TL;DR: Meta’s Muse, OpenAI’s Dots and Uber’s driver assistant share one guess: the agent speaks first. That strikes the arduous drawback from what to reply to when to interrupt, on which channel, and with what supply. Traditional ML and new choice fashions can clear up it.
Three launches, one sample
- Meta Muse (Sept 8). A private agent that books, emails and retains working with the app closed. It remembers particulars, makes unprompted ideas and checks in for approval, in its personal app and in WhatsApp.
- OpenAI Dots (Sept 29). At all times-on brokers that run “proactive analysis,” read-only monitoring of your apps, to catch a forgotten bill or a bug in Slack. They attain you in ChatGPT, Slack and Groups, with textual content and voice coming.
- Uber’s driver assistant. It turns dwell market indicators into recommendation. In a latest speak, Uber’s product staff described a driver idle for 33 minutes being pointed to a greater zone, with proof. Their rules: keep “at all times on the driving force’s facet,” floor alternatives proactively, and measure whether or not drivers acted. A hands-free voice model was introduced Sept 24.
From pull to push
Chatbots have been a pull interface: the person selected the second, the channel and the query. Proactive brokers invert that. Interrupt too typically and customers mute you. Interrupt too late and the surge has ended or the bill is overdue. Choose the improper channel and message fails. The LLM can write the message; it’s the improper device to resolve whether or not to ship it.
The rule: worth should beat interruption price
Each proactive message is a guess. Ship solely when its anticipated worth to the person exceeds the price of interrupting them. Worth has 4 elements: how a lot is at stake, how doubtless the person is to behave, how briskly the chance expires, and whose worth it’s, the person’s or the platform’s.
Worth then units each choices. Excessive-value, expiring messages go now; modest ones anticipate a digest; the remaining are by no means despatched. And the extra a message is value, the extra intrusive a channel it has earned.
“When” and “how” are prediction issues
Notification and development groups have solved variations of this for years:
- Uplift fashions estimate whether or not a nudge causes the motion.
- Contextual bandits study the very best second and channel for every person.
- Receptivity fashions predict whether or not the person can interact proper now: driving or parked, in a gathering or free.
Channels have a worth of admission. An in-app card prices little, chat prices extra, SMS extra nonetheless, and a voice name is reserved for pressing, hands-busy moments. Uber’s final result monitoring provides the labels these fashions want.
The place choice fashions match
A brand new mannequin class fits this job. Choice fashions equivalent to TypeSafe’s Jev and the open-weights Julia 1 from Supersonic Labs return no textual content, solely a typed judgment: a selection from a brief listing, a rating, or a sure/no, every with a chance. Julia 1 runs on a CPU, from about 33 ms per choice.
An always-on agent evaluates way more triggers than it sends, so every candidate can get low-cost, structured questions earlier than any LLM name: Is that this value an interruption? How beneficial, how pressing? Ship now, wait, batch or drop? Card, chat, SMS or voice?
The bounds matter. These fashions decide solely the context they’re given, so they won’t study that one person ignores mornings;
Cross-sell, fastidiously.
An agent that speaks first is a strong distribution channel. Meta says it’s exploring commerce in Muse and has already launched Muse for Small Enterprise; OpenAI introduced Dots alongside a brand new $500 month-to-month tier; Uber has supply and different providers to advertise. However a proposal is simply one other message. It should clear the identical bar, measured on worth to the person. As soon as customers suspect the agent is promoting as a substitute of serving, each message loses credibility.
The takeaway
All three firms have robust fashions. The race might be received on judgment about consideration: talking not often, on the proper second, in the suitable place, on the person’s facet. Corporations with years of notification information might have a head begin.
