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

rationale:

Over the brief 8.5-month horizon to June 2027, frontier AI practice will overwhelmingly rely on context injection and memory features rather than live weight updates. Since August 2026, every major release from top labs has centered on expanding memory and context without altering model weights. Recent examples include GPT-6 Astra allowing Codex to maintain full session context https://www.ithinkdiff.com/gpt-6-astra-launch-codex-memory-cybersecurity/, Google's Gemini agent introducing multi-type episodic and semantic memory Gemini at Work 2026: Introducing Gemini agentGemini at Work 2026: Introducing Gemini agent, xAI's Team Bots learning via files and skills Team Bots: shared AI teammates that learn as they work, and Anthropic's categorized memory expansions Release notes | Claude Help Center. Aligning with this structural trend away from live weight updates and the rapid expansion of these features by major providers, we assign 75% to "Context and memory only, with no weight change" definitively representing the dominant commercial practice.

The prospect of per-customer weight forks (2%) or even adapter-only personalization (5%) is actively shrinking. Notably, the market leader is moving in the opposite direction: OpenAI is entirely phasing out self-serve fine-tuning, prohibiting new jobs across all customers by January 6, 2027 https://developers.openai.com/api/docs/deprecations. OpenAI's enterprise guidance explicitly steers customers away from weight modification, pointing them instead toward RAG, memory, and evaluation OpenAI is shutting down self-serve fine-tuning. Furthermore, known technical barriers such as catastrophic forgetting, sample inefficiency in gradient updates, and alignment drift remain persistent bottlenecks Google's Gemini co-leads explain what comes nextEntering the era of AI models that can continually self-learn. Even Dwarkesh Patel's essay concedes that single-session data is currently insufficient for true per-user continual learning 8 Predictions for the Era of Continual Learning.

While we allocate 13% to "Periodic aggregate retraining, monthly or faster," true high-frequency weight updates are currently restricted to non-frontier models. Systems like Shopify—which compresses merchant data into a shared model daily—and Cursor Tab operate well below the frontier scale How Shopify built a continual learning loop with PyTorch ...Sidekick's continual learning loop (2026). The primary risk to this forecast is definitional: frontier labs maintain a roughly six-week release cadence (e.g., OpenAI releases in May, July, August, and September 2026), and a lenient resolver might interpret this rapid post-training iteration on pooled consumer data as monthly aggregate retraining GPT-5.6 — August Updates - Deployment Safety Hub - OpenAI.

Finally, we assign "No meaningful change from 2026" a low 5% probability because the current pace of memory and context feature development is intense. The resolution criteria dictate that this option only wins if June 2027 is indistinguishable from October 2026. With the rapid rollout of features like OpenAI's Dreaming V3 Dreaming: Better memory for a more helpful ChatGPT and persistent agent workspaces, context management is demonstrably advancing. The future Dwarkesh Patel dismissed—where AI personalization scales entirely via sophisticated memory and context without touching the weights—is exactly the paradigm frontier labs are successfully commercializing.