.fyi/ltwc Good tools help people do their work. You don’t have to do their work for them.
11. OUR DEVELOPMENT PHILOSOPHY OUR DEVELOPMENT PHILOSOPHY “Let T h e m W rite Code” spacy.fyi/ltwc Good tools help people do their work. You don’t have to do their work for them. ["go", "swim"]
12. OUR DEVELOPMENT PHILOSOPHY OUR DEVELOPMENT PHILOSOPHY “Let T h e m W rite Code” spacy.fyi/ltwc Good tools help people do their work. You don’t have to do their work for them. ["go", "swim"] spaCy
13. OUR DEVELOPMENT PHILOSOPHY OUR DEVELOPMENT PHILOSOPHY “Let T h e m W rite Code” spacy.fyi/ltwc Good tools help people do their work. You don’t have to do their work for them. You can reinvent the wheel, but don’t try to reinvent the road. ["go", "swim"] spaCy
14. ^ first commit to spaCy spaCy is first released spacy.io
15. ^ first commit to spaCy spaCy is first released spacy.io everyone gets excited about chat bots
17. “knocker-uppers”
18. The Window K nocking Machine Tes t ines.io/blog/window-knocking-machine-test “knocker-uppers”
19. The Window K nocking Machine Tes t ines.io/blog/window-knocking-machine-test Are you designing a window-knocking machine or an alarm clock? “knocker-uppers”
20. Hello, I ’ m Toni ’ s virtual assistant and I help schedule appointments. Do you have time at 1pm on Monday? No, but Tuesday would work for me. Okay, please confirm: Tuesday at 1pm? 1pm is unideal but 3pm would work. Toni doesn ’ t have availability at 3pm but I could offer a slot at 4pm or 5 : 30pm. Which time zone is this by the way? I ’ m in CET.
21. Hello, I ’ m Toni ’ s virtual assistant and I help schedule appointments. Do you have time at 1pm on Monday? No, but Tuesday would work for me. Okay, please confirm: Tuesday at 1pm? 1pm is unideal but 3pm would work. Toni doesn ’ t have availability at 3pm but I could offer a slot at 4pm or 5 : 30pm. Which time zone is this by the way? I ’ m in CET. Calendly
22. Hello, I ’ m Toni ’ s virtual assistant and I help schedule appointments. Do you have time at 1pm on Monday? No, but Tuesday would work for me. Okay, please confirm: Tuesday at 1pm? 1pm is unideal but 3pm would work. Toni doesn ’ t have availability at 3pm but I could offer a slot at 4pm or 5 : 30pm. Which time zone is this by the way? I ’ m in CET. Calendly “window-knocking machine” “alarm clock”
23. ^ first commit to spaCy spaCy is first released spacy.io everyone gets excited about chat bots
24. ^ first commit to spaCy spaCy is first released spacy.io deep learning is widely adopted everyone gets excited about chat bots
25. Software 1.0 Software 1.0 📄 code 💾 program compiler
26. Software 1.0 Software 1.0 📄 code 💾 program compiler Software 2.0 Software 2.0 📊 data 🔮 model algorithm
27. Software 1.0 Software 1.0 📄 code 💾 program compiler Software 2.0 Software 2.0 📊 data 🔮 model algorithm ✅ tests 📈 evaluation
28. Software 1.0 Software 1.0 📄 code 💾 program compiler Software 2.0 Software 2.0 📊 data 🔮 model algorithm ✅ tests 📈 evaluation refactoring refactoring
29. Software 1.0 Software 1.0 📄 code 💾 program compiler Software 2.0 Software 2.0 📊 data 🔮 model algorithm ✅ tests 📈 evaluation refactoring refactoring iteration iteration
30. Prodigy is first released prodigy.ai
31. language model pre-training works ^ ^ Prodigy is first released prodigy.ai
32. language model pre-training works ^ ^ Prodigy is first released prodigy.ai few-shot in-context learning works ^ ^
33. language model pre-training works ^ ^ Prodigy is first released prodigy.ai few-shot in-context learning works ^ ^
34. spaCy v3 is first released
35. i U se cases i n industr y generative tasks 📖 single/multi-doc summarization 🧮 reasoning ✅ problem solving ✍ paraphrasing 🖼 style transfer ⁉ question answering predictive tasks 🔖 entity recognition 🔗 relation extraction 👫 coreference resolution 🧬 grammar & morphology 🎯 semantic parsing 💬 discourse structure 📚 text classification
36. i U se cases i n industr y generative tasks 📖 single/multi-doc summarization 🧮 reasoning ✅ problem solving ✍ paraphrasing 🖼 style transfer ⁉ question answering predictive tasks 🔖 entity recognition 🔗 relation extraction 👫 coreference resolution 🧬 grammar & morphology 🎯 semantic parsing 💬 discourse structure 📚 text classification structured data many industry problems have remained the same, they just changed in scale
37. spaCy v3 is first released
38. spaCy v3 is first released in-context learning gains traction
39. human-facing systems machine-facing models ChatGPT GPT-4 A I products are m ore t h an jus t a model
40. human-facing systems machine-facing models ChatGPT GPT-4 most important di erentiation is product, not just technology A I products are m ore t h an jus t a model
41. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization most important di erentiation is product, not just technology A I products are m ore t h an jus t a model
42. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization most important di erentiation is product, not just technology swappable components based on research, impacts are quantifiable A I products are m ore t h an jus t a model
43. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization speed accuracy latency cost most important di erentiation is product, not just technology swappable components based on research, impacts are quantifiable A I products are m ore t h an jus t a model
44. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization speed accuracy latency cost But what about the data? most important di erentiation is product, not just technology swappable components based on research, impacts are quantifiable A I products are m ore t h an jus t a model
45. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization speed accuracy latency cost But what about the data? User data is an advantage for product, not the foundation for machine-facing tasks. most important di erentiation is product, not just technology swappable components based on research, impacts are quantifiable A I products are m ore t h an jus t a model
46. human-facing systems machine-facing models ChatGPT GPT-4 UI / UX marketing customization speed accuracy latency cost But what about the data? User data is an advantage for product, not the foundation for machine-facing tasks. You don’t need specific data to gain general knowledge. most important di erentiation is product, not just technology swappable components based on research, impacts are quantifiable A I products are m ore t h an jus t a model
47. spaCy v3 is first released in-context learning gains traction
48. spacy-llm is first released github.com/explosion/spacy-llm spaCy v3 is first released in-context learning gains traction
49. task-specific output 💬 prompt 📖 text LLM spacy.io/usage/large-language-models spac y -llm
50. task-specific output 💬 prompt 📖 text LLM prompt model & transform output to structured data spacy.io/usage/large-language-models spac y -llm
51. task-specific output 💬 prompt 📖 text LLM prompt model & transform output to structured data spacy.io/usage/large-language-models spac y -llm config.cfg Structured Data {} LLM Text
52. task-specific output 💬 prompt 📖 text LLM prompt model & transform output to structured data spacy.io/usage/large-language-models unified, model-agnostic API spac y -llm config.cfg Structured Data {} LLM Text
53. task-specific output 💬 prompt 📖 text LLM prompt model & transform output to structured data spacy.io/usage/large-language-models unified, model-agnostic API spac y -llm config.cfg Structured Data {} LLM Text entity recognition entity linking text classification relation extraction and more…
54. spacy-llm is first released github.com/explosion/spacy-llm spaCy v3 is first released in-context learning gains traction
55. spacy-llm is first released github.com/explosion/spacy-llm spaCy v3 is first released in-context learning gains traction LLMs and Generative AI fully hit the mainstream ChatGPT ⏺ ⏺ ⏺
56. E cono m ies of scale of scale output costs
57. E cono m ies of scale of scale output costs OpenAI Google
58. E cono m ies of scale of scale output costs OpenAI Google access to talent, compute etc.
59. E cono m ies of scale of scale output costs OpenAI Google access to talent, compute etc. API request batching
60. E cono m ies of scale of scale output costs OpenAI Google high tra ff ic 💧 💧 💧 💧 💧 💧 💧 💧 low tra ff ic batch 💧 💧 💧 💧 💧 💧 💧 💧 … access to talent, compute etc. API request batching
61. E cono m ies of scale of scale output costs OpenAI Google you 🤠 high tra ff ic 💧 💧 💧 💧 💧 💧 💧 💧 low tra ff ic batch 💧 💧 💧 💧 💧 💧 💧 💧 … access to talent, compute etc. API request batching
62. human-in-the-loop distillation is promising prodigy.fyi/distillation
63. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation LLM
64. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM
65. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting
66. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting
67. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning CO M PO N EN T
68. in the loop H uma n explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning CO M PO N EN T distilled model
69. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
70. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes • high-security environment 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
71. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes • high-security environment • used LLM during annotation 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
72. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes • high-security environment • used LLM during annotation • 10× data development speedup with humans and model in the loop 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
73. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes • high-security environment • used LLM during annotation • 10× data development speedup with humans and model in the loop • 8 market pipelines in production 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
74. 99% 99% Case Stud y : S&P Global • real-time commodities trading insights by extracting structured attributes • high-security environment • used LLM during annotation • 10× data development speedup with humans and model in the loop • 8 market pipelines in production 6mb 6mb model size 16k+ 16k+ words/second F-score explosion.ai/blog/sp-global-commodities
75. human-in-the-loop distillation is promising prodigy.fyi/distillation
76. human-in-the-loop distillation is promising prodigy.fyi/distillation everyone is excited about chat bots again
77. ? ines.io/blog/window-knocking-machine-test
78. What ’ s the total services revenue from 2023? $2,923,531 How many clients is that in total? 29 ⏺ ⏺ ⏺ ? ines.io/blog/window-knocking-machine-test
