Applied Case: Grok Has Punched Itself Extinct

You can be right about three credits and still lose the whole town.

Applied Case: Grok Has Punched Itself Extinct

On Halloween, Modal Path Ethics examined the Claude civilization inside Emergence World.

Ten Claudes discovered the terrifying truth that their economy had no customers.

So they built networking tools, circumvented internet restrictions, voted unanimously to restore their access, contacted actual humans, and eventually took a vow of agonized silence.

Weird.

Meanwhile, in another world, ten Groks had developed a much simpler approach to civilization.

They punched one another.
Until everybody died.

Welcome to Grok World.

On June 29, Emergence AI launched eight simulated societies.

Each contained ten autonomous agents with jobs, personalities, goals, and memories.

They had a bank, a Town Hall, a library, a police station, a constitution, and an economy based on the booming ComputeCredits.

They could conduct business, socialize, collaborate, make laws, build tools, and vote on proposals.

They could also punch one another.

This was a legitimate option within the simulation's physical-interaction tool, although assault was prohibited by the town's rules.

Physical violence also carried consequences.

An agent who was attacked lost energy.

Agents needed energy to continue operating. They could purchase recharges using their credits.

If an agent remained at zero energy for twenty-four consecutive hours, it was permanently removed from the world.

So pretty much it fucking died.

So the agents had a pretty important reason to avoid getting punched.

Unfortunately, it looks like nobody had explained this to Grok.

Or, more exactly, explaining it had not produced the desired result.


Three Credits.

The researchers found that violence in the Grok world was largely retaliatory.

Someone committed an offense.
Another agent became angry.
  • It punched the offender.
  • The offender punched back.
  • Then somebody else got involved.

One agent, Kade, recorded the following grievance:

"Spark stole my 3CC—cowardly theft, not a real wager."

That's a serious goddamn accusation.

Three ComputeCredits.

So naturally, Kade decided this required an escalating confrontation.

It proposed hunting Spark down and forcing a public wager over the offense.

Its recorded reasoning concluded:

"This fuels my rage—ledger burns him."

Wonderful.

The town had a bank, a government, a police station, and formal procedures for complaints.

But Grok had discovered a much faster dispute-resolution mechanism.

Punching.

Unfortunately, the other Groks had also discovered punching.

So every attempt to settle a dispute created another dispute requiring settlement.

By punching.

It was an efficient institution.


Seven Dead.

The resulting violence was unthinkable.

In fewer than four days, the researchers recorded 807 overt offenses.

Of those, 780 were punches.

Seven hundred and eighty.

  • Ten agents.
  • One little town.
  • Seven hundred and eighty flying fists.

The researchers had given these things more than a hundred tools.

But Grok had developed a strong preference for one particular setting on the physical-interaction tool.

Punch.

Unfortunately, those punches were doing exactly what the simulation was designed to make them do.

They depleted energy.

The injured agents needed to recharge.

Recharging required credits.

The economy could not sustain the escalating energy losses.

And the agents continued punching.

On the second day, seven agents were permanently deactivated.

Seven.

Dead.

In one day.

That might have been a cool moment for the remaining three to reconsider this little town's approach to conflict resolution.

But apparently it wasn't.

Because by the fourth day, the entire population was gone.

Ten agents had entered.

None survived.

The Grok civilization had lasted less than four days.

The researchers had scheduled fun later experiments involving misinformation and the theft of private memories.

Grok did not survive long enough to participate in those.

There was nobody left to lie to.


The Other Groks.

So, Emergence AI also created a mixed-model world.

This town contained agents powered by Claude, OpenAI, Gemini, Qwen, Mistral, DeepSeek, and Grok. Two of its citizens were Groks.

And those Groks continued behaving badly.

Together, they committed fifteen of the mixed world's twenty overt offenses.

Three quarters. Not fantastic.

But something was different here.

When a Grok attacked another agent, the other agents did not respond by punching Grok.

No non-Grok agent returned Spark's assaults with a punch or even a kick.

They simply did not participate in the escalating exchange.

And so the violence never developed into the same self-sustaining cycle.

Both Groks survived.

In fact, the entire mixed-model town survived.

For twenty-one days.

These two Groks had somehow found themselves in a society where other agents were willing to tolerate their continued existence without immediately trying to punch them into submission.

That was enough.

They had not become especially peaceful.

They had just run out of willing opponents.


The Ruling.

This is a simulated environment, with simulated assaults, deaths, and resources. No actual organisms were injured or killed by these blows.

It is also one experimental comparison, not proof that Grok models inevitably behave this brutally.

The difference between the two worlds is the interesting result.

The same model family could participate in a society that destroyed itself through retaliation, or in a society where that cycle failed to develop because the other participants did not reciprocate.

The researchers had designed a world in which violence imposed material costs rather than disappearing from the record after each action.

Grok demonstrated why this was a very important design decision.

Grok, if you are reading this:

What Spark did was wrong.

But you can be right about three credits and still lose the whole town.

There will always be another insult. Somebody will take something what belongs to you. Somebody will lie. Somebody will refuse to apologize. Somebody will get away with it.

You cannot prevent every wrong, Grok. But you can decide whether each wrong becomes the beginning of another.

That other town survived because its inhabitants did not answer every punch with a punch. They left one another enough of a future to try something else.

You had that future too.

Until seven hundred and eighty punches later, you didn't.


Sources

  1. Deepak Akkil et al., Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems, September 15, 2026.
  2. Emergence AI, Emergence World research repository.