Transforming OpenAI’s Coding Tools: The Evolution into Dots, Your Ever-Present Assistant

This year marks a pivotal moment for AI personal assistants, especially those that operate continuously in the background, providing insights and assistance without constant prompting. The startup Instinct has quickly garnered attention, while Meta has heavily promoted its assistant, Muse.

OpenAI has joined the fray with the launch of its assistant, dots, which debuted during its developer day in September. Alexander Embiricos, who heads OpenAI’s enterprise product team, shared insights at TechCrunch Disrupt about the impetus for creating this product, particularly why many tech firms are favoring whimsical mascots to represent their advanced AI systems.

“There seems to be a natural evolution happening,” said Embiricos regarding the development of AI assistants. “In our product design, we recognized the potential to frame it as a singular entity engaged in one ongoing conversation, rather than scattered dialogues. This approach makes it feel more natural.”

Available through the ChatGPT desktop and mobile applications for Pro subscribers starting at $100 monthly, the assistant connects to your calendar, email, and other applications to assist you in completing tasks.

Embiricos noted that while similar functions could be performed using OpenAI’s Codex coding tool introduced earlier this year, it requires more user engagement. He recounted celebrating when his wife began leaving her computer on overnight, allowing the assistant to operate effectively.

“My focus on the product team is to ensure intuitive usability for everyone,” he told the audience at Disrupt. “Much of that doesn’t involve new technology; rather, it’s about how we present it to the user.”

To achieve an intuitive experience, OpenAI aimed to reframe its models’ capabilities into a user-friendly format. This initiative traces back to their development of ChatGPT Work from coding tools initially designed for software developers. The assistant’s name even draws on the subtle elements of OpenAI’s logo, reflecting a hope that it will become a common tool in daily life. Embiricos revealed he only recently discovered that Elon Musk owns the dot.com domain name.

The concept behind a persistent agent is that it operates without needing prompts. For OpenAI, this translates to increased token consumption, while users benefit from a proactive assistant that is constantly considering how to assist them more effectively, as Embiricos described.

During testing, the assistant demonstrated its forward-thinking capabilities, although there is still room for improvement. It alerts users about calendar conflicts and approaching deadlines, though not always promptly.

“In the early development phase, it once alerted me about a conflict only three minutes prior,” Embiricos said. “It was like, ‘Hey, this is coming up. You may want to act on this.’ But three minutes isn’t much time!”

He anticipates enhancements in timing but emphasizes the importance of the proactive nature of the product. “I heard from an engineer at a drone delivery service who mentioned that while he was oblivious to an outage, his assistant had already detected and analyzed the situation, providing him with constructive insights on how to address it.”

Empowering the Assistant

A key concern with AI agents is the balance between the personal information they require to function effectively and users’ privacy. Muse, for instance, has faced scrutiny for its ability to compile extensive profiles on users’ acquaintances, potentially benefiting Meta’s advertising initiatives. Does OpenAI’s dots gather similar information?

“Indeed, we’ve integrated what we call proactive research into dots,” Embiricos explained. “At regular intervals, it evaluates what it knows about you and the conversations it has had. Based on this, it will conduct research on how it can offer more assistance and make pertinent suggestions.”

Distinguishing the data needs of the assistant from those of OpenAI itself is complicated. Hark, another startup crafting its own AI assistant, explicitly states it does not collect information for advertising purposes like OpenAI and Meta.

According to OpenAI, dots ensures the same privacy protocols as ChatGPT: users can disable model training to stop the AI from learning from their interactions, and they can manage what resources the assistant accesses. However, these controls can lack granularity, and users may feel pressured; for instance, it does not allow read-only access to cloud storage.

“We recognize the need for more detailed controls,” Embiricos noted while also emphasizing the importance of the assistant’s intelligence and discretion.

“We encountered some criticism upon launch for having conservative defaults that often required user permission,” he recalled. “While we anticipated this feedback, we believed it could be adjusted over time.”

Can AI Assistants Become Widespread?

Another significant question surrounding AI agents is their scalability. While they are engaging when only a handful of people utilize them, businesses might find them challenging to manage. Many individuals also prefer direct human interaction rather than automated services. Embiricos, for example, maintains a human assistant since OpenAI’s partners typically want to connect with people rather than AI agents.

As the industry asserts that AI agents represent the future, concerns arise about whether we can integrate them into existing infrastructures, including businesses and payment systems, without overwhelming the system.

“At OpenAI, we might not always be the fastest to deploy new functionalities for the assistants,” he mentioned. “However, we collaborate closely with the broader ecosystem. While I can’t disclose details, we are actively in discussions with various payment processors.”

Enabling an AI assistant to conduct financial transactions through partnerships with payment systems could be a significant step for OpenAI, but it poses a considerable ask for users. Many already feel uneasy granting an assistant access to their email, let alone to sensitive bank information.