Part of: this research — view all 7 rows
Current Deployments and Strict Requirements Nothing qualifying exists today at a top-five lab by revenue (OpenAI, Anthropic, Google, xAI, or Mistral). All shipped memory features at the frontier—such as ChatGPT Memory, Claude memory, and Gemini Personal Context—rely on retrieval or context injection rather than weight updates Frontier model uses daily weight updates / online learning. The resolution criteria specifically require merging per-customer weight updates back into a shared base model. This excludes systems that merely pool data for centralized retraining, such as Shopify's Sidekick (which runs a daily full-parameter fine-tune on failures pooled across millions of merchants) Sidekick's continual learning loop (2026) or Cursor Tab's online reinforcement learning Improving Composer through real-time RL. The upstream requirement of generating live-session weight updates remains entirely absent at the frontier Models shift to continuous/iterative training - AI 2027 Tracker.
Technical and Safety Bottlenecks Merging per-customer forks back into a shared model introduces severe safety and security risks that will likely delay public deployment, pushing my median expectation to June 1, 2031 . Pooling untrusted customer-derived updates is dangerous: data-poisoning research shows that roughly 250 poisoned documents can implant a backdoor regardless of model size, and narrow fine-tuning on insecure code has produced broadly misaligned behavior. Furthermore, "poisoning by dilution" demonstrates that hidden backdoors can be created when clean contrasting data is added. On the technical side, gradient updates are highly sample-inefficient, and model merging does not reliably prevent catastrophic forgetting. Revealingly, no top-five lab's safety framework currently defines an evaluation trigger for continuous weight updates.
Business and Privacy Disincentives Market trends are actively moving away from the prerequisites for this architecture. OpenAI is winding down its self-serve fine-tuning, with no new jobs permitted after January 2027, and Anthropic offers no fine-tuning API at all. Major providers exclude API and enterprise data from training by default to satisfy enterprise customers, who expect strict data separation. These customers would almost certainly reject architectures that blend their specific tuned weights into a shared model served to competitors.
Accelerants and Resolution Risks Despite these hurdles, AI-accelerated R&D could pull the timeline forward, supporting a 10% probability of a qualifying occurrence by January 1, 2028, and a 25% probability by June 1, 2029. Capability growth is rapid, with Anthropic's internal capability advancing at roughly 2x its earlier rate, and labs have demonstrated interest in learning from deployment data through initiatives like Google's API Contributor Tier. Furthermore, a lenient reading of the resolution criteria could trigger an early resolution. If a lab explicitly describes a federated-tuning pipeline for a smaller model (such as Gemini Nano) or deploys a pooled online-RL loop loosely framed as merging user weight updates, it might satisfy the literal text of the question.
The Long Right Tail The distribution extends deep into the 2030s because this specific architectural paradigm—forking and merging customer weights—may never be necessary. Labs might solve continual learning through entirely different methods, such as relying indefinitely on massive context windows, discrete centralized retraining phases When AI Systems Learn During Deployment, Our Safety ..., or external agentic memory files The components of agent memory. Even if true continual learning is eventually achieved, the industry may simply bypass the highly specific mechanism of merging individual user weight updates into a shared base. Together with the safety barriers to public deployment, this supports a 75% probability by January 1, 2034, and a 90% probability only by December 31, 2038.
Adjusted timelines outward to reflect the severe safety and poisoning risks that will likely delay any public deployment of architectures that merge untrusted customer updates into a shared base model .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited