The race for a digital employee

THE RACE to build software that operates a computer the way a human does — clicking, typing, navigating between applications — accelerated sharply this week when two very different entrants staked their claims within hours of each other. On March 11th Elon Musk unveiled Digital Optimus, a joint xAI-Tesla project cheekily codenamed Macrohard, that pairs the Grok large language model with a Tesla-built agent capable of processing a continuous five-second video stream of a user's screen in real time. Hours later, at its inaugural developer conference in a converted San Francisco church, Perplexity announced Personal Computer, software that turns a spare Mac mini into an always-on, locally controlled AI agent with persistent access to files, apps and sessions.

Both systems are chasing the same prize: an AI that does not merely answer questions but performs the work, navigating graphical interfaces the way a human employee would. The concept, loosely called "computer use," has become the defining battleground in agentic AI. Anthropic's Claude Cowork, which launched in January on macOS and expanded to Windows in February, has already rattled Wall Street — triggering roughly $285 billion in enterprise-software market-capitalization losses as investors repriced companies whose core offerings overlapped with what a desktop agent could automate. Microsoft responded by building its own Copilot Cowork on top of Anthropic's Claude, bundled into a new $99-per-user E7 tier. And OpenClaw, the viral open-source agent that surpassed 250,000 GitHub stars in 90 days, demonstrated just how ravenous the demand is — even as security researchers documented over 135,000 publicly exposed instances, a poisoned skills marketplace and a critical one-click remote-takeover vulnerability.

Ctrl-Alt-Compete

Yet the real story is not convergence on a concept; it is the wildly divergent theories of how to deliver it. Musk is betting on edge computing and proprietary silicon. Digital Optimus would run primarily on Tesla's in-house AI4 chip, priced at around $650, using xAI's Nvidia-based cloud hardware only when heavier reasoning is required. He frames the architecture through Daniel Kahneman's dual-process model: Tesla's component as System 1 (fast, instinctive screen interactions) and Grok as System 2 (deliberative reasoning). It is a clever pitch, albeit one that remains entirely undemonstrated — no live demo accompanied the announcement, and no audited corporate filing names Macrohard as a formal project. Musk reckons the system could be available in about six months.