AIAcademy update cohort · archived

2017 Q2 transformer and feedback roots

The archived 2017 Q2 root snapshot preserves the two research ideas that later made modern assistants teachable: the transformer architecture as a reusable attention-based substrate and human-preference reinforcement learning as an early behavior-shaping recipe.

2 updates · 2 source-linked · 0 product-created

Key questions

• Which 2017 research claims became the architecture and alignment roots for later foundation-model lessons?
• Which ideas were still narrow research mechanisms rather than consumer assistant guidance?
• Which later lessons should keep 2017 as historical root context instead of treating it as current practice?

Updates

• Transformer architecture root archived
  The transformer paper remains visible as the 2017 root of modern general-purpose model architecture, before GPT-scale pretraining, few-shot prompting, retrieval, instruction tuning, and multimodal assistant products changed the curriculum.
• Human-preference learning root archived
  The original OpenAI/DeepMind human-preference reinforcement-learning paper stays archived as the behavior-shaping root behind later RLHF, InstructGPT, RLAIF, Constitutional AI, and DPO lessons.