The purpose of this curriculum is to help new Elicit employees learn background in machine learning, with a focus on language models. I’ve tried to strike a balance between papers that are relevant for deploying ML in production and techniques that matter for longer-term scalability.

If you don’t work at Elicit yet - we’re hiring ML and software engineers.

Recommended reading order:

1. Read “Tier 1” for all topics
2. Read “Tier 2” for all topics
3. Etc

✨ Added after 2024/4/1

• Fundamentals
  • Introduction to machine learning
  • Transformers
  • Key foundation model architectures
  • Training and finetuning
• Reasoning and runtime strategies
  • In-context reasoning
  • Task decomposition
  • Debate
  • Tool use and scaffolding
  • Honesty, factuality, and epistemics
• Applications
  • Science
  • Forecasting
  • Search and ranking
• ML in practice
  • Production deployment
  • Benchmarks
  • Datasets
• Advanced topics
  • World models and causality
  • Planning
  • Uncertainty, calibration, and active learning
  • Interpretability and model editing
  • Reinforcement learning
• The big picture
  • AI scaling
  • AI safety
  • Economic and social impacts
  • Philosophy
• Maintainer

Tier 1

• A short introduction to machine learning
• But what is a neural network?
• Gradient descent, how neural networks learn

Tier 2

• ✨ An intuitive understanding of backpropagation
• What is backpropagation really doing?
• An introduction to deep reinforcement learning

Tier 3

• The spelled-out intro to neural networks and backpropagation: building micrograd
• Backpropagation calculus

Tier 1

• ✨ But what is a GPT? Visual intro to transformers
• ✨ Attention in transformers, visually explained
• ✨ Attention? Attention!
• The Illustrated Transformer
• The Illustrated GPT-2 (Visualizing Transformer Language Models)

Tier 2

• ✨ Let's build the GPT Tokenizer
• ✨ Neural Machine Translation by Jointly Learning to Align and Translate
• The Annotated Transformer
• Attention Is All You Need

Tier 3

• A Practical Survey on Faster and Lighter Transformers
• TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second
• Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets
• A Mathematical Framework for Transformer Circuits

Tier 1

• Language Models are Unsupervised Multitask Learners (GPT-2)
• Language Models are Few-Shot Learners (GPT-3)

Tier 2

• ✨ LLaMA: Open and Efficient Foundation Language Models (LLaMA)
• ✨ Efficiently Modeling Long Sequences with Structured State Spaces (video) (S4)
• Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5)
• Evaluating Large Language Models Trained on Code (OpenAI Codex)
• Training language models to follow instructions with human feedback (OpenAI Instruct)

Tier 3

• ✨ Mistral 7B (Mistral)
• ✨ Mixtral of Experts (Mixtral)
• ✨ Gemini: A Family of Highly Capable Multimodal Models (Gemini)
• ✨ Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Mamba)
• Scaling Instruction-Finetuned Language Models (Flan)

Tier 2

• ✨ Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer
• Learning to summarise with human feedback
• Training Verifiers to Solve Math Word Problems

Tier 3

• ✨ Pretraining Language Models with Human Preferences
• ✨ Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision
• Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning
• LoRA: Low-Rank Adaptation of Large Language Models
• Unsupervised Neural Machine Translation with Generative Language Models Only

Tier 2

• Chain of Thought Prompting Elicits Reasoning in Large Language Models
• Large Language Models are Zero-Shot Reasoners (Let's think step by step)
• Self-Consistency Improves Chain of Thought Reasoning in Language Models

Tier 3

• ✨ Chain-of-Thought Reasoning Without Prompting
• ✨ Why think step-by-step? Reasoning emerges from the locality of experience

Tier 1

• Supervise Process, not Outcomes
• Supervising strong learners by amplifying weak experts

Tier 2

• ✨ Tree of Thoughts: Deliberate Problem Solving with Large Language Models
• Factored cognition
• Iterated Distillation and Amplification
• Recursively Summarizing Books with Human Feedback
• Solving math word problems with process-based and outcome-based feedback

Tier 3

• ✨ Factored Verification: Detecting and Reducing Hallucination in Summaries of Academic Papers
• Faithful Reasoning Using Large Language Models
• Humans consulting HCH
• Iterated Decomposition: Improving Science Q&A by Supervising Reasoning Processes
• Language Model Cascades

Tier 2

• AI safety via debate

Tier 3

• ✨ Debate Helps Supervise Unreliable Experts
• Two-Turn Debate Doesn’t Help Humans Answer Hard Reading Comprehension Questions

Tier 2

• ✨ Measuring the impact of post-training enhancements
• WebGPT: Browser-assisted question-answering with human feedback

Tier 3

• ✨ AI capabilities can be significantly improved without expensive retraining
• ✨ Automated Statistical Model Discovery with Language Models

