GPT-3 in 2020 didn't just improve language models — it fundamentally changed what data scientists do, shifting from training custom models to orchestrating foundation models.
GPT-3 (2020) demonstrated a single foundation model with competence across an astonishingly broad range — rewriting how data science is practised.
Traditional ML demanded thousands of labelled examples per task. Foundation models arrive pre-trained on vast corpora — the data scientist's role shifts from labelling to prompting.
MLOps focused on training pipelines and feature stores. GenAI applications need an entirely new operational discipline — LLMOps.
Accuracy, precision, and recall gave clear objective scores. Generative outputs resist simple metrics — evaluation becomes application-specific.
RAG became the dominant pattern for enterprise AI — combining foundation model language capability with dynamic information retrieval.
Zero mentions in 2020 job postings — nearly 20% by 2024. GenAI created an entirely new skill category alongside existing data science roles.
Many organisations experiment with ChatGPT for simple tasks but struggle to integrate GenAI into core business processes. The gap between proof-of-concept demonstrations and production-grade systems remains substantial — and is where skilled practitioners create the most value.
Key Insight — Universal metrics of traditional ML give way to bespoke evaluation frameworks tailored to specific use cases — customer service prioritises accuracy, creative writing values originality, code generation demands correctness.