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:
"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.
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).
"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.
1,000
3.00Why Resolved Yes:
1. “Stumbling Agents” Are Here
Major product launches: In January 2025, OpenAI unveiled “Operator,” an AI agent that can autonomously book flights, order groceries, and complete multi-step web tasks by simulating browser interactions BNN.
Commercial deployments: Baidu released its “Xinxiang” agent on Android, marketed as a personal assistant for complex tasks like travel planning and data analysis, with corporate adoption underway Reuters.
Industry consensus: Analyst write-ups and developer toolkits (e.g., OpenAI’s “building blocks for agents”) confirm that multi-step, agentic systems are no longer prototypes but public products OpenAI.
These developments go beyond simple chatbots, fulfilling the scenario’s requirement for agents that perform genuine multi-step tasks.
2. High Cost for Top Agents
OpenAI pricing: The “Operator” agent is gated behind ChatGPT Pro, which costs US$200 per month for individual users TechCrunch.
Anthropic’s premium tier: Anthropic’s new “Max Plan” for its Claude agent offers 20× the usage of its Pro tier at either US$100 or US$200 per month The Verge.
Specialized enterprise agents: Reports indicate OpenAI plans to charge up to US$2,000–$20,000 per month for high-end “knowledge worker” and “PhD-level” agents TechCrunch.
With two leading labs (OpenAI and Anthropic) offering agent subscriptions above US$100/month, the high-cost milestone is unequivocally met.
3. “World’s Most Expensive AI” Infrastructure Push
Stargate announcement: In January 2025, President Trump and company executives unveiled the US$500 billion “Stargate” project—a joint venture of OpenAI, SoftBank, Oracle, and others—to build dozens of AI-optimized datacenter campuses in the US Reuters.
Site expansions: Reuters reports OpenAI is actively scouting sites across 16 states for these new AI datacenters, with initial construction begun in Texas and plans for multiple additional campuses Reuters.
Compute scale: The scale of investment and stated purpose—to train models at orders-of-magnitude beyond GPT-4’s ~2e25 FLOP—matches the scenario’s “>>1e27–1e28 FLOP” target The Verge.
This unprecedented build-out clearly qualifies as the “world’s most expensive AI” infrastructure push described in the AI 2027 scenario.