The LLM reading list Elicit uses to onboard engineers

What it does

This repository is Elicit’s internal machine-learning onboarding curriculum, open-sourced as a tiered reading list. It links to papers, videos, and technical reports—ranging from introductory neural-network explainers to the DeepSeek-R1 and Llama 3 papers—and arranges them by topic and difficulty. The goal is to take a new hire from scratch to the research frontier without drowning in undifferentiated arXiv feeds.

The interesting bit

The maintainer explicitly balances “deploy ML in production” against “longer-term scalability,” so the list mixes engineering hygiene with speculative research. The strict tiering—finish Tier 1 on every topic before advancing—treats the repository like a syllabus with prerequisites rather than a passive bibliography.

Key highlights

• Covers the full stack: transformer fundamentals, key architectures (GPT-2 through Qwen2.5), training and finetuning recipes, reasoning strategies, and applied domains like science and forecasting.
• Includes both canonical sources (Attention Is All You Need, Karpathy’s video lectures) and recent additions such as multi-token prediction, QLoRA, and the Byte Latent Transformer paper.
• Organized by topic with mandatory tiered progression, preventing readers from jumping to advanced interpretability or world models before digesting core mechanics.
• Explicitly bridges research and engineering, mixing production-deployment notes with scaling theory, benchmarks, and AI-safety philosophy.

Caveats

• The list is overwhelmingly language-model-centric; vision and multimodal coverage appear only sporadically.
• The public README is truncated, so the full depth of Tier 4+ material is only partially visible.
• There are no exercises, implementations, or comprehension checks—this is pure curation, not a course platform.

Verdict

New hires joining an LLM team and self-directed learners who want an opinionated path through the paper firehose should bookmark this. If you need runnable code, interactive notebooks, or a quick reference card, it will not help; this is a bibliography with a point of view.