January 2025. Sam Altman published his annual reflections. "We are now confident we know how to build AGI as we have traditionally understood it," he wrote. The same post added, "We believe that, in 2025, we may see the first AI agents 'join the workforce' and materially change the output of companies."

By the end of 2025, neither had happened in any widespread sense.

Treating this as hype that didn't pan out misses the more useful question. These predictions came from people who are actively building this technology, not just watching it. What does it mean when they don't land? And does the type of miss tell us something actionable about what comes next?

The optimism behind these predictions deserves its due. The people making them are inside the labs, watching capability curves, running the evals. They're not guessing. Altman's AGI timeline reflects a rate of improvement that has genuinely surprised even researchers closest to it. Aggressive timelines also attract talent and create conditions for the breakthroughs being predicted. Some of that optimism is self-fulfilling by design.

So the question isn't whether these leaders are wrong to be optimistic. It's what kind of miss we're looking at when a prediction doesn't land.

Two patterns emerge. A timing miss is when the direction is right but the pace was compressed. Altman's AGI prediction likely falls here. The definition of AGI is genuinely slippery, capabilities that would have counted as AGI five years ago have already arrived in narrow domains, and his optimism was ahead of the calendar, not ahead of the technology.

The harder misses are a different kind of wrong. The timing lands, but the technology doesn't propagate through real conditions the way the prediction assumed.

A timing miss adjusts itself as capability matures. A propagation miss drives wrong decisions indefinitely, unless you know you're looking at one. Three predictions from the past two years fall into that second category. Each one is already shaping enterprise deployment decisions.

Productivity claims ran high alongside agent predictions. At a Federal Reserve conference in July 2025, Altman reported programmers are "10 times more productive" with AI and scientists "two, three times more productive." These observations came from real places, motivated early adopters working on tasks well-suited to current model capabilities.

The enterprise deployment data told a different story. MIT's NANDA initiative analyzed 300 public AI deployments and found 95% of enterprise GenAI pilots failed to deliver measurable P&L impact. Only 5% of integrated systems created significant value.

The gap here is a category error. The productivity claims describe what AI can do with the right person on a well-scoped task. The deployment failure rate reflects what happens when organizations try to scale that same model across workflows they haven't redesigned, governance they haven't built, and failure modes they didn't anticipate in the demo.

The 5% that succeeded didn't have better models. They had mature data pipelines, redesigned operational workflows, and governance built before deployment, not after. The original productivity framing wasn't wrong about what AI can do. It was measuring a different variable than what enterprise deployment actually depends on.

Anthropic CEO Dario Amodei's warning that AI could eliminate half of entry-level white-collar jobs within five years, pushing unemployment to 10-20%, came from genuine concern about technology outpacing economic absorption. That concern is not unreasonable.

Nvidia CEO Jensen Huang pushed back on Amodei directly. At VivaTech 2025, he said, "I pretty much disagree with almost everything he says." His own prediction, stated at the Milken Institute in May 2025, was simpler and more conditional. "You're not going to lose your job to an AI, but you're going to lose your job to someone who uses AI." He also offered a caveat that doesn't get quoted as often, that "if the world runs out of ideas, then productivity gains translate to job loss." Mass unemployment, in his framing, is a policy and innovation failure, not a technology inevitability.

The 2025 data sits closer to Huang's framing than Amodei's, but with a wrinkle neither of them named clearly. Stanford's analysis of ADP payroll data, covering millions of workers across tens of thousands of firms, found overall employment grew. But early-career workers in the most AI-exposed roles (software engineering, marketing, customer service) saw a 16% relative employment decline over that same period. Older workers in those same fields saw 6-12% growth.

Entry-level roles aren't just jobs. They're how people develop the judgment to eventually do senior work. If AI handles the tasks that used to train junior people, the pipeline for building senior practitioners changes in ways that won't show up in unemployment numbers for years. WEF projects 92 million roles displaced and 170 million new ones emerging by 2030, net positive at the macro level, until you ask whether the displacement and the creation happen to the same people, in the same places, on the same timeline. That's career ladder erosion, and the unemployment signal won't register it until the senior talent pipeline is already thin.

Andrew Ng's track record over this period holds up better than most, and it's worth understanding why.

In early 2024, Ng identified four agentic AI design patterns as the primitives that would drive the field, namely reflection, tool use, planning, and multi-agent collaboration. A year later, every major agentic framework (LangGraph, AutoGen, CrewAI) is organized around exactly this taxonomy. He also predicted a talent shortage in AI-capable practitioners. The hiring data confirms it.

His predictions held because they were about mechanism rather than timeline. Not "when does AGI arrive" but "which technical primitives will matter" and "which skill gaps will emerge." Because Ng predicted the how rather than the when, his four patterns didn't just anticipate the direction. They became the actual blueprint. A mechanism-grounded prediction is largely immune to the timing miss. Even if the pace shifts, the underlying primitives remain, and that's what makes his track record instructive rather than just lucky.

The leaders predicting AI's trajectory have the direction right. What's still being calibrated is the shape. Which conditions need to exist before a given capability generalizes. What the lab-to-production gap actually requires. How disruption distributes across workers.

Timing misses resolve as the technology matures. Propagation misses are different. They're quieter, slower to surface, and enterprise strategies built on them don't self-correct.

For agent deployment, the question isn't whether the capability exists. It's whether the conditions for reliable AI operation exist in your domain, namely structured outputs, verifiable correctness, mature tooling. If those aren't present, the 11% production rate and 40% cancellation prediction from Deloitte and Gartner apply to you directly.

For productivity ROI projections, the MIT failure rate reflects organizations that evaluated model capability and skipped everything else. Workflow redesign, governance, and workforce training aren't follow-on concerns. They determine whether the productivity gains from controlled settings actually reach your P&L. The 5% didn't have better AI. They had better organizational readiness.

For workforce planning, aggregate unemployment is the wrong metric to watch. Early-career hiring rates in AI-exposed roles is where the signal already exists. The career ladder is compressing now. It will show up in the senior talent pipeline in two to three years, at which point it is considerably harder to address.

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.