AIAcademy update archive

Update archive and lesson diffs

Historical cohorts stay visible so current AI guidance can change without hiding what learners saw before.

Cohorts

• 2026 Q2 frontier refresh
  The first public AIAcademy update cohort tracks frontier-lab source refreshes, product-created learning surfaces, and the lessons that need review before the next quarterly refresh.
• 2025 Q4 foundation snapshot
  The archived 2025 Q4 snapshot preserves the first frontier-model, AI-workflow, and household-safety guidance that later cohorts compare against instead of silently replacing.
• 2024 Q4 assistant baseline
  The archived 2024 Q4 baseline preserves the first multimodal, instruction-following, safety, and computer-use guidance that later cohorts revise as AI assistants become routine household, workplace, and builder infrastructure.
• 2023 Q4 foundation-model baseline
  The archived 2023 Q4 baseline preserves the first AIAcademy framing for GPT-3 to GPT-4 capability demos, broad benchmark literacy, tool-use research, early coding-agent evaluation, and workplace Copilot evidence before later cohorts move those topics into system-card, workflow, and governance practice.
• 2022 Q4 pre-ChatGPT foundation stack
  The archived 2022 Q4 baseline preserves the pre-ChatGPT foundation stack: transformer architecture, GPT-3 few-shot prompting, instruction tuning, compute-optimal scaling, ReAct-style tool reasoning, and early multimodal research before consumer assistants made those ideas visible to everyday learners.
• 2021 Q4 research-frontier bridge
  The archived 2021 Q4 bridge preserves the research frontier just before the 2022 assistant stack: CLIP and DALL-E text-image learning, Codex and HumanEval code generation, WebGPT browsing with citations and human feedback, and DeepMind Gopher/RETRO scale-plus-retrieval evidence.
• 2020 Q4 few-shot foundation roots
  The archived 2020 Q4 baseline preserves the pre-assistant foundation-model roots: transformer architecture as the reusable substrate, GPT-3 few-shot prompting as the first widely visible general model pattern, and MMLU-style benchmark literacy before instruction tuning, chat products, and system cards existed.
• 2019 Q4 generation-control baseline
  The archived 2019 Q4 baseline preserves early evidence that language-model behavior depends on decoding choices, before production builders had mature observability, structured-output, and evaluation routines for everyday AI products.
• 2018 Q4 task-impact and teach-back bridge
  The archived 2018 Q4 bridge preserves two pre-assistant foundations for AIAcademy: task-level machine-learning impact for work and learning-by-teaching evidence for study patterns before chat assistants became household learning partners.
• 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.

Lesson version diffs

• 6.5.1.1 · Reading frontier system cards without overclaiming
  4 cohort versions · needs-review
• 3.2.2.1 · Designing AI-supported knowledge-work workflows
  3 cohort versions · current
• 1.1.3.3 · Instruction tuning as the adoption bridge
  5 cohort versions · current
• 1.2.2.1 · Stage 1: pretraining, the read-the-internet step
  5 cohort versions · current
• 4.6.2.1 · RLHF -> RLAIF -> Constitutional AI — the lineage
  3 cohort versions · current
• 2.2.1.1 · Talking to kids about everyday AI use
  2 cohort versions · current
• 2.4.3.1 · Personal AI and health-adjacent boundaries
  2 cohort versions · needs-review
• 2.5.1.1 · Generated images as everyday media
  2 cohort versions · current
• 2.5.2.1 · Generated audio and video identity checks
  3 cohort versions · scheduled
• 6.4.1.1 · Veo, Kling, Runway: what video generation can do now
  4 cohort versions · needs-review
• 4.5.2.1 · Computer-use agents and irreversible actions
  2 cohort versions · current
• 4.2.2.1 · Tool calling
  3 cohort versions · current
• 4.1.2.1 · Context engineering before long-context products
  2 cohort versions · current
• 4.4.2.1 · Code-generation evals from HumanEval to agents
  2 cohort versions · current
• 5.7.2.1 · Prompt-injection boundaries for agents
  2 cohort versions · needs-review
• 7.1.1.1 · Compute and hardware literacy
  4 cohort versions · current
• 2.3.1.1 · The "feels like studying" trap
  2 cohort versions · current
• 4.4.4.4 · Temperature is a reliability knob
  2 cohort versions · current
• 5.5.2.2 · Brynjolfsson's "augmentation, not automation"
  2 cohort versions · current