As of 2027-06-30, which of these will best describe continual learning in commercial practice at the frontier?

rationale:

As of late 2026, dominant commercial practice relies entirely on retrieval and context injection. Given the rapid pace of development in agentic memory and cross-session synthesis, the mid-2027 landscape will likely see advanced context features scaling further. Set against related questions, the distribution remains mostly unchanged, maintaining the strong likelihood that context and memory injection will dominate over per-customer weight forks by mid-2027, keeping the bulk of probability (60%) in 'Context and memory only, with no weight change.' The alternative of 'Periodic aggregate retraining, monthly or faster' (23%) captures the possibility that frontier labs pivot to heavily rely on deployment-driven retraining loops operating on a monthly cadence or faster. 'Adapter-only personalization' (9%) faces structural headwinds as major providers have begun phasing out self-serve fine-tuning. Most importantly, 'Per-customer weight forks updated weekly or faster' (3%) is highly unlikely on this ten-month horizon. Severe safety constraints, enterprise privacy demands, and the lack of existing frontier production systems push the timelines for continuous shared-weight updates (10th percentile ~2030) and per-customer merging (10th percentile ~2031) well into the next decade. This leaves minimal runway for rapid per-customer weight forks to dominate by mid-2027. Finally, 'No meaningful change from 2026' receives 5% to account for a completely stalled environment where even memory injection features fail to advance.