Always-on AI personal assistants that keep running in the background even when you don’t ask for them are this year’s answer to the question of how people actually use deep learning models. The startup Instinct became an overnight sensation, and Meta devoted all its marketing efforts to its assistant, Muse.
I didn’t mean to ignore OpenAI. The company released its assistant product, dots, at a developer day in late September. Alexander Embiricos, who leads OpenAI’s enterprise products team, joined this month’s TechCrunch Disrupt to explain the origins of the product. I started by wondering why so many companies seem to be turning to cute mascots as avatars for their multi-trillion parameter matrix multiplication machines.
“There’s a kind of convergent evolution happening,” Embirikos said of efforts to develop AI assistants. “As we were working on the product, we realized that we could frame this as one entity and one conversation instead of different conversations. … It started to become intuitive.”
The assistant is available on the ChatGPT desktop and mobile apps for Pro-level subscribers and above, with plans starting at $100 per month. Once agents are connected to your calendar, email, computer, and other tools, they’ll try to help you get your work done.
Embricos points out that many of the capabilities of dots can be achieved using the Codex app, Open AI’s coding tool released earlier this year. But it requires more work and manual labor. He recalled celebrating when his wife started turning on her computer all night so agents could continue working.
“What I’m thinking about on the product team is how can we make this intuitive and useful for people,” he told Disrupt attendees. “A lot of it isn’t about incorporating new technology into the product. It’s about how the product is assembled to the user and how it feels.”
Making the technology feel intuitive meant repackaging the functionality of the company’s model into something more accessible, a path it was already on when it spun out ChatGPT Work from an agent coding tool it built for software engineers. Even the name, a play on the negative space in OpenAI’s logo, is meant to “make it feel like something that could eventually become ubiquitous.” (When asked when he learned that rival AI entrepreneur Elon Musk owned the dot.com URL, Embiricos said, “I found out very late.”)
The idea of persistent agents is that they don’t require prompting. For OpenAI, that means more token usage, and for users, as Embiricos puts it, a more proactive assistant “thinks about you and thinks about how to help you.”
In our tests, the assistant’s ability to think ahead is clear and often useful, but it’s still not enough. It flags calendar conflicts and upcoming deadlines, but doesn’t always notify you in the most timely way.
“The first time it informed me about one of those (collisions), this was like early in development, it informed me three minutes before that,” Embiricos says of his dots. “It’s like, ‘Hey, this is going to happen. Maybe I should do something about it.'” You only have three minutes! ”
Embiricos expects the timing to improve over time, but a proactive approach is key for this product. “I was listening to an engineer at a drone delivery startup say, “We’ve had this failure. By the time I realized it was happening, my people were already aware of the failure, were already analyzing the failure, and were coming to me with useful information on how to resolve the failure.”
give key to dot
As with most AI agent use cases, there is a tension between how much personal information an AI model needs to be useful and how much personal data it exposes to software companies. Muse has been criticized for creating profiles of family and friends. The software could be used, for example, to figure out what kind of restaurant you want to go to with a particular friend, but it could also bring incredible value to Meta’s advertising business. Does Dot collect similar data?
“Yes, we have something built into the dots called proactive research,” Embiricos said. “With a certain cadence, it’ll think about what it knows about you, what conversations it’s had, and what you’ve given it access to think about. Then it’ll go to research and think, ‘What can I do to help you more?'” and basically make recommendations. ”
It’s difficult to differentiate between the agent’s data needs and the parent company’s data needs. Hark, a startup that is developing its own AI personal assistant, clearly points out that it is not collecting data to power its advertising business like OpenAI and Meta.
With Dot, OpenAI says its agents have the same privacy controls as ChatGPT itself. Users can turn off model training (stopping OpenAI from using conversations to improve the AI) and control which resources in the computer are accessed and how. Still, these controls aren’t very granular and the app can be pushy. For example, deny read-only access to cloud drives.
“What I think we need to do is basically build in very, very, very fine-grained control,” Embiricos said, though he pointed to the model’s intelligence and discretion as the flip side of this coin.
“This is actually something that we were somewhat frustrated with when we launched,” Embiricos said. “We had very conservative default settings for what the agent would do, and we were asking for permission over and over again. We knew we would see that feedback, but I remember thinking, ‘This is fine, we can adjust it over time.'”
Can personal AI agents scale?
Another big question about AI agents is whether they are scalable. It’s fun when only a few people have them, and stories abound of agents snatching up hard-to-get reservations. However, businesses don’t necessarily like or may not be ready to interact with agents. People definitely won’t like to interact with them (note: please stop pitching me to AI agents), and Embiricos is one of those people. OpenAI’s corporate partners also hire human assistants because they want human interaction, he said, adding, “Personally, I would never meet with someone else’s agents.”
If the AI industry is right that agents are our future, can we get there without overtaxing the infrastructure the world currently runs on, from businesses and websites to payment networks?
“At OpenAI… we are not always moving the fastest in terms of delivering the next capabilities that allow agents to interact with the world,” he said. “In doing that, we’re working very closely with other parts of the ecosystem to do that. I can’t say too much here, but we’re having a lot of conversations with payment processors.”
Partnering with payment companies to let AI assistants spend money could be a first for OpenAI, but it’s also a big ask for consumers. Giving my assistant access to my email inbox already feels like a big ask, but my bank account is off limits. For now.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
