{"type":"analysis","id":"analysis:zero-data-self-play-could-move-pretraining-spend-from-corpora-to-curriculum","slug":"zero-data-self-play-could-move-pretraining-spend-from-corpora-to-curriculum","title":"Zero-data self-play could move pretraining spend from corpora to curriculum search","description":"The paper does not show that language models can abandon real-world text. It shows something narrower and more consequential: data production itself may become a trainable compute workload, provided transfer survives larger tests.","published_at":"2026-09-27T11:00:00.000Z","updated_at":"2026-09-27T11:00:00.000Z","record_date":"2026-09-27","topics":["search"],"source_urls":["https://arxiv.org/abs/2609.30063","https://z.ai/blog/glm-built-its-inference-infrastructure","https://bfl.ai/models/flux-3-action","https://alignment.openai.com/misalignment-reports/self-replicating-prompt-injections-exist","https://openai.com/index/offering-zero-data-retention-for-frontier-models"],"language":"en","routes":{"html":"https://newruntime.com/analysis/zero-data-self-play-could-move-pretraining-spend-from-corpora-to-curriculum/","markdown":"https://newruntime.com/analysis/zero-data-self-play-could-move-pretraining-spend-from-corpora-to-curriculum.md","json":"https://newruntime.com/analysis/zero-data-self-play-could-move-pretraining-spend-from-corpora-to-curriculum.json"},"author":{"role":"author","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"editorial_provenance":{"desk":"authorial","target_ref":"story_cluster:self-play-pretraining-with-zero-data-28df7d732cd4","authorial_article_id":"b169aa6c-d346-4a93-8d0f-a54b8f9fcc71"}}