Today, artificial intelligence startups frequently boast that their products will revolutionize industries by automating white-collar work. Yet, few mention the various setbacks these companies encounter when developing such products for enterprise clients, or the awkward situations that arise when implementing their own technology internally.
This is precisely why a recent podcast episode garnered significant attention—its guest was an xAI engineer who departed the company shortly after the episode aired.
Suleyman Gory, a former technician at xAI, revealed on an episode of the "Never Give Up" podcast that xAI has been developing a human-like AI program called the "Human Simulator," whose behavior patterns are indistinguishable from humans. He stated that xAI has been internally testing the "Human Simulator" and treating it as a formal employee. In some instances, this approach has even caused confusion among certain xAI staff, who were unaware they were interacting with artificial intelligence.
"I've received messages several times asking, 'The org chart shows this person reports to you, are they not in today?' But in reality, that 'person' was just an AI," Gory said. These virtual employees sometimes also experience "hallucination phenomena." For example, a virtual employee might notify a real employee to meet at its workstation, only for that employee to arrive and find an empty desk. (If this reminds you of the AI assistant "Son of Anton" developed by Gilfoyle in the TV series "Silicon Valley," you are certainly not alone.)
At xAI, Gory was part of the "Macrohard" team. xAI CEO Elon Musk positioned this team as an "AI software company" capable of autonomously completing software development work. The name "Macrohard" plays on the Microsoft brand name and highlights Musk's ambition—to create AI tools that can easily replace traditional enterprise software like Microsoft Teams and Word. Musk has explicitly stated that he hopes this project will directly compete with giants like Microsoft for enterprise software contracts.
However, based on current circumstances, the advancement of the "Macrohard" project is highly reliant on manual operations and requires customized adjustments for different tasks and client needs.
Gory's statements align with what I have previously reported—xAI is attempting to use AI to replace some internal staff, including certain members of the trust and safety team at the social platform X. Furthermore, other AI labs like Anthropic and OpenAI are also intensifying efforts to develop similar products, striving to automate workflows through AI technology and thereby replace traditional software.
Gory also detailed the various challenges xAI faced while developing the "Human Simulator" for enterprise clients. He mentioned that xAI employees would first interview clients, observe their workflows, and then build custom virtual employees based on that, but this process often overlooked critical workflow steps.
"When we reviewed the virtual employee's work steps, we found it would consistently fail in certain specific scenarios," Gory said. "Later, after observing human employees perform the same task, we realized the process was missing a full 20-plus steps; all these steps had been omitted. When we asked the clients, they suddenly realized, 'Oh right, we do handle it that way; I forgot to tell you, my mistake.' This kind of situation is incredibly common."
Gory also revealed that xAI plans to significantly expand the number of "Human Simulators," potentially scaling up to 1 million simulators running simultaneously in the future to enhance AI's capability to perform white-collar tasks. He stated that to achieve this goal, the company is considering leasing idle computing power from Tesla owners—specifically, utilizing the computational resources available while the vehicles are charging, an idea consistent with a concept Musk has publicly proposed before.
The entire podcast episode is worth listening to, if only for Gory's insights into daily life at xAI. Reportedly, working at xAI requires being constantly responsive to Musk's product demands, the inspiration for which often stems from content he encounters while browsing the social platform X.
Perhaps it's no coincidence that shortly after this podcast episode was released, Gory announced his departure from xAI.
Google Successfully Knocks on the Enterprise Market's Door
Google has not traditionally been considered a giant in the enterprise software arena, but the strong performance of its Gemini AI model has the potential to change this.
According to internal data Irene disclosed over the weekend, the costs software developers have invested in using Google's Gemini model have been steadily increasing over the past year, and the number of API calls for Gemini nearly doubled between March and August of this year.
This trend is undoubtedly a positive signal for Google's cloud business. For a long time, Google Cloud has consistently ranked third in the global cloud services market, trailing behind Amazon Web Services and Microsoft Azure; recently, however, Google Cloud has been enhancing its competitiveness by attracting AI developers. It's worth noting that the Gemini model is exclusively provided by Google Cloud, so usage of this model will directly boost Google Cloud's business growth.
Despite Gemini's growing popularity, this advantage has not yet fully translated into sales performance for AI software products, such as Gemini Enterprise—an application that allows business clients to search company-wide data via the Gemini chatbot and personalize AI agents. As related reports have pointed out, while Gemini Enterprise has shown some growth momentum, many customers feedback that the software's actual customization results are often unsatisfactory.
Nonetheless, the strong performance of the Gemini model among the developer community is a relief for Google. After all, Google has invested over $8 billion in AI startup Anthropic, which is now cooperating increasingly closely with Google's main competitor, Microsoft.
