Meta's AI Model Delay Reveals Shifting Power in

So Meta's latest AI model isn't coming out this month. Or next month. Maybe May, if things go well. The model—codenamed Avocado, because apparently we're naming AI after produce now—has been delayed because it's just not good enough. And honestly? That might be the most interesting story here.

According to The New York Times, Avocado is falling short on basically every metric that matters for modern large language models: reasoning, coding, writing. It reportedly beats Google's Gemini 2.5 but can't touch Gemini 3. That's a problem when you've spent nine months on development and the goalposts have moved three times since you started.

Meta's official statement tries to sound optimistic—"our next model will be good, but more importantly, show the rapid trajectory we're on"—but here's the thing that makes me raise an eyebrow: they're apparently considering licensing Gemini as a stopgap. That's not what you do when you're confident about your pipeline.

The Three-Way Race (And Who's Not In It)

Ethan Mollick captured the vibe when he tweeted that frontier AI models are "really a three-way race at this point." He's talking about OpenAI, Google, and Anthropic. Not Meta. Not XAI. The leaders are pulling away, and the companies that aren't in that top tier are having to make some uncomfortable decisions about what comes next.

Meta's been here before—remember when everyone thought they'd dominate VR? But AI feels different. There's no clear second-place prize, no participation trophy for trying hard. Either your model is competitive or people use something else. The switching costs are essentially zero.

What's wild is that Meta has resources most companies can only dream about. They've got compute, they've got talent, they've got distribution through their platforms. And they're still struggling to keep pace. That tells you something about how technically difficult this actually is.

XAI's Controlled Demolition (Or Is It?)

Meanwhile, Elon Musk's XAI is going through what he's calling a "controlled demolition." Six of the company's twelve co-founders have left this year. Six. Only three remain, and one of them is Musk himself.

Musk's framing is that "XAI was not built right the first time around so is being rebuilt from the foundations up. Same thing happened with Tesla." Which, okay, but Tesla didn't lose half its founding team in a few months. There's rebuilding, and then there's whatever this is.

The company just poached two senior leaders from Cursor—Andrew Militch and Jason Ginsburg—to help catch up on coding capabilities. Musk admitted at a conference that XAI is behind but expects to "catch up and exceed our competitors" by mid-year. That's an aggressive timeline given that they're also losing institutional knowledge every few weeks.

Here's the tension: Cursor itself is raising money at a $50 billion valuation, suggesting they're trying to compete independently rather than getting acquired. Their CEO told employees in January that they're in "wartime," planning to build their own state-of-the-art models to reduce dependency on other labs. So XAI is trying to catch up by hiring from a company that's simultaneously trying to become a direct competitor. It's recursive competition.

The Implementation Gap

But maybe the most underreported story is this: enterprises have no idea how to actually use any of this stuff.

Anthropic is apparently in talks with Blackstone to launch an AI consulting venture specifically because corporate customers need help implementing the technology. Blackstone wants to deliver these services to their hundreds of portfolio companies. They had similar talks with OpenAI. This isn't about the models being insufficient—it's about the organizational muscle memory not existing yet.

Think about what that means. These companies can buy access to cutting-edge AI tomorrow. The constraint isn't capability or cost. It's knowing what to do with it. It's having the internal processes, the change management, the human expertise to actually integrate these tools into workflows.

That's why 81% of doctors now use AI but mostly for administrative tasks—keeping up with research, generating discharge instructions, documenting appointments. Only 17% are using it for assisted diagnosis. The American Medical Association is so careful about this distinction that they call it "augmented intelligence" instead of artificial intelligence, hammering home that it's not replacing human judgment.

AMA CEO John White said usage has more than doubled since 2023, which is genuinely fast adoption for healthcare. But notice what's being adopted: the boring stuff. The paperwork. The parts of the job that nobody went to med school to do.

Altman's Meter Is Running

Sam Altman gave a talk at a BlackRock conference this week that's worth sitting with. He said OpenAI's business model is basically "selling tokens"—treating intelligence as a metered utility like electricity or water. "Fundamentally, our business is going to look like selling tokens. We see a future where intelligence is a utility like electricity or water, and people buy it from us on a meter."

That's a very different framing from "AGI will solve all problems" or "intelligence too cheap to meter." He's saying: we're going to charge you for every API call, and given how demand is scaling, those prices might stay high for a while.

Altman also said the term AGI has "lost all meaning," which feels significant coming from the CEO whose company has been using that term constantly. Instead, he's watching for two milestones: when most of the world's intelligence is in data centers (maybe 2028), and when leaders can't do their jobs without heavy AI reliance.

That second one is interesting because it's about dependency, not capability. It's the moment when opting out stops being viable. And Altman thinks it might happen soon—though he joked that for some jobs, "that pretty much describes my job already."

The economic implications are what he kept circling back to. "The entire structure of capitalism is designed to manage scarcity," he said. "If AI delivers true abundance, then society will need to rapidly adjust to a new paradigm." More immediately, he noted that AI is "disrupting the balance between labor and capital that keep society functioning."

His prediction: "I am not a long-term jobs doomer. I think we will figure out new things to do, but I think the next few years are going to be a painful adjustment."

What Actually Matters Here

So we've got Meta delaying models, XAI losing cofounders while raiding Cursor for talent, Cursor raising at valuations that suggest they're planning to compete forever, consulting firms seeing an opening because nobody knows how to implement any of this, and the CEO of OpenAI quietly shifting from "AGI will save us" to "get ready for a painful adjustment."

The pattern I'm seeing: the technology is advancing faster than our ability to integrate it, but not as fast as the hype suggested. We're in this weird middle zone where the tools are genuinely useful but also not transformative enough to justify some of the wilder claims. And the companies building these tools are discovering that being technically impressive doesn't automatically translate to market dominance.

Meta can delay Avocado because releasing something mediocre would be worse than waiting. That's probably the right call. But it also means they're falling further behind in a race where second place might not matter. XAI can call it a controlled demolition, but when half your cofounders leave, you're rebuilding trust as much as technology. And Cursor can pursue independence, but training competitive models from scratch is a bet that most startups don't survive.

Meanwhile, 81% of doctors are using AI for paperwork, and Sam Altman is telling Wall Street that the next few years will be painful. Those things are connected. We're getting the technology before we've figured out what it's for—or what it costs.

—Zara Chen