Skip to content

The papers that built generative AI, in plain language

1.4M words of research - Transformers to diffusion, RLHF to reasoning models - read in full and rewritten as 203,000 words. About 8x shorter, and written to be finished.

No maths degree assumed, no paywall, no signup. Press / to search all 225,000+ words.

Start the roadmap Browse all 107 papers Look up a term

  • 1.4M words in
  • 203,000 words out
  • 8x shorter
  • 107 papers
  • 13 years covered
  • $0 to read

Source figure measured, not estimated: every word pdftotext extracts from 92 of the 107 papers. The other 15 sit behind journal paywalls or are published as web pages with no PDF, so they are excluded - the real total is higher. See the script.

Four ways in

Pick your entry point

  • Follow a path

    A staged reading order that assumes nothing, from Word2Vec through to today's reasoning models. Start here if the field is new to you.

    Learning roadmap

  • Browse the library

    All 107 summaries grouped by category, or filtered down to the 57 topic tags - retrieval, alignment, efficiency, agents, scaling.

    Paper index

  • Compare approaches

    LoRA or full fine-tuning, RAG or long context, DPO or PPO. Decision guides for the choices you actually face when building.

    Comparisons

  • Get the gist fast

    One-line takeaways for every paper, plus which ones still matter in production and which are history you can safely skim.

    Quick reference

The library

Papers by category

Full index with titles, authors, and years

Quick links

Jump to what you need

What each summary gives you

Every paper here was read in full, then rewritten to a fixed shape so you can move between them without relearning the format each time:

  • Why this matters - the one thing that changed, stated before any notation.
  • The problem - what people were stuck on, and why the obvious fix did not work.
  • The core innovation - the actual idea, with diagrams instead of proofs.
  • Key results - the numbers the paper reported, not the numbers people remember.
  • Real-world impact - what shipped because of it.
  • Limitations - what it does not do, including where the field has since moved past it.
  • Related in this collection - generated cross-links, so following a thread never dead-ends.

Summaries are not a substitute for the paper. Every one links to the original, and the good ones are worth your time once the summary has told you what to look for.

Who made this

Built by Patrick Wiloak - ex-AWS Solutions Architect, 10 years in tech, 18x multi-cloud certified. YouTube · LinkedIn · Source on GitHub

Learning the cloud, data, and security side too? Cloud, Data, AI and Security - Zero to Hero is the sibling site: concepts, hands-on builds, and the most comprehensive certification library on GitHub.

We build custom software and products at Nobler Works. Open-source training like this is how we give back - we are nothing without the community that supports us. If you need software built, get in touch.

Want the reps as well as the reading?

gitGood.dev - practice questions, coding challenges, system design and cloud certification practice exams for engineers, data, security, DevOps and product technologists

This site gives you the material. gitGood gives you the reps, and tells you whether you actually know it - including ML and AI practice paths built for exactly the material on this site.

10 days free, then $5/month or $40/year. The free tier needs no card.

Fine print

Summaries are original work, released under CC BY 4.0 - free to use with attribution. The papers themselves belong to their authors and are linked, never reproduced.

This is an independent educational resource. It is not affiliated with, endorsed by, or sponsored by OpenAI, Google, Anthropic, Meta, or any other organisation whose work is summarised here. All trademarks belong to their respective owners.

Found something wrong or out of date? Contributions welcome.