rationale: Serving 1,000+ distinct weight variants of a single flagship model that diverge based on accumulated customer learning faces immense structural hurdles. While multi-adapter serving techniques make hosting thousands of LoRA adapters technically cheap, the true barriers are privacy and safety. Enterprise contracts default to zero-data-retention, and updating deployed weights creates unresolved data poisoning and sycophancy risks. Even in an aggressive scenario where continual learning takes root, labs are more likely to update a single shared model than to fork thousands of distinct weight variants, because individual user sessions generate too little signal to meaningfully update a frontier model on their own. Given that we expect per-customer weight forks to represent only a 4% share of commercial practice by mid-2027 and the correspondingly low baseline probability of per-customer weight customization dominating commercial practice, reaching the 1,000-variant threshold for a flagship model by mid-2028 holds a 7% probability.
futuresearch