2017 to present — the architectures that reshaped machine intelligence
Self-attention replaces recurrence entirely. Enables massively parallel training on long sequences. Every production LLM today — GPT, Claude, Gemini — descends from this architecture.
foundationalMasked language modeling and next-sentence prediction. Bidirectional context captures richer semantics than left-to-right GPT. Transfer learning for NLP becomes standard practice.
encoderScale reveals in-context learning as an emergent property. Few-shot prompting replaces fine-tuning for many tasks. API-first deployment establishes the LLM-as-service model.
scaleReinforcement Learning from Human Feedback teaches models to follow instructions and refuse harmful requests. Consumer adoption goes mainstream: 100M users in 60 days.
alignmentVision inputs, 100K+ context windows, and bar-exam-level reasoning. Models begin passing professional licensing exams. Constitutional AI and process supervision improve safety.
multimodalStandardises tool and resource interfaces so agents can compose capabilities across vendors without custom glue code. Marks the shift from chatbots to autonomous agentic systems.
agentic