OpenAI launched dots in September as an assistant that keeps working when the user isn’t giving it instructions. Available in the ChatGPT desktop and mobile apps for subscribers to the Pro plan or higher —with plans starting at $100 a month—, it can connect to calendars, email, computers, and other tools.
Alexander Embiricos, head of OpenAI’s enterprise product team, explained at TechCrunch Disrupt that dots didn’t emerge from completely new technology: much of what it does was already possible with Codex, the company’s coding tool. The difference, he said, lies in reducing the work and oversight required to use those features.
“A lot of that isn’t building some new piece of tech into the product; it’s how you frame it to the user and how it feels.”
— Alexander Embiricos, head of OpenAI’s enterprise product team
From coding tools to a persistent agent
The evolution involves packaging OpenAI’s model capabilities into a simpler experience: from coding agents for software engineers to ChatGPT Work and, later, dots. The idea is for the agent to maintain an ongoing conversation and presence, rather than relying on a series of isolated queries.
This approach also means the system uses more tokens. For users, the promise is to get help before they ask for it: for example, alerts about scheduling conflicts or upcoming deadlines. However, the alerts don’t always arrive in time. Embiricos recalled that, at an early stage, his own dot flagged a calendar conflict just three minutes beforehand.
In another case he recounted, an engineer at a drone delivery startup said the agent had detected a service outage, analyzed it, and offered useful information to resolve it before he noticed. Embiricos expects the alerts to become more timely; he considers the assistant’s initiative central to the product.
Data access, permissions, and next steps
To anticipate needs, dots can use information it knows about the user, their conversations, and the resources they have granted it access to. Embiricos described a feature called proactive research: the agent periodically reviews this data, looks for ways to help, and proposes actions.
Privacy controls are the same as those in ChatGPT: users can disable the use of conversations to train models and control which computer resources are available. But they don’t offer the level of detail Embiricos considers necessary. Permissions also create friction: the app rejects read-only access to cloud drives. “What I think we need to do is basically build out super, super, super granular controls”, he said.
OpenAI had opted for conservative permissions and numerous authorization requests when introducing the product, a decision Embiricos said the company expected to adjust over time. The tension remains: the assistant needs access to personal information to anticipate needs, but that access expands the data OpenAI handles.
Expansion into external actions is also being discussed. Embiricos said OpenAI is in talks with payment processors, though he did not provide further details. Allowing an assistant to spend money through partnerships with payment companies could be potentially lucrative for OpenAI, but would also place significant demands on consumers.
Source: TechCrunch
