Building AI with AI

AI-powered coding assistants have transformed the way we build software – and they can be even more impactful for AI development itself. In this talk, I'll show you why we should use LLMs to build systems instead of as systems, and why code and the open-source ecosystem is more important than ever, not less.

Transcript

1. Ines Montani Explosion
2. Open-source library for industrial-strength natural language processing spacy.io 470m+ downloads
3. Open-source library for industrial-strength natural language processing spacy.io LLMs are really good at spaCy code! 470m+ downloads
4. 12k+ users 1000+ companies Modern scriptable annotation tool for machine learning developers prodigy.ai
5. 12k+ users 1000+ companies fully scriptable in Python Alex Smith Developer Kim Miller Analyst GPT-5 API Modern scriptable annotation tool for machine learning developers prodigy.ai
6. of coding assistants
7. help developer implement code for the given tools of coding assistants
8. help developer implement code for the given tools of coding assistants Jay Alammar: PyData London Keynote
9. help developer implement code for the given tools of coding assistants Jay Alammar: PyData London Keynote
10. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants
11. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants solve a business problem
12. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants solve a business problem “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” 2025.pdf 2024.pdf 2023.pdf
13. reverse these strings
14. reverse these strings write a script to reverse strings
15. reverse these strings write a script to reverse strings prompt program
17. reproducible
18. scalable reproducible
19. scalable maintainable reproducible
20. scalable maintainable reproducible extensible
21. scalable maintainable faster reproducible extensible
22. scalable maintainable cheaper faster reproducible extensible
23. scalable maintainable cheaper faster reproducible extensible OSS ecosystem
24. scalable maintainable cheaper faster reproducible extensible OSS ecosystem coding assistants
25. scalable maintainable cheaper faster reproducible extensible OSS ecosystem coding assistants experience
26. as the system
27. as the system to build
28. list all company names in the text write a script to extract company names from text
29. natural language structured data vs.
30. natural language structured data vs. consumed by humans
31. natural language structured data vs. consumed by humans consumed by machines
32. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.”
33. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs
34. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses
35. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type
36. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math
37. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table
38. natural language structured data vs. consumed by humans consumed by machines “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table
39. natural language structured data vs. consumed by humans consumed by machines Most industry applications of NLP are part of a larger system. “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table
40. natural language structured data vs. consumed by humans consumed by machines Most industry applications of NLP are part of a larger system. “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table “The results will then be added to our internal database so we can predict future spending.”
41. natural language structured data vs. consumed by humans consumed by machines Most industry applications of NLP are part of a larger system. “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table populate database “The results will then be added to our internal database so we can predict future spending.”
42. natural language structured data vs. consumed by humans consumed by machines Most industry applications of NLP are part of a larger system. “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” parse PDFs extract expenses classify expense type do math create table populate database “The results will then be added to our internal database so we can predict future spending.” model predictions
43. At their core, many NLP systems consist of flat classifications. You can shove them into a single prompt, or you can decompose them into smaller pieces. Many classification tasks are straightforward to solve nowadays – but they become vastly more complicated if one model needs to do them all at once. explosion.ai/blog/human-in-the-loop-distillation
44. explosion.ai/blog/human-in-the-loop-distillation LLM Human-in-the-loop
45. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM Human-in-the-loop
46. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting Human-in-the-loop
47. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting Human-in-the-loop
48. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning Human-in-the-loop
49. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning distilled model Human-in-the-loop
50. explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning distilled model Human-in-the-loop production
51. sort these documents into custom categories create data and train a classifier for custom categories
53. Pareto Frontier for AI models Cost Accuracy
54. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation!
55. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation! runtime → development use LLMs to create runtime system
56. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation! write code runtime → development use LLMs to create runtime system
57. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation! write code create data runtime → development use LLMs to create runtime system
58. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation! write code create data train classifiers runtime → development use LLMs to create runtime system
59. Pareto Frontier for AI models Cost Accuracy LLMs as developer tools change the calculation! write code create data train classifiers strategize runtime → development use LLMs to create runtime system
60. Use LLMs to build the system, not as the system.
61. Use LLMs to build the system, not as the system. There’s no need to compromise on development best practices or privacy.
62. Use LLMs to build the system, not as the system. There’s no need to compromise on development best practices or privacy. Code is more important than ever – not less!
63. Use LLMs to build the system, not as the system. There’s no need to compromise on development best practices or privacy. Code is more important than ever – not less!* * This includes the open-source ecosystem!
64. Explosion spaCy Prodigy Bluesky Mastodon explosion.ai spacy.io prodigy.ai @inesmontani.bsky.social @ines@sigmoid.social LinkedIn