Source note
AI 2027
Summary
AI 2027 is a concrete forecast scenario for rapid AI agent progress through 2027, with special focus on AI systems that automate coding, AI research, cyber work, and model development.
Problem
- The paper addresses a forecasting problem: near-term AGI claims are often vague, so the authors give a dated scenario that can be debated and checked later.
- The stakes matter because the scenario links AI coding agents, automated AI R&D, model-weight theft, cyber capability, labor disruption, and U.S.-China competition.
- It also frames alignment as an unresolved control problem: companies can train models against written rules, but they cannot directly verify the model's internal goals.
Approach
- The authors build an iterative scenario rather than a benchmarked model: they write a timeline, revise it, and branch it into two endings called “slowdown” and “race.”
- Their inputs include trend extrapolation, about 25 tabletop exercises, feedback from over 100 people, expert review in AI governance and technical AI work, and prior forecasts by the authors.
- The core mechanism in the scenario is recursive AI acceleration: better agents help automate AI R&D, which produces stronger agents, which then speed up the next round of research.
- The technical story uses named capabilities and mechanisms: autonomous coding agents, synthetic data, reinforcement learning on long-horizon tasks, neuralese recurrence and memory, and iterated distillation and amplification.
- The geopolitical mechanism is a compute and model-weight race: OpenBrain scales datacenters and internal agents, while China tries to close the gap through centralization, chip routing, and theft of model weights.
Results
- The excerpt provides no experimental benchmark results or dataset scores. Its strongest claims are dated quantitative forecasts.
- Training compute is projected to rise from GPT-4 at about 3 x 10^23 FLOP, to Agent-0 at about 2 x 10^25 FLOP, to a later OpenBrain cluster able to train with about 4 x 10^27 FLOP, roughly 1,000x GPT-4.
- In early 2026, Agent-1 is forecast to make OpenBrain's algorithmic AI progress 50% faster, described as a 1.5x AI R&D progress multiplier.
- By January 2027, Agent-2 is forecast to roughly triple OpenBrain's algorithmic progress and reach near top-human research-engineering ability, with research taste near the 25th percentile OpenBrain scientist.
- China is forecast to hold about 12% of world AI-relevant compute in mid-2026, run about six months behind OpenBrain models, direct almost 50% of its AI-relevant compute into a DeepCent-led collective, and route over 80% of new chips to the Centralized Development Zone.
- The scenario claims Agent-1-mini is 10x cheaper than Agent-1, the 2026 stock market rises 30% led by AI-linked firms, and a 2027 model-weight theft can exfiltrate a roughly 2.5 TB half-precision checkpoint in under two hours using about 25 servers leaking about 100 GB each.