Semantic programming by example with pre-trained models
summary
A system combines FlashFill-style deductive synthesis with GPT-3 to handle semantic string ops by adding three semantic operators (e.g. SemMap, SemPos) and a deferred query execution algorithm plus clustering. It treats GPT-3 as an oracle, avoids thousands of expensive (up to 10s) calls during search, and keeps only similar query outputs. Evaluation shows GPT-3 alone needs many more examples, FlashGPT3 cuts examples in many semantic benchmarks (typically syntactic part done after two examples) and scaling problems are mitigated.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:48What is GPD3
- 2:30How Flashfield works
- 5:34Semantic operators
- 7:47Semantic position operator
- 9:47Ranking
- 11:07Challenges
- 14:18Future work
related talks
22mSemantic vs keyword search as context for GPT
18mData Extraction via Semantic Regular Expression Synthesis
59mOrchestrate your AI with Semantic Kernel
23mPanel
01h 01mC++hatGPT: How AI Tools Impact Your Life as a Systems Programmer
45mHow to use GPT in your projects: A Case Study
28mTalking to Objects in Natural Language: Toward Semantic Tools for Exploratory(…)
51mLarge Language Models for Real-World Applications - A Gentle Intro