Tesla's Digital Optimus Can Play 'About Halfway' Through Diablo, Says the Person Who Would Know, in a Reply on X

Elon Musk said on X on Saturday that Digital Optimus, Tesla’s computer-using AI agent, can now “play about halfway through the Diablo campaign so far just by looking at the screen like a human.” It is also, he said, good at Counter-Strike and other fast games, and League is in training. The goal, he wrote, is to generalize across all games.

Halfway through Diablo. A series named after the Lord of Terror, which is also how some of us describe a Monday.

First, the precise part. The statement came as a reply to a post about a Tesla job listing, per Drive Tesla Canada and Crypto Briefing. The listing is for a machine learning engineer in Palo Alto on the agent’s “Gaming Track,” and the stated work is improving the agent’s memory, planning and real-time decision-making, plus testing whether skills learned in games carry over to other software.

Second, the other precise part. Per TeslaNorth, Musk did not share video, a release date or scores, and did not name which Diablo or which League title. Neither Tesla nor its partner company has published details about the gaming results. Crypto Briefing adds that transfer of gaming skills to non-gaming software has not been demonstrated yet, and that generalization remains a challenge.

So the claim stands as a claim. It might be accurate. It might be a boss-fight screenshot from a very good Saturday. We report it with the confidence of a person who has been told their friend “is basically at endgame.”

Here is what the system is, as described. Digital Optimus was unveiled in March as part of the Macrohard project, a collaboration of Tesla and xAI. It watches the last five seconds of real-time screen video and responds with keyboard and mouse actions. Unlike systems that read a game’s internal data, per Drive Tesla Canada, it has to interpret what is on screen, as a person would. Musk described a two-part setup: Digital Optimus handles fast reactions while Grok acts as what he called the “master conductor/navigator” that decides what to do next. He said it would run mostly on Tesla’s AI4 computer, which TeslaNorth puts at about $650.

Let us admire the org chart. A fast part that reacts, a slow part that decides. Reflex and judgment, split into two systems and rejoined by a messaging layer. It is the architecture of a competent person who has been reading the comments section: react immediately, think later.

Why games? The job listing’s answer, as Drive Tesla Canada reports it, is that Tesla wants to integrate these agents into Full Self-Driving and the Optimus robot. Games are a place to practice perception, planning and quick decisions with a score attached and a very forgiving reset button. A dungeon crawler is a surprisingly efficient way to learn that a thing coming at you is probably a problem.

There is a caveat the hiring post itself implies: the skills have to transfer. Learning to click on skeletons does not automatically produce a better spreadsheet operator, or a better driver. (Though one imagines that if the agent encounters a spreadsheet, it will treat the cells as loot.)

Musk has also said, per Drive Tesla Canada, that Digital Optimus could eventually run on AI4 hardware in Tesla vehicles, allowing parked cars to do computer-based work. This would make the parked car the first vehicle in history to have a side hustle. It would also raise the question of whether, if the car is playing Diablo to improve its generalization, we should call that “work” or “personal time.” Labor law has not caught up.

It is worth noting the unusual standard of evidence in this genre. When a lab reports a benchmark, there is usually a paper, a leaderboard, a footnote that says vendor-reported. Here we have a sentence, a hiring post, and a halfway point. Even the halfway point is soft: the story of Diablo is told in acts, and “halfway” could mean a boss, a town, or the point at which the inventory fills up with unidentified rings.

The “just by looking at the screen like a human” part is, however, genuinely the interesting claim, if it holds. Systems that need privileged access to a game’s internals are useful for research but tell you little about whether an agent can handle software it has never seen. Pixels in, keyboard and mouse out, per Crypto Briefing, is the harder and more general interface, and it’s also the interface of every program ever written for people. If an agent masters that, the ceiling is any screen.

The floor is a necromancer who has just died, again, to something with a lightning aura.

For now, the facts are: a Saturday post, a job listing, a half-finished campaign, and a goal of all games. The independent verification is pending, and the League of Legends training has no finish line. The agent, in the meantime, is presumably still working on its dodge timing.

Halfway through Diablo. Come back when it reaches the end, and then, as is traditional, play it again on a harder difficulty.

Sources: