In brief

This directory collects comprehensive syntheses of major scientific and technical publications in the field of artificial intelligence. Each document is structured by chapter and available for PDF download for in-depth reading.

Available syntheses

Attention Is All You Need — Vaswani et al. (NIPS 2017)

The founding paper. Twelve pages published in June 2017 that gave us the Transformer — the architecture underpinning GPT, Claude, Gemini, Llama and Mistral today. Its contribution was not winning two points of translation score: it was making training parallelisable, by removing the recurrence that forced words to be read one after another. Once that bottleneck was gone, the limiting factor shifted from time to budget — and scaling became possible.

The synthesis covers what the paper actually says: the attention mechanism and its scaling factor, multi-head attention, the three uses of attention in the architecture, positional encoding, the complexity table behind the quadratic cost of long contexts, the training protocol and the results. It closes by separating what has held for nine years from what has been replaced — and what the paper does not claim, despite what is often attributed to it.

Note: the synthesis is written in French.

Source: arXiv:1706.03762

Download: FR synthesis — PDF, 10 pages

Multi-Agent Systems with LLMs — A. Gulli et al. (2025)

Complete synthesis of the reference work on multi-agent systems powered by large language models. Covers design patterns, orchestration, tools, memory, and deployment.

Download: FR Synthesis — PDF

HyperAgents: Self-Referential Self-Improving Agents — Zhang et al. (2026)

Synthesis of the Meta FAIR research paper on self-referential agents capable of modifying their own improvement mechanism. Introduces the Darwin Gödel Machine concept with HyperAgents (DGM-H) and demonstrates open-ended self-improvement beyond the code domain.

Source: arXiv:2603.19461 | GitHub

Download: FR Synthesis — PDF

Towards End-to-End Automation of AI Research — C. Lu, C. Lu, R.T. Lange et al. (Nature, 2026)

Synthesis of the Nature paper on The AI Scientist — a fully agentic pipeline for automating scientific research, from idea generation to peer review submission. Covers both operating modes (template-based and open-ended), quantitative results, failure modes, and ethical implications.

Source: Nature Vol. 651, 26 March 2026

Download: FR Synthesis — PDF

Conversationally Cognizant AI — J. Morrissette (2026), Blocks A+D

Synthesis of the 13 foundational concept papers (Foundations and Architecture blocks) of the "Conversationally Cognizant AI" series. Covers Morrissette's architectural vision: conversation as state, intention as the unifying mechanism, the MAD pattern, progressive cognitive pipeline, deterministic routing, code agent-optimal architecture, and inter-agent communication.

Source: GitHub repository jmorrissette-RMDC/portfolio — withdrawn by its author; link removed on 2026-09-06 after verification (HTTP 404). The PDF synthesis below is now the only access to the content.

Download: FR Synthesis Blocks A+D — PDF

Conversationally Cognizant AI — J. Morrissette (2026), Blocks B+C

Synthesis of the remaining 10 concept papers (Cognition & Autonomy and Core Components blocks) of the "Conversationally Cognizant AI" series. Covers collective reasoning (Quorum pattern), ecosystem learning, progressive agent autonomy, autonomous development cycle, autoprompting, the meta-architect, and the three brokers (Context, Inference, Conversation).

Source: GitHub repository jmorrissette-RMDC/portfolio — withdrawn by its author; link removed on 2026-09-06 after verification (HTTP 404). The PDF synthesis below is now the only access to the content.

Download: FR Synthesis Blocks B+C — PDF

MIRAGE — The Illusion of Visual Understanding in LLMs — Asadi et al. (Stanford, 2026)

Synthesis of the Stanford study demonstrating that state-of-the-art multimodal models produce confident visual descriptions without actually analysing images. Covers the "mirage reasoning" phenomenon, contamination of 74% of multimodal benchmarks, and the B-Clean framework for robust evaluation.

Source: arXiv:2603.21687

Read: Full article

Key takeaways

  • Each synthesis covers the full source document, chapter by chapter
  • Technical terms are kept in their original language with translations where needed
  • PDF format allows comfortable offline reading
  • Original sources are systematically credited

Sources