Could OpenAI build the operating system like the one in ‘Her’?

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week here.

Will native-AI operating systems run our computers in the near future? 

Samantha, the AI that Theodore falls in love with in the 2013 movie Her is actually an OS. That’s how he meets her: He buys a new OS called OS1 (“it’s not just an operating system, it’s a consciousness”) and “she” is its persona. Samantha becomes his intuitive and personalized companion to all his digital stuff, from email to video games. 

ChatGPT set off the generative AI boom in part because consumers (not just academics and developers) saw shades of Her. Two years later, the chatbot is a major consumer destination and is the AI interface with which people are most comfortable and familiar. More than 400 million people around the world now use ChatGPT every week, OpenAI COO Brad Lightcap recently told CNBC. For comparison, Apple’s iOS operates on more than 2 billion active devices worldwide, including more than 1.5 billion iPhones. Those are impressive numbers, but iOS is standing relatively still while ChatGPT is steadily moving up the adoption curve.

Historically, Apple’s superpower has been its uncanny understanding of the user interface—the technology that acts as the mediator between a human user and their digital tools and content. Apple’s user interfaces rely mainly on older input technologies, such as mouse-clicking and touchscreen-tapping (though Siri remains a problem child). ChatGPT feels more and more like a new generation of user interface, one that’s powered by a very advanced mastery of language and the ability to intuit the intent behind a user’s input. Increasingly, it can act as a personal assistant with the ability to “reason” through problems. It can reason about things a user says with their voice or shows to the AI via their phone cam. Now OpenAI (and Anthropic) are rapidly developing the AI’s ability to control computer functions and apps. 

At the moment, OpenAI is packaging all this stuff inside ChatGPT. But how long before a chatbot app no longer contains it? Throughout the day, I call up the ChatGPT desktop app (by hitting <Option> and <Space Bar>) to run quick research lookups. For $200 per month, it can conduct deep research projects and control my mouse as it uses the web. It’s not a far leap from there to the AI using an app on my computer or printing something. Soon enough, I’ll likely have a running voice dialog with my AI OS throughout the day. Like Samantha (presumably) did for Theodore, the OS will learn my workflows, habits, and preferences. (In fact, many people in the field of natural language are working hard to build “EI,” or emotional intelligence into these systems, so that your AI operating system can be your friend too—as Sam was to Theodore.) 

The analyst Ben Thompson said in a recent Stratechery newsletter that OpenAI itself has realized that ChatGPT (and everything contained in it) is taking over the focus of the company. “Consumer tech companies . . . require a completely different culture and value chain than a research organization with an API on the side,” he wrote. “That is the fundamental reality that I suspect has driven much of the OpenAI upheaval over the last two-and-a-half years . . .” 

In short, OpenAI started out as an AI research lab, then became a provider of one-size-fits-all foundation models to enterprises and developers, then watched as ChatGPT stole the show and became its main source of revenue and fame. The question is, where does ChatGPT go from here? I think it’s the OS.

LLMs could help solve the glacial change of government systems

Jen Pahlka has a big idea: She says we should use large language models (LLMs) to eat through the government red tape that often stymies badly needed change to bureaucratic systems. She should know. She served as deputy chief technology officer (DCTO) under the Obama administration and helped found the United States Digital Service (USDS). Pahlka says that change agents like the USDS often encounter thousands of pages of policy and regulations that may or may not apply to a proposed change to a government website or system. 

Those pages pile up administration by administration,  often creating overlapping and conflicting rules or a tangle of ambiguities. And agency bureaucrats often use that ambiguity to put new proposals into a state of perpetual review, saying neither yes nor no, effectively freezing attempts at change that they don’t really want anyway. That’s why interagency organizations like USDS need lawyers as much as they need top-notch designers and coders.

LLMs could be another tool to cut through the ambiguity and shorten timelines. “You can use an LLM to figure out why there are 7,719 pages [of rules and regulations around a proposed system change] and what of that could get reduced,” Pahlka told Robert Safian in a recent Rapid Response podcast. “You can pretty quickly get to ‘this stuff is conflicting, this stuff is vestigial, this stuff is really controversial but you’ve got to deal with it’.”

Pahlka has done work with Stanford’s Regulation, Evaluation, and Governance Lab (RegLab), which is studying the development of LLMs that specialize in swimming through years of policies, guidelines, and regulations. “It’s still pretty early days, but it’s starting to get used that way,” Pahlka says. “I want to see somebody like the labor commissioner in a state saying, ‘Oh yeah, we actually now have a regulatory environment that allows us to serve the needs of our state.’”

Nvidia earnings: Good enough to keep the Boom booming

Nvidia’s fiscal fourth-quarter earnings report on Wednesday marked the company’s first since the Chinese company DeepSeek showed the world that it could train a world-class AI model with far fewer Nvidia GPUs than anyone thought possible. If any big AI companies ran to the phone to cancel their orders of Nvidia’s new Blackwell processors it wasn’t apparent in the numbers the company reported. Revenues came in at $39.3 billion (35.6% from data center sales), up 78% from a year ago and up 12% from the previous quarter. Analysts had expected $38.04 billion. 

The company also said the good times would continue at least through the current quarter, in which it expects $43 billion in revenues. Analysts on the earnings call did, however zero in on the one weak point in the report: Nvidia’s profit margin, which remained flat for the second quarter in a row. But  that metric is indicative of overall demand for GPUs only in an indirect way. Also, the company’s stock sagged somewhat in after-hours trading Wednesday, which is very dependent on the (mysterious) assumptions that institutional investors had already baked into the value of the stock. 

The bottom line is that the tech companies betting big on AI—think Google, Meta, OpenAI, and Microsoft—are not hedging on their infrastructure spending. Not yet, anyway. “Blackwell generating ‘billions in sales’ validates Nvidia’s top position just as the market is expanding on the agentic and physical AI fronts,” wrote eMarketer’s Jacob Bourne in a research note published Tuesday. “Short-term volatility is still on the horizon, but Nvidia’s market command remains unmatched.”

More AI coverage from Fast Company: 

• Anthropic’s new Claude AI model can decide between speed and deep thinking
• Autonomous AI agents are amazing—and scary
• People are most worried about AI replacing these 2 jobs
• This new AI tool helps Walmart’s merchandising team plan what’s in stores

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.

View the full article