79. What ’ s the total services revenue from 2023? $2,923,531 How many clients is that in total? 29 ⏺ ⏺ ⏺ 🔮 LLM 📚 database 🤖 agents ⚙ query Retrieval-Augmented Generation ? ines.io/blog/window-knocking-machine-test
80. 2023 Year Services Type ACME Inc. FooBar GmbH NLPCorp XKCD Ltd. Python AG 432,032 82,000 1,500 193,000 91,320 $ 2,625,032 Clients (28) Revenue What ’ s the total services revenue from 2023? $2,923,531 How many clients is that in total? 29 ⏺ ⏺ ⏺ 🔮 LLM 📚 database 🤖 agents ⚙ query Retrieval-Augmented Generation ? ines.io/blog/window-knocking-machine-test
81. 2023 Year Services Type ACME Inc. FooBar GmbH NLPCorp XKCD Ltd. Python AG 432,032 82,000 1,500 193,000 91,320 $ 2,625,032 Clients (28) Revenue A I still needs produc t decisions! Kim Miller Analyst What ’ s the total services revenue from 2023? $2,923,531 How many clients is that in total? 29 ⏺ ⏺ ⏺ 🔮 LLM 📚 database 🤖 agents ⚙ query Retrieval-Augmented Generation ? ines.io/blog/window-knocking-machine-test
82. human-in-the-loop distillation is promising prodigy.fyi/distillation everyone is excited about chat bots again
83. Explosion goes back to independent-minded and self-su ff icient explosion.ai/blog/ back-to-our-roots human-in-the-loop distillation is promising prodigy.fyi/distillation everyone is excited about chat bots again
84. Explosion goes back to independent-minded and self-su ff icient explosion.ai/blog/ back-to-our-roots human-in-the-loop distillation is promising prodigy.fyi/distillation everyone is excited about chat bots again What’s next?
85. Cycle A doptio n
86. Cycle A doptio n rules and conditional logic
87. Cycle A doptio n rules and conditional logic applied workflow
88. Cycle A doptio n rules and conditional logic linear models applied workflow
89. Cycle A doptio n rules and conditional logic linear models applied workflow applied workflow
90. Cycle A doptio n rules and conditional logic linear models applied workflow applied workflow combine new techniques with established workflows
91. Cycle A doptio n rules and conditional logic deep learning linear models applied workflow applied workflow combine new techniques with established workflows
92. Cycle A doptio n rules and conditional logic deep learning linear models chat bots applied workflow applied workflow combine new techniques with established workflows
93. Cycle A doptio n rules and conditional logic deep learning linear models chat bots applied workflow applied workflow applied workflow combine new techniques with established workflows
94. Cycle A doptio n rules and conditional logic deep learning transfer learning linear models chat bots applied workflow applied workflow applied workflow combine new techniques with established workflows
95. Cycle A doptio n rules and conditional logic deep learning transfer learning linear models chat bots trans- formers applied workflow applied workflow applied workflow combine new techniques with established workflows
96. Cycle A doptio n rules and conditional logic deep learning transfer learning linear models chat bots trans- formers applied workflow applied workflow applied workflow applied workflow combine new techniques with established workflows
97. Cycle A doptio n rules and conditional logic deep learning transfer learning in-context learning linear models chat bots trans- formers applied workflow applied workflow applied workflow applied workflow combine new techniques with established workflows
98. Cycle A doptio n rules and conditional logic deep learning transfer learning in-context learning linear models chat bots LLMs and GenAI trans- formers applied workflow applied workflow applied workflow applied workflow combine new techniques with established workflows
99. Cycle A doptio n rules and conditional logic deep learning transfer learning in-context learning linear models chat bots LLMs and GenAI trans- formers applied workflow applied workflow applied workflow applied workflow applied workflow combine new techniques with established workflows
100. Summar y NAVIGATING AI & NLP NAVIGATING AI & NLP
101. Summar y NAVIGATING AI & NLP NAVIGATING AI & NLP Think beyond chat bots or human-shaped tasks. You don’t want to build a “window-knocking machine”.
102. Summar y NAVIGATING AI & NLP NAVIGATING AI & NLP Think beyond chat bots or human-shaped tasks. You don’t want to build a “window-knocking machine”. Structured Data {} Focus on your application. Consider what it really needs and let your data guide you.
103. Summar y NAVIGATING AI & NLP NAVIGATING AI & NLP Think beyond chat bots or human-shaped tasks. You don’t want to build a “window-knocking machine”. Stay ambitious. Don’t compromise on best practices, e iciency and privacy. Structured Data {} Focus on your application. Consider what it really needs and let your data guide you.
104. Summar y NAVIGATING AI & NLP NAVIGATING AI & NLP Think beyond chat bots or human-shaped tasks. You don’t want to build a “window-knocking machine”. Stay ambitious. Don’t compromise on best practices, e iciency and privacy. LLM Keep filling up your toolbox. Know the techniques you have available and apply the best ones to get the job done. Structured Data {} Focus on your application. Consider what it really needs and let your data guide you.
105. Explosion spaCy Prodigy Twitter Mastodon Bluesky explosion.ai spacy.io prodigy.ai @_inesmontani @ines@sigmoid.social @inesmontani.bsky.social LinkedIn