Insights · 7 min read

The future of AI training data is environments, not text

The scraping era is ending. The practice era has started.

Bearish AI · October 21, 2026

The text is running out

Epoch AI estimates the effective stock of public human-written text at around 300 trillion tokens, and projects that, if trends continue, language models will fully use it sometime between 2026 and 2032. More scraping won’t carry the next leap.

Synthetic data can’t fill the gap alone

Generating training data with models is cheap, but it has limits. A study in Nature found that models trained repeatedly on generated data lose the rare, unusual parts of the original distribution, a failure the authors call model collapse. The rare parts are often exactly what matters: the edge case, the odd bug, the decision that went against the obvious choice.

Labs are moving to environments

The industry has shifted from showing models text to letting them act. TechCrunch reported in 2025 that Anthropic leaders discussed spending more than $1 billion on reinforcement learning environments over a year, and quoted Andreessen Horowitz partner Jennifer Li: “All the big AI labs are building RL environments in-house,” while also looking to outside vendors.

Epoch AI’s research describes demand spreading from math and coding into enterprise workflows, and rising interest in longer, multi-step tasks.

What the next generation of data needs

• Verifiable: every task tied to a check that can say pass or fail automatically.
• Fresh: drawn from work that was never public, so models can’t have memorized it.
• Realistic: real systems with real complexity, not simplified replicas.
• Process-rich: the reasoning and the mistakes, not just the final answer.
• Multilingual by design: written natively, not translated after the fact.

Where Bearish AI fits

Bearish AI was built for this shift. We supply verifiable tasks rebuilt from real development problems, a real enterprise platform as a practice environment, the complete build history of that same system, and professionals who make original business scenarios across finance, accounting, procurement, logistics, sales, customer service and administration, Arabic-first for the Middle East and North Africa.

Text taught models to talk. Environments will teach them to work. We intend to supply both the practice and the proof.