The Canon of IA
There is a fundamental difference between those who use AI and those who understand AI. The Library of PrezenceAI is the curation of knowledge that builds real understanding — foundational papers, books that have withstood the test of time, and documents that reveal what lies behind interfaces.
Why Reading Original Papers Matters
Most practitioners of AI learn through tutorials, Twitter threads, and YouTube videos. This is sufficient to use existing tools — it is not sufficient to understand why they work or to build something new.The difference between an advanced user and a genuine AI engineer is the ability to read and understand original papers.
The fundamental corpus of modern AI fits in fewer than 50 papers. "Attention is All You Need" (2017), "Language Models are Few-Shot Learners" (GPT-3, 2020), "Training language models to follow instructions" (InstructGPT, 2022), "Constitutional AI" (2022), "Llama 2" (2023) — Those who have read these papers carefully understand 80% of what is happening in the field today. The rest is variation.
The paradox of the AI learning ecosystem: There has never been so much material available, and never has it been so easy to learn superficially.The filter that separates genuine knowledge from superficial familiarity is exactly the willingness to sit with a dense paper, understand every equation, and reconstruct the intuition that generated it. The Library helps identify what worth this investment.
15 Terms that Define Library
Bibliographic and academic terminology for the Pointy Library. Used to contextualize resources within the intellectual trajectory of the field.
| Term | Editorial Definition | Level |
|---|---|---|
| Preprint | Paper submitted to arXiv before peer review — standard in ML/IA; speed > formality | Gold |
| Peer Review | Peer Review — NeurIPS, ICML, ICLR are the main conferences; process takes 3-6 months | Gold |
| arXiv | Preprint Repository — cs.AI, cs.LG, cs.CL as relevant sections; free and open | Diamond |
| Citation Count | Citation Count | Silver |
| Ablation Study | Experiment that removes components to measure individual contribution — standard rigor in AI papers | Diamond |
| Reproducibility | Ability to replicate results from a paper — the reproducibility crisis affects 40-60% of ML papers | Diamond |
| Hugging Face Papers | Aggregator of papers with interactive demos and code — democratizes access to frontier research | Gold |
| Connected Papers | Network citation visualization tool — navigates literature by similarity | Silver |
| Papers With Code | Repository of papers with associated code — best implementations and benchmarks | Gold |
| Semantic Scholar | Academic search engine with AI — automatic summaries and recommendations | Silver |
| Google Scholar | Academic indexer — tracks citations and paper versions | Silver |
| Survey Paper | Review paper of the literature — efficient entry point for a new area | Gold |
| Technical Report | Technical document published directly by a company — OpenAI, Anthropic, Google; no formal review | Silver |
| Benchmark Paper | Paper proposing a new evaluation suite — MMLU, HumanEval, GPQA as influential examples | Gold |
| Scaling Law Paper | Papers that quantify the relationship between parameters/data/compute — Kaplan (2020), Chinchilla (2022) as milestones | Diamond |
15 Mandatory Readings for AI Solution Architects
An AI reading list can have 1,000 items — or 15. This is the version of 15: the texts that, when read with real attention and understanding, build a solid intellectual foundation for navigating any future development in the field. Curation prioritizes texts that age well and explain fundamental mechanisms rather than techniques that will become obsolete in 18 months.
The Mathematical Foundations
1. "Attention is All You Need" — Vaswani et al. (2017). The paper that created the transformer. Read section 3 carefully — the multi-head attention mechanism explained from scratch. 2. "An Introduction to Neural Networks" — Nielsen (2015, online free). The best entry point for someone who wants to truly understand backpropagation. 3. "The Annotated Transformer" — Harvard NLP (haveibeenpwned.com). Line-by-line implementation with comments — the Vaswani paper translated into readable Python code.
The Revolution of LLMs
4. "Language Models are Few-Shot Learners" — Brown et al. / OpenAI (2020). The GPT-3 paper — proof that scaling produces emergent behaviors. 5. "Training language models to follow instructions" — Ouyang et al. / OpenAI (2022). InstructGPT — how RLHF transformed capable models into useful ones. 6. "Constitutional AI" — Bai et al. / Anthropic (2022). Alternative to RLHF — alignment via explicit principles.
Scale, Efficiency and Open Source
7. "Scaling Laws for Neural Language Models" — Kaplan et al. / OpenAI (2020). Laws that govern the relationship between size, data, and performance. 8. "Training Compute-Optimal LLMs" (Chinchilla) — Hoffmann et al. / DeepMind (2022). Redefined the ideal proportions of parameters × tokens. 9. "Llama 2" — Touvron et al. / Meta (2023). The paper that democratized frontier models — and the most well-written technical document of its generation.
What's Coming Next
10. "ReAct: Synergizing Reasoning and Acting" — Yao et al. (2022). The technical foundation of agents that reason and act. 11. "Toolformer" — Schick et al. / Meta (2023). How models learn to use external tools. 12. "Chain-of-Thought Prompting" — Wei et al. / Google (2022). Simple and powerful: show the reasoning, the model learns to reason. 13. "Sparks of AGI" — Bubeck et al. / Microsoft (2023). Controversial, but rigorously analyzes GPT-4 with rare depth for a corporate paper. 14. "Risks from Learned Optimization" — Hubinger et al. (2019). The reference paper on table alignment. 15. "Deep Learning" — Goodfellow, Bengio & Courville (2016). The standard textbook — free online at deeplearningbook.org.