Description:

This question resolves to YES if at least TWO of the following three key milestones, predicted for Mid-to-Late 2025 in the "AI 2027" scenario (published April 3, 2025, by Kokotajlo et al. at ai-2027.com), are demonstrably achieved by the resolution date.

Resolution Criteria:

1. "Stumbling Agents" Become Available & Used: Publicly available AI agents marketed as "personal assistants" capable of multi-step tasks (e.g., ordering food online, basic budget spreadsheet manipulation) exist, OR specialized coding/research agents demonstrably function autonomously within companies, making substantial code changes or conducting multi-source research (beyond simple 2024-level instruction following). Evidence requires widespread reporting, product releases from major AI labs, or significant documented corporate adoption. Simple chatbots or single-function tools are insufficient.

2. High Cost for Top Agents: Leading, publicly available AI agents (analogous to the scenario's advanced agents) have a typical subscription cost exceeding US$100 per month for individual prosumer or enterprise use. Evidence requires pricing information from at least two major AI providers (e.g., OpenAI, Google DeepMind, Anthropic, Microsoft, or plausible analogues).

3. "World's Most Expensive AI" Infrastructure Push: At least one leading AI company (e.g., OpenAI, Google DeepMind, Anthropic, or a clear analogue to the fictional "OpenBrain") publicly announces or is credibly reported (e.g., via major news outlets, satellite imagery analysis) to be constructing new datacenter infrastructure explicitly stated or widely understood to be for training models significantly larger (>10x compute) than GPT-4 class models (i.e., targeting >>2e25 FLOP, towards the 1e27-1e28 FLOP scale mentioned in the scenario).

Verification must rely on credible public sources (major news outlets, official company announcements/pricing, reputable industry analysis) available by December 31, 2025.