How modern AI happened, from the 2017 Transformer to 2026 reasoning agents.
Most explanations of modern AI are either a pop-science blur or a stack of disconnected papers. This is the path between: one continuous story for someone who knew CNNs and watched attention arrive in NLP around 2019, and wants the rest of it - RLHF, DPO, GRPO, prompt optimization, reasoning models, agents - as a single arc rather than a reading list.
The through-line is one question: where does the optimization happen - in the weights, or in the context? Eight chapters, each turning one lever, every claim cited to its primary source, with live interactive demos.
| # | Chapter | Era | The lever it turns |
|---|---|---|---|
| 0 | Where you left off | 2017–2019 | Architecture: the Transformer, BERT vs GPT |
| 1 | Scale is a strategy | 2020–2022 | Pretraining: GPT-3, scaling laws, Chinchilla, ViT/CLIP |
| 2 | Teaching it to be helpful | 2022 | Alignment: RLHF, ChatGPT, the chain-of-thought seed |
| 3 | The recipe gets cheaper | 2023–2024 | Alignment, cheaper: DPO, GRPO, LLaMA, LoRA/QLoRA |
| 4 | Steering without retraining | 2022–2025 | Context: APE, OPRO, DSPy, GEPA |
| 5 | Learning to think | 2024–2025 | Test-time compute: o1, DeepSeek-R1, CoT trained in |
| 6 | Learning to act | 2024–2026 | Agency: tool use, ReAct, coding agents |
| 7 | Computational cognition | 2026 | Synthesis, plus a voice-tool capstone |
What makes it a story instead of a timeline - motifs that recur and pay off across chapters:
- Chain-of-thought, three times: a prompt trick in Ch2, a prompt-optimization target in Ch4, then trained straight into the weights in Ch5.
- The cost curve: every advance gets cheaper and more open. RLHF → DPO → GRPO. Full fine-tune → LoRA → QLoRA. Closed → LLaMA → DeepSeek.
- The eval shadow: how we measure progress (BBH → MMLU → SWE-bench → LLM-as-judge), with Goodhart's law as the recurring villain.
- Multimodality underneath: it surfaces where it matters (ViT/CLIP early, native-multimodal models late) instead of taking its own chapter.
- Live demos, not pictures of demos: PPO vs DPO vs GRPO side by side, a propose→score→select prompt search, reflective evolution on a Pareto frontier, a pairwise→Elo leaderboard.
- A glossary sidebar: click any underlined term for a first-principles nugget plus links to the primary sources, without losing your scroll position.
- Cited, not vibes: every date and figure links to arXiv, ACL, or the original report.
- Read it either way: drop into a single chapter, or follow the prev/next arc end to end.
- Calm by design: light warm-paper, all-sans, low-stimulation. Color is rare, so it carries meaning.
Chapter 4, Steering without retraining: sidebar TOC, cited prose, and a compare-and-contrast table.
Plain HTML, CSS, and vanilla JS. No build step, no dependencies. KaTeX and Google Fonts load from a CDN, so file:// won't work - serve the folder:
python3 -m http.server 8000 # then open http://localhost:8000Architecture, design tokens, and how to add a chapter live in CLAUDE.md - the onboarding doc for humans and agents working on the site.
MIT. See LICENSE.