|
Jev is built to answer exactly those kinds of questions. Give it a support ticket, a review or a document, and it returns a category, a rating or an estimate of how likely something is true. Its unusual feature is that it cannot write a sentence: the output
is a decision software can use directly. The model comes from TypeSafe AI, whose founder, Diogo Almeida, helped develop the research behind ChatGPT at OpenAI.
In a customer service queue, that could mean sending billing questions to finance, flagging complaints that need a person, or grouping reviews by the problem customers describe. Your team defines the categories and what belongs in each. Jev then applies those
instructions to the incoming work.
Each of those choices may take a colleague only a moment. Across a busy team, though, the reading, sorting and forwarding add up. Decision models are designed to handle those routine choices quickly and cheaply. Their value depends on how much work they save
while making decisions your team can trust.
Databricks sees a similar opportunity in the information companies already hold on its platform. About two weeks after Jev launched, it announced ai_decide, giving data teams a way to apply decision models to that information. Databricks points to Jev as an
example of this approach and says ai_decide is compatible with TypeSafe’s API. That gives your existing data team a way to explore the same kind of decisions within Databricks.
To show the idea in action, Databricks had ai_decide play Snake. At each turn, it looked at the board and chose the next direction. The same pattern carries over to everyday work: look at what has arrived, decide what should happen next, and pass that decision
to the software handling it.
Of course, a wrong turn in Snake ends a game. A wrong decision at work could delay a customer or send a document to the wrong person. Jev gives no written explanation, which can make mistakes harder to understand, and another AI can miss the same error. With
ai_decide still in Beta, Databricks’ early testing stage, it makes sense to start with a task your team knows well and can easily check.
Our take: a useful first win would be a queue that takes less effort to manage, with routine requests reaching the right team and difficult cases reaching a person sooner.
Let that be the measure of success: does it make the work easier without creating more mistakes to clean up? Our
AI Functions page follows what Databricks adds next.
|
