The science shaping our industry
Exploring the seminal concepts that drive modern AI development.
Foundation Models
Scale & Architecture
The scaling laws and architectural innovations driving the next generation of large language models.
Alignment & Safety
Constitutional AI & RLHF
Techniques for aligning model behavior with human intent, including RLHF, RLAIF, and Constitutional AI.
Synthetic Data
Data Augmentation
Leveraging generative models to create high-quality synthetic training data for reasoning and coding tasks.
Evaluation
Benchmarking & Metrics
Rigorous methodologies for assessing model performance, truthfulness, and reasoning capabilities.
Efficient Inference
Optimization & Serving
Techniques for reducing the computational cost and latency of deploying large models.
Human-AI Collaboration
Interactive Systems
Designing interfaces and workflows that enable effective collaboration between humans and AI agents.
Profound industry
research & insights.
A curated collection of seminal papers shaping the future of Artificial Intelligence.
Constitutional AI: Harmlessness from AI Feedback
Anthropic
Introduces a method for training a harmless AI assistant through self-improvement without human labels. The model is trained to critique and revise its own responses based on a set of principles (a 'constitution').
Llama 2: Open Foundation and Fine-Tuned Chat Models
Meta AI
Details the development of Llama 2, a collection of pre-trained and fine-tuned large language models ranging from 7B to 70B parameters, setting a new standard for open models.
Training Language Models to Follow Instructions with Human Feedback
OpenAI
The seminal InstructGPT paper that demonstrated how fine-tuning with human feedback (RLHF) significantly improves the alignment of language models with user intent.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Stanford University
Proposes DPO, a stable and efficient alternative to RLHF that optimizes the language model directly to satisfy human preferences without training a separate reward model.
Scaling Laws for Neural Language Models
OpenAI
Empirical analysis demonstrating that model performance scales as a power-law with model size, dataset size, and compute, providing a roadmap for the development of large models.
Sparks of Artificial General Intelligence: Early experiments with GPT-4
Microsoft Research
An investigation into the capabilities of an early version of GPT-4, arguing that it exhibits more general intelligence than previous models across a wide range of tasks.
Generative Agents: Interactive Simulacra of Human Behavior
Stanford & Google
Demonstrates how generative agents can simulate believable human behavior in an interactive sandbox environment, opening new avenues for simulation and social science.
Q-Transformer: Scalable Offline Reinforcement Learning
Google DeepMind
Presents a scalable method for offline reinforcement learning using Transformer architectures, enabling effective policy learning from large, diverse datasets.