Meta’s mixed signals to staff on tokenmaxxing [Report]

Imagine a job where you shoot at targets all day, and someone tells you your worth to the company depends on how many bullets you fire. So you fire like crazy. Then the ammo runs short and a memo lands saying nobody expects you to empty the magazine anymore. And right around then, someone hands you the minigun from “Terminator 2: Judgement Day.”

Something roughly like that is happening with AI at Meta, according to Wired.

The scoreboard is getting switched off

The Information reported Wednesday that Meta engineers had been told “AI adoption dashboards” and “token counts” would no longer factor into evaluations.

Wired, working from anonymous Meta insiders, reports that newly published performance review guidance for employees drops “usage of AI” as an assessment criterion. The guidance apparently describes desired outcomes that “can be supported by AI or other means.” Several of Wired’s sources inside Meta said they’re relieved to no longer feel pressure to use AI constantly.

That’s a real reversal, and it didn’t come out of nowhere.

How tokenmaxxing became a competition in the first place

A March column by Kevin Roose in the New York Times introduced most people to the idea that certain AI-happy workplaces, Meta and OpenAI among them, were pushing staff to maximize AI usage. Hence “tokenmaxxing.” Roose pointed to OpenClaw, the token-chugging agentic AI platform that was popular at the time, as a possible driver.

Roose said Meta was “rewarding workers who make heavy use of A.I. tools and chastening those who don’t.” The report also mentioned leaderboards, which turned AI profligacy into a game with a scoreboard.

A few months later, in June, the Financial Times reported that Google was throttling Meta’s AI tokens, which forced the company to rethink tokenmaxxing to some degree. Turns out the supply of bullets was never unlimited.

The minigun is called Hatch

While all this was unfolding, Meta has been trying to build a competitive AI tool of its own.

Back in 2024, Mark Zuckerberg said, “This is going to be the year when a highly intelligent and personalized AI assistant reaches more than 1 billion people, and I expect Meta AI to be that leading AI assistant.” It isn’t clear that ever happened with Meta’s ChatGPT-style app.

I’ve used the Meta AI web app exactly once, for a story. Amanda Siberling of TechCrunch apparently also used Meta AI for a story, and had a worse time of it, because Instagram notified her friends that she was using it and they made fun of her.

So Meta may be hoping for better luck with a productivity tool instead of another swing at the ChatGPT market. In May, the Information reported Meta was developing something called Hatch, its own OpenClaw-like agentic AI platform. A consumer-friendly version of exactly the kind of token-guzzling agentic platform that helped create tokenmaxxing to begin with.

Per Wired, alongside the apparent sunsetting of tokenmaxxing incentives, Meta “has been encouraging—but not requiring—employees to use its most advanced AI experiment.” That experiment is Hatch. It still hasn’t been released, but employees have had access “for the past several weeks,” Wired claims. The tool apparently burns more AI tokens than standard chatbots and AI coding tools.

Some employees apparently like it. Others in Wired’s group of sources point to an April news event, when it emerged that Meta planned to use employee keystrokes to train AI, as a reason to be wary of the thing. Meta later paused that program.

Agents are still the house religion

Heavy use of AI agents remains part of Meta’s workplace culture regardless of what the review rubric says. A memo earlier this week from AI chief Alexandr Wang explained that the company’s switch from Google Chat to Slack was meant to further enable agent use.

Slack, Wang wrote, is “the strongest platform available today for agents.” And even though “not all of us are building agents today,” he wrote, “everyone at the company will benefit from a robust and useful agent ecosystem.”

What employees are actually worried about

The fear, according to Wired, is that a successful internal rollout of Hatch is followed by layoffs.

A Reuters report last week said Meta planned to cut up to 60% of its workforce in certain areas. That plan never moved forward, according to the report, because of bugs in the AI software that would have taken over for the dismissed employees.

Which is the part worth sitting with. The cuts weren’t called off on principle. They were called off because the replacement didn’t work yet.

We reached out to Meta for information about the Reuters report and about internal testing of Hatch. We did not receive a reply.