Bigger models aren't necessarily more helpful. OpenAI used a three-step RLHF pipeline to make a 1.3B InstructGPT preferred over a 175B GPT-3. We unpack SFT, reward modeling, and PPO in detail.
T5's unified text-to-text framework swallows all NLP tasks; BART fuses BERT's bidirectional understanding with GPT's generation via denoising — the peak of the encoder-decoder line.
GPT-3's 175B parameters proved scale is capability: no fine-tuning, no gradient updates — just a few examples in the prompt beat then-SOTA fine-tuned models.
DeepSeek-R1 showed that pure RL (no SFT cold start) can make a base model reason — and that GRPO with rule-based rewards is enough to produce emergent chain-of-thought.
Source-level analysis of containerd gRPC, TTRPC, and UDS layers — how a container runtime manages process lifecycles.