Tier 2

• ✨ Self-critiquing models for assisting human evaluators

Tier 3

• ✨ What Evidence Do Language Models Find Convincing?
• ✨ How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions

Tier 3

• ✨ Can large language models provide useful feedback on research papers? A large-scale empirical analysis
• ✨ Large Language Models Encode Clinical Knowledge
• ✨ The Impact of Large Language Models on Scientific Discovery: a Preliminary Study using GPT-4
• A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers

Tier 3

• ✨ AI-Augmented Predictions: LLM Assistants Improve Human Forecasting Accuracy
• ✨ Approaching Human-Level Forecasting with Language Models
• ✨ Are Transformers Effective for Time Series Forecasting?
• Forecasting Future World Events with Neural Networks

Tier 2

• Learning Dense Representations of Phrases at Scale
• Text and Code Embeddings by Contrastive Pre-Training (OpenAI embeddings)

Tier 3

• ✨ Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting
• Not All Vector Databases Are Made Equal
• REALM: Retrieval-Augmented Language Model Pre-Training
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Task-aware Retrieval with Instructions

Tier 1

• Machine Learning in Python: Main developments and technology trends in data science, machine learning, and AI
• Machine Learning: The High Interest Credit Card of Technical Debt

Tier 2

• ✨ Designing Data-Intensive Applications
• A Recipe for Training Neural Networks

Tier 2

• ✨ GPQA: A Graduate-Level Google-Proof Q&A Benchmark
• ✨ SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
• TruthfulQA: Measuring How Models Mimic Human Falsehoods

Tier 3

• FLEX: Unifying Evaluation for Few-Shot NLP
• Holistic Evaluation of Language Models (HELM)
• Measuring Massive Multitask Language Understanding
• RAFT: A Real-World Few-Shot Text Classification Benchmark
• True Few-Shot Learning with Language Models

Tier 2

• Common Crawl
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling

Tier 3

• Dialog Inpainting: Turning Documents into Dialogs
• MS MARCO: A Human Generated MAchine Reading COmprehension Dataset
• Microsoft Academic Graph
• TLDR9+: A Large Scale Resource for Extreme Summarization of Social Media Posts

Tier 3

• ✨ Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task
• ✨ From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought
• Language Models Represent Space and Time

Tier 2

• ✨ Experts Don't Cheat: Learning What You Don't Know By Predicting Pairs
• A Simple Baseline for Bayesian Uncertainty in Deep Learning
• Plex: Towards Reliability using Pretrained Large Model Extensions

Tier 3

• ✨ Active Preference Inference using Language Models and Probabilistic Reasoning
• ✨ Eliciting Human Preferences with Language Models
• Active Learning by Acquiring Contrastive Examples
• Describing Differences between Text Distributions with Natural Language
• Teaching Models to Express Their Uncertainty in Words

Tier 2

• Discovering Latent Knowledge in Language Models Without Supervision

Tier 3

• ✨ Interpretability at Scale: Identifying Causal Mechanisms in Alpaca
• ✨ Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks
• ✨ Representation Engineering: A Top-Down Approach to AI Transparency
• ✨ Studying Large Language Model Generalization with Influence Functions
• Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

Tier 2

• ✨ Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• ✨ Reflexion: Language Agents with Verbal Reinforcement Learning
• Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (AlphaZero)
• MuZero: Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model

Tier 3

• ✨ Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
• AlphaStar: mastering the real-time strategy game StarCraft II
• Decision Transformer
• Mastering Atari Games with Limited Data (EfficientZero)
• Mastering Stratego, the classic game of imperfect information (DeepNash)

Tier 1

• Scaling Laws for Neural Language Models
• Takeoff speeds
• The Bitter Lesson

Tier 2

• AI and compute
• Scaling Laws for Transfer
• Training Compute-Optimal Large Language Models (Chinchilla)

Tier 3

• Emergent Abilities of Large Language Models
• Transcending Scaling Laws with 0.1% Extra Compute (U-PaLM)

Tier 1

• Three impacts of machine intelligence
• What failure looks like
• Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover

Tier 2

• ✨ An Overview of Catastrophic AI Risks
• Clarifying “What failure looks like” (part 1)
• Deep RL from human preferences
• The alignment problem from a deep learning perspective

Tier 3

• ✨ Scheming AIs: Will AIs fake alignment during training in order to get power?
• Measuring Progress on Scalable Oversight for Large Language Models
• Risks from Learned Optimization in Advanced Machine Learning Systems
• Scalable agent alignment via reward modelling

Tier 3

• ✨ Explosive growth from AI automation: A review of the arguments
• ✨ Language Models Can Reduce Asymmetry in Information Markets

Tier 2

• Meaning without reference in large language models

[email protected]