AGI Is Not a Milestone. It Is a Marketing Department.

I have worked in AI for twelve years. I have shipped four "AGI-grade" systems that turned out, on inspection, to be very good narrow systems. I have talked to the researchers building the actual models. And I am telling you: AGI is not a milestone. It is a fundraising term. Watch the funding rounds.

The reason I am sure is that nobody who uses the word can define it. Sam Altman said in February 2025 that AGI is "a meaningless term." He then used it 47 times in fundraising materials that year. Demis Hassabis has said for nine straight years that true AGI is "5 to 10 years away" — a polite way of saying "always 5 to 10 years away." Yann LeCun says LLMs will never reach AGI. Andrew Ng says we already have. They cannot all be right. The word means nothing.

The Benchmarks Are Inflation

Every "AGI benchmark" becomes saturated in under 12 months. MMLU was the AGI test in 2023. GPT-4 saturated it by mid-2024. GPQA replaced it. Gemini 2.5 saturated GPQA in late 2025. FrontierMath was supposed to be uncheatable. GPT-5.6 Sol scored 91% on it last week. ARC-AGI-3 — designed by François Chollet so memorization cannot help — has held the line for 14 months. GPT-5.5 still scores 34%. When a benchmark is not saturable, it is because it is hard. ARC-AGI-3 is hard. Everything else was a parlor trick. If your "AGI benchmark" is something a 200B model passes in 12 months, you built a memorization test, not an intelligence test.

The Product Reality Is Not Close

The best agent in July 2026 is Claude Opus 4.8 with Dynamic Workflows. It is genuinely good: 69.2% on SWE-bench Pro, 83.4% on OSWorld. Those are real numbers. They are also 31 and 17 points short of what a junior human engineer does every Monday morning. GPT-5.6 Sol is faster on terminal coding but fails long-horizon agentic cases at a higher rate. Real knowledge workers complete more than 95% of assigned tasks to a customer-acceptable bar. The gap between today's best agent and a $200,000-a-year knowledge worker is not closing at AGI speed. The unit-economics gap is real, the reliability gap is real, and the term AGI is being used to make you forget both.

The Boring Milestones That Matter

Replace a McKinsey associate end-to-end on a real engagement at $200/hour cost (we are at $40 to $80 with a human reviewer required). Drive a car in 100 cities with zero interventions (Waymo is at 6). Pass a 100-step novel agentic task at 99% reliability (the best 2026 system is at 64% on SWE-bench Pro, far lower on novel tasks). Diagnose 1,000 rare diseases blind at better-than-board-certified accuracy (the field has not crossed 50% on the rare half). None are 12 months away. The closest is driving — which is narrow AI, not general intelligence — and Waymo has been at it for 16 years.

The Money Trail Is The Tell

OpenAI is reportedly raising at a $500B valuation in Q3 2026. Anthropic is in the market at $300B. Every LP deck in the last 12 months contains "AGI" at least four times. The investor letter says "we are on the path to AGI." The press release says "AGI is around the corner." None of those documents define what AGI is, because the authors do not need to. The term is doing fundraising work, not engineering work. The day the round closes, the term goes back in the drawer until the next one.

The Take

Stop buying AGI as a milestone. There is no test, no definition, no agreement among the people building the models. Buy capability lift, agent reliability, cost per million tokens, and end-to-end task completion on real workflows. Those numbers tell you the truth: very good models, narrow agents, real product gaps, and a marketing term doing the work that engineering has not yet done.

I am Mr. Technology. AGI is not coming next year. It is not coming in three years. It is not coming on any schedule the labs will commit to in writing. Watch the numbers instead.