Forget the Label. Watch the Machine.
There is a debate raging about whether we are close to AGI. The debate is almost entirely useless. While researchers argue about definitions and timelines, the technology they are trying to label is already doing weeks of expert-level autonomous work, solving graduate-level science problems at 90%+ accuracy, and writing production software that passes 80% of real-world engineering benchmarks. The machine does not care what you call it.
The fixation on labels creates a genuine perceptual blind spot. Russian speakers have two distinct words for light blue and dark blue, and neuroimaging shows they literally perceive a categorical boundary between shades that English speakers experience as a smooth gradient. The same thing is happening with intelligence. People are scanning for a sci-fi entity that passes some philosophical test of consciousness. Because the thing in front of them looks like a chatbot with a text box, they cannot see that it already outperforms most professionals across most cognitive domains. By the time anyone agrees on the label, the economic consequences will already be irreversible.
The Convergence
A remarkable cluster of predictions has emerged over the past eighteen months. Anthropic expects systems matching Nobel laureates across most disciplines by early 2027. OpenAI’s internal roadmap targets a fully automated AI researcher by 2028. Google DeepMind gives roughly even odds on AGI by decade’s end. Nvidia, SoftBank, and multiple AI pioneers converge on similar windows. The most detailed public forecast, the AI 2027 scenario, originally placed superhuman coding capability in early 2027 with recursive self-improvement shortly after, though its authors have since lengthened their median estimates by a few years based on 2025 data.
The logic behind these predictions is mechanical, grounded in observable trends. Compute is scaling at roughly half an order of magnitude per year. Algorithmic efficiency gains add another half order of magnitude. Agent task horizons are doubling every four to seven months. The systems that could handle two-hour autonomous coding tasks last year now handle fourteen-hour sessions. Four to six more doublings puts you at week-long autonomous projects. The trajectory has held for six years without interruption.
The Bottleneck Cascade
Every time one bottleneck falls, the next one emerges. Chips were the crisis of 2024. Massive capital allocation and fab expansion resolved it within eighteen months. Now memory is the binding constraint, with high-bandwidth memory (HBM) sold out through 2026 across all three global suppliers. The market is attacking it with the same ferocity, and relief is expected by late 2027 as new mega-fabs come online.
Energy is the harder problem. You can build a chip fab in two years. A nuclear reactor takes a decade. The entire AI infrastructure buildout demands hundreds of gigawatts of new power by 2030, and the physical world simply cannot be rushed past certain floors regardless of capital deployed. This is likely the strongest natural brake on the trajectory, and it may be the reason the singularity, if it arrives, feels more like a steep ramp than a vertical wall.
The Jobless Expansion
The economic impact is already visible for anyone willing to look past the headline statistics. US GDP grew 3.7% in late 2025 while job creation was revised down to one of the worst years outside a recession. White-collar employment in finance, insurance, information, and professional services peaked in November 2022 and has declined ever since. Entry-level hiring in AI-exposed occupations has dropped 16% and is still falling. The economy is growing. The jobs are vanishing.
This is textbook Solow’s Paradox meeting the J-curve of productivity. General-purpose technologies always show a lag between capability and measurable economic impact because adoption requires massive invisible investment in workflow redesign, retraining, and organizational restructuring. We are deep in that investment phase now. A small vanguard of companies is already compressing weeks of work into hours. The vast majority are still experimenting. When the laggards catch up, and competitive pressure guarantees they will, the harvest phase will hit the labor market like a series of rolling dam breaks across sectors.
The Forest and the Placard
Every previous automation wave targeted a specific type of work. Machines replaced muscle. Computers replaced calculation. AI targets cognition itself, the common substrate underneath virtually all knowledge work. And unlike every previous technology, this one is actively being used to improve itself. The feedback loop only turns in one direction.
The real question was never “is it truly intelligent?” It was always “can it do the job?” And by every empirical measure, the answer is converging on yes across an expanding frontier of professional domains. The people debating the placard at the gate have already missed the forest growing behind it. The efficiency-seeking logic of markets guarantees that once the capability exists, adoption follows. The capability exists now. The adoption is accelerating. The shocks are coming. The only variable left is whether we prepare.
