Reading list

Papers we’re reading

A curated, annotated library of the work most relevant to forecasting AGI — with a short note on why each one earns a place. Updated continually.

Paper · 2026

Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents

Tianshi Xu, Huifeng Wen, Meng Li · Peking University

Harness that aims to "unfreeze" models through remembering and injecting knowledge about past failures.

arxiv.org

Paper · 2020

Scaling Laws for Neural Language Models

Jared Kaplan, Sam McCandlish, et al. · OpenAI

The paper that turned capability into an engineering forecast: loss falls as a smooth power law in compute, data, and parameters. The empirical backbone of the "scaling is enough" thesis we continually stress-test.

ScalingFoundational arxiv.org

Paper · 2022

Training Compute-Optimal Large Language Models

Jordan Hoffmann, Sebastian Borgeaud, et al. · DeepMind

The "Chinchilla" result recalibrated the scaling recipe: most large models were badly undertrained on data. It reset everyone's compute-to-data ratios and remains essential for any honest forecast of where the curves go next.

ScalingCompute arxiv.org

Paper · 2023

Sparks of Artificial General Intelligence: Early Experiments with GPT-4

Sébastien Bubeck, Varun Chandrasekaran, et al. · Microsoft Research

A sprawling, contested case that GPT-4 shows early signs of generality. Read it alongside its critics: it is as much a document about how hard AGI is to evaluate as it is a claim about the model.

AGIEvaluation arxiv.org

Paper · 2022

Emergent Abilities of Large Language Models

Jason Wei, Yi Tay, et al. · Google Research

Documents capabilities that appear abruptly at scale rather than improving smoothly. Central to the optimistic forecast — and to the debate over whether "emergence" is real or an artefact of how we measure it.

EmergenceScaling arxiv.org

Paper · 2023

Are Emergent Abilities of Large Language Models a Mirage?

Rylan Schaeffer, Brando Miranda, Sanmi Koyejo · Stanford University

The essential counterpoint: argues apparent "emergence" can be an artefact of discontinuous metrics, and smooths away under better-chosen ones. We keep it paired with the emergence literature to keep ourselves honest.

EmergenceSkeptical arxiv.org

Paper · 2022

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Jason Wei, Xuezhi Wang, et al. · Google Research

Showed that simply asking models to reason step by step unlocks latent capability — the seed of today's test-time-compute paradigm and the second scaling axis we now track on the dashboard.

ReasoningMethods arxiv.org