Even Lighter Than Lightweiht: Augmenting Type Inference with Primitive Heuristics and AI-support
- 1. Even Li hter Than Li htwei ht Augmenting Type Inference with Primitive Heuristics and AI-support Larisa Safina, Jan Blizničenko IWST 25 – Gdańsk, 2025
- 2. Type In erence? In Dynamic Lan ua es? Why? Improves correctness, maintainability, performance Hard in dynamic languages like Smalltalk - Polymorphism - Trade-off: precision vs. speed - Computational resources 2
- 3. And no one did it be ore? 3
- 4. A bit o history… 4 Impressive timeline (from 1981 by Suzuki) From less to more formal Compiler-oriented vs IDE-oriented Pluggable Types Gradual Typing Heuristics …
- 5. What about the present 5 - IDE: RoelTyper - IDE: JInferer - TypeInfoTools - VM-level type feedback - … However, IDE-integrated type inference is not widely used by Pharo developers on a daily basis
- 6. Why this insufficient adoption? - Problems of early solutions - Human factor - end of funding - lost of interest - phd/postdoc ends - Not user-friendly 6
- 7. So what do we want ? And how ? - Balance precision and speed - Make inference interactive and IDE-friendly Step 1: Lightweight heuristics to pre-calculate types and speed up stronger tools Simple, fast, surprisingly effective 7
- 8. Goin back to heuristics Heuristics on method titles (127k methods) - Implicit self returns (~57k methods) - Type by name (~7,8k methods) - Single return methods (~10k methods) - Collection-specific methods - Test methods (~22,5k methods) 8
- 9. Goin back to heuristics Heuristics on method titles - Type by name (~7,8k methods) - Boolean (isEmpty, hasElements, includes, ..) - String (asString, …) - Numerical (size, …) 9
- 10. Goin back to heuristics Heuristics on method titles - Single return methods (~10k methods) - AST LiteralNode as String, Number (8745 methods) - class reference (982 methods) - class new (394 methods) - self - self new (34 methods) - nil (278 methods) 10
- 11. Goin back to heuristics Heuristics on method titles - Collection-specific methods - Numeric e.g indexOf - String e.g. concatenation - Boolean e.g. anySatisfy: - Transformation e.g asCollection - Collection-preserving e.g addAll:, copy 11
- 12. Goin back to heuristics Heuristics on method titles (127k methods) - Implicit self returns (~57k methods) - Type by name (~7,8k methods) - Single return methods (~10k methods) - Collection-specific methods - Test methods (~22,5k methods) 12
- 13. Evaluation Inference takes ~1 minute on Mac M2 with 24 GB of RAM In Pharo 13 image heuristics cover 97,3k methods of 127k Validation - Sanity Check - Comparison with TypeInfoTools 13
- 14. AI-Assisted In erence (Experimental) - Queries to OpenAI/Mistral for type prediction - Promising, but slow and sometimes wrong - Can complement heuristics - Need to learn lessons from Python 14
- 15. IDE Inte ration - Type results shown on demand - CSV-based or loaded into Pharo (negligible overhead) - AI-specific settings 15
- 16. Type results shown on demand 16
- 17. What's Next? Combine everything! (Heuristics, other TI tools, AI, CI, ADHD, …) - Interactive Type Refinement - Create a Smalltalk “typeshed” - Allow devs to refine suggestions - Improve AI part - … 17