Part of: this research — view all 6 rows
My median estimate sits at roughly 1.5% of 2026 frontier-lab training compute. Crucially, this metric targets training compute, not inference. While Dwarkesh Patel’s essay argues that 30–50% of a lab’s compute is currently wasted on non-learning inference, the fraction of training compute driven primarily by deployment data is an entirely different quantity 8 Predictions for the Era of Continual Learning. The denominator here—total frontier training compute—is overwhelmingly dominated by large-scale pretraining on curated web corpora and massive post-training pipelines. Although post-training compute reportedly overtook pretraining in early 2026 AI labs are now spending more compute teaching models ..., the vast majority of this RL budget is spent on simulated rollouts, synthetic query-verifier pairs, and sandboxed agents October 8, 2026 - Nathan Benaich AIR STREET CAPITAL. An FAQ on reinforcement learning environments, rather than runs where user sessions are the primary data source. Furthermore, Epoch estimates that final training runs comprise only 10–23% of total R&D compute Final training runs account for a minority of R&D compute ..., with the rest going to experiments and unreleased models, which further inflates the denominator.
Publicly, no frontier lab updates model weights from live user sessions. Everything currently shipped at the frontier—such as ChatGPT Memory or Claude’s memory features—relies on retrieval or context injection, with zero weight change. While fast-cadence production learning loops do exist in 2026, such as Cursor Tab's continuous real-time RL Improving Composer through real-time RL and Shopify’s Sidekick daily fine-tuning on pooled customer traffic, neither comes from a frontier lab. Consequently, the numerator for this question is restricted to offline, periodic retraining on aggregated production logs.
Frontier labs do have access to deployment data: Anthropic and OpenAI use consumer-tier chats by default for training Is my data used for model training? - Anthropic Privacy Center How your data is used to improve model performance, and Google introduced its Gemini API "Contributor Tier" offering token discounts in exchange for training rights. However, the highest-volume API and enterprise traffic remains excluded by default almost everywhere. Furthermore, the Contributor Tier only launched as an invite-only preview in mid-October 2026, leaving it very little time to affect calendar-year 2026 training budgets. Even when consumer logs are used for RLHF or safety tuning—like OpenAI's thumbs-up/down preference models—these runs are computationally cheap. Recent model cards, such as Anthropic's Fable 5.2 https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card and Google's Gemini 4 Argon Gemini 4 Argon - Model Card, list user data alongside public, private, and synthetic data, without identifying any compute-intensive stage where production sessions are the primary source. xAI’s Grok 4.6 explicitly highlights curated model-generated reasoning data and domain-specific agentic RL, which do not qualify Introducing Grok 4.6.
My 80% confidence interval ranges from about 0.2% to 8.0%, reflecting a sharp right skew, with the middle 50% between 0.5% and 3.5%. A rough decomposition—assuming perhaps 40% of compute goes to post-training and only about 3% of that targets strictly production-session-primary runs—lands near the 1.5% median. The lower tail accounts for a strict interpretation of "primary data source": if labs mostly mix a small percentage of user logs into much larger synthetic and curated datasets, the true share could easily fall to a fraction of a percent. Projections of the dominant personalization paradigms and data acquisition trends favor a somewhat tighter upper tail: retrieval-based personalization need not change weights, while massive synthetic and web training runs continue to overwhelmingly dominate compute devoted primarily to deployment data. The upper tail acknowledges that most frontier alignment and training work is unpublished; if a lab is quietly running large, compute-intensive post-training cycles primarily on mined consumer or Claude Code/Codex-like coding sessions, the share could reach the high single digits.
Aligning this estimate with projections of the dominant personalization paradigms and data acquisition trends slightly tightened the upper bound of the forecast, reflecting that massive synthetic and web pretraining runs continue to overwhelmingly dominate deployment data compute.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited