|
Welcome back! AI startups talk a lot these days about how their products will be revolutionary by automating white collar work. But it’s rarer to hear messy stories about the hiccups AI startups face when trying to build such products for businesses, or the awkward situations that can arise when they roll out their own technology internally. That makes a recent podcast with an xAI engineer—who left the company shortly after it was published—especially notable. Sulaiman Ghori, who was until recently a member of the technical staff at xAI, said on an episode of the Relentless podcast that the company has been developing AI employees that act like humans, known as “human emulators.” xAI has been testing human emulators internally and treating them as regular employees, he said. In some instances, that’s led to confusion among xAI staff who don’t realize they’re working with AI. “Multiple times I’ve gotten a ping saying like, ‘Hey this guy on the org chart reports to you. Is he not in today or something?’ And it’s an AI,” Ghori said. The virtual employees sometimes hallucinate. For example, Ghori said, a virtual employee might tell a real employee to come over to its desk to talk—and the real employee will do so, only to find an empty desk, he said. (If this reminds you of Gilfoyle’s AI worker Son of Anton from the TV series “Silicon Valley,” you aren’t alone.) At xAI, Ghori has been working as part of the Macrohard team, which xAI CEO Elon Musk has described as an “AI software company” that can autonomously build software. The name, a play on Microsoft’s name, is part of Musk’s ambition to build AI tools that can easily replace older versions of enterprise software, such as Microsoft Teams or Word. Musk has said he wants the venture to compete directly for enterprise software contracts with the likes of Microsoft. So far, though, building Macrohard sounds like a very manual process tailored to individual tasks and customers. Ghori’s comments are in line with what I’ve reported about xAI trying to replace its own staff with AI, including some trust-and-safety staff at X. Other AI labs, like Anthropic and OpenAI, have similarly been working on products that use AI to automate work and replace older versions of software. Ghori also detailed some of the hiccups xAI has faced while trying to build human emulators for enterprise customers. xAI staff interview customers and watch them work before building virtual employees for them, but it’s easy to overlook important tasks when doing so, Ghori said. “We’ll look at the steps the virtual employee is making and realize, well it’s always making mistakes in these places in these specific cases,” Ghori said. “And we go watch the human doing the same thing and there’s like 20 different steps that are missing that they just totally left out and we go to them and they’re like, ‘Oh yeah, we do that, I forgot to tell you. My bad.’ It happens all the time.” Ghori also said xAI plans to dramatically increase the number of human emulators, with as many as one million running at once, to improve how well the AI can complete white collar tasks. To do so, he said the company is considering leasing computing power from Tesla owners whose cars are charging, echoing claims Musk himself has made. The full podcast is worth a listen, if only for Ghori’s anecdotes about what it’s like to work at xAI, which apparently involves scrambling to respond to product requests from Musk based on what he sees on his X feed. Perhaps not coincidentally, Ghori announced he was leaving xAI just after the podcast was released. Google Gets a Foot in the Door with Enterprises Google has never been known as an enterprise software heavyweight, but the success of its Gemini AI models could change that. Specifically, software developers have been forking over more cash for Google’s Gemini models over the past year, with requests sent to the Gemini application programming interface roughly doubling between March and August, according to internal data Erin published over the weekend. That’s a good omen for Google’s cloud business, which has always held a distant third place in the cloud market behind Amazon and Microsoft but has more recently been wooing AI developers to its platform. Google’s Gemini models aren’t available on other cloud providers, so their use directly boosts Google Cloud. While its models are growing in popularity, that hasn’t necessarily translated into success in selling AI software products such as Gemini Enterprise, an application that lets corporate customers use a Gemini chatbot to search across company data and configure AI agents. As our story notes, Gemini Enterprise has shown some signs of growth, but buyers still complain that customizing the software for doesn’t always work well in practice. Nevertheless, Gemini’s strength among developers is likely a relief for Google given that Anthropic, the AI startup it has backed with more than $8 billion, has been getting increasingly close with Microsoft, a Google archrival.
Thank you for reading the Applied AI newsletter! I’d love your feedback, ideas and tips: aaron@theinformation.com. If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.
|
