XNews

Databricks Observes GPT-6 Astra Benefits Concentrated on Hard Tasks

Databricks Observes GPT-6 Astra Benefits Concentrated on Hard Tasks

After deploying GPT-6 Astra to all employees, Databricks found that the model's gains are concentrated on complex tasks, while simple tasks show no improvement. Overall coding expenditure rose 60% due to increased usage on hard tasks. Databricks used a gateway with queue experiments for rollout and implemented budget routing to steer model selection.

Source posts