Sources
Sebastian Raschka's LLM Eras Framework: 2026 = Inference-Time Scaling Year — 2027 Predicted as Continual Learning
Raschka published a compact LLM training paradigm periodization: 2022 RLHF/PPO, 2023 LoRA SFT, 2024 mid-training, 2025 RLVR/GRPO, 2026 inference-time scaling (current), and 2027 continual learning (predicted). With his 168K-subscriber Ahead of AI newsletter as signal amplifier, this framework is already shaping how practitioners prioritize research direction. The implication: inference-time compute optimization — test-time search, chain-of-thought scaling, streaming compute — is the highest-ROI skill to develop now.
↳ Follow the thread