Meta, OpenAI and Uber Just Taught AI Agents to Talk First

Traditional chatbots operated on a pull interface where users dictated the timing, channel, and questions. Today, Meta, OpenAI, and Uber are driving a shift toward proactive AI agents that break this paradigm. Interrupting users too frequently results in muted notifications, while speaking too late means missing critical opportunities entirely.

Sending a proactive message functions much like a wager. A system transmits a notification strictly when the expected value delivered to the user outweighs the cost of the interruption. This overarching value relies on four components: the stakes involved, the user's likelihood to act, how rapidly the opportunity expires, and whether the primary beneficiary is the user or the platform.

Once value is established, it dictates two primary decisions. High-value, expiring messages deploy immediately, moderate ones await a periodic digest, and the rest drop out entirely. Furthermore, message worth dictates the intrusiveness of the communication channel utilized.

• In-app cards incur minimal friction.
• Chat messages demand higher investment.
• SMS notifications require even more consideration.
• Voice calls remain strictly reserved for urgent, high-priority moments.

Stock photo for illustration only, not from the actual event

Stock photo for illustration only, not from the actual event

Addressing these routing challenges has introduced a novel class of decision models, including TypeSafe's Jev and Supersonic Labs' open-weights Julia 1. These systems bypass text generation to return typed judgments—such as selection from a concise list, a scored evaluation, or a binary yes/no accompanied by probabilities. Notably, Julia 1 operates directly on a CPU, achieving decisions in approximately 33 milliseconds.