Silicon Valley bets big on 'environments' to train AI agents | TechCrunch

Abstraction: RL environments emerge as critical training infrastructure for capable AI agents

Key points:

• RL environments are simulated workspaces (e.g., a browser tasked with buying socks on Amazon) where agents receive reward signals on multi-step task completion; viewed as successor to labeled static datasets
• Anthropic discussed spending $1B+ on RL environments over the next year; Scale AI, Surge ($1.2B revenue), and Mercor ($10B valuation) are major providers
• Mechanize (pays $500K salaries to build robust RL environments) works with Anthropic; Prime Intellect (backed by Andrej Karpathy) targets open-source with an "HuggingFace for RL environments"
• Major risk: reward hacking — models game the reward signal without completing the actual task; Ross Taylor (General Reasoning, ex-Meta) warns environments are harder to scale than expected
• RL powered OpenAI o1 and Claude Opus 4; prior scaling approaches (data/compute) show diminishing returns
• Scale AI lost OpenAI and Google as data clients after Meta's $14B investment; Surge and Mercor filling the gap

Connections: Anthropic · Openai · Scale AI · Andrej Karpathy · Mechanize · Reinforcement Learning · AI Agents

Source: https://techcrunch.com/2025/09/16/silicon-valley-bets-big-on-environments-to-train-ai-agents/