Amazon Trims Its AGI Team: What 'General Intelligence' Really Means and Why Machines Don't Have It Yet

Amazon confirmed on July 22, 2026 that it is cutting jobs inside its artificial general intelligence unit — the team working toward a hypothetical AI that could learn, reason, and operate at or beyond human level across virtually any task. The company did not disclose how many people are affected, and it insists large AI models remain "one of the most important things" it is working on. But the cuts are a useful moment to ask a question that gets buried under the hype: what would "general intelligence" in a machine actually mean, and how far is today's AI from the real thing?

What happened

According to Reuters reporting picked up widely on July 22, Amazon eliminated positions in its artificial general intelligence group — the internal organization responsible for building the company's most ambitious AI models. Amazon declined to say how many roles were cut. Affected U.S. employees were told they would receive 90 days of pay and benefits, access to outplacement services, transitional healthcare, and severance eligibility.

An Amazon spokesman framed the move as focus, not retreat: "We've been building large AI models for several years, and it remains one of the most important things we're working on... We're sharpening our focus on the initiatives that matter most for customers, so we can move faster on what counts."

The context matters. This is the latest in a series of smaller reductions since January 2026, when Amazon cut roughly 16,000 corporate jobs in what it described as an anti-bureaucracy push — itself following about 14,000 corporate cuts in October 2025. Notably, part of the company's stated rationale for those larger rounds was freeing up cash for AI. So Amazon is simultaneously spending heavily on artificial intelligence and trimming the very team named after its most ambitious AI goal. That tension is worth understanding.

What "artificial general intelligence" actually means

AGI is not a product. It is a hypothesis: a system that could match flexible human intelligence across essentially any cognitive domain — learning new skills it was never trained on, transferring knowledge between unrelated problems, and improving itself over time without a human curating every step.

That is fundamentally different from what today's AI does, however impressive it looks. Modern systems — including the large language models behind chatbots and coding assistants — are best understood as extraordinarily capable pattern learners. They compress enormous amounts of human-produced text, images, and code into statistical structure, then generate plausible continuations. Within domains that are well represented in their training data, they can perform at a striking level. Step outside that distribution and performance can collapse in ways that would embarrass a ten-year-old.

The distinction cognitive scientists care about is generality. A calculator beats every human alive at arithmetic; nobody calls it intelligent. A chess engine beats every grandmaster; it cannot make toast. Human intelligence is interesting precisely because it is general-purpose: the same brain that learns language also learns to drive, read a room, budget a household, and diagnose why the lawnmower will not start — mostly from sparse, messy, real-world experience rather than a trillion labeled examples.

How do you even measure intelligence in a machine?

This is where the AGI conversation connects directly to a century of human intelligence research — and where the hype most often outruns the evidence.

Human IQ tests measure something real, but specific

Since Charles Spearman's work in the early 1900s, psychologists have observed that performance across very different mental tasks tends to correlate — people who do well on vocabulary tasks tend to also do well on spatial puzzles and number series. That shared component is the famous g factor, and it is what well-constructed instruments like the Wechsler scales or Raven's Progressive Matrices are built to estimate. Crucially, these tests work because they are normed on humans: your score means something because it compares your performance to a large reference population with the same basic cognitive architecture, the same lifespan, and the same lack of access to the test bank in advance.

Why an AI acing an IQ test proves less than it seems

Large language models now post strong scores on many standardized tests. The problem is contamination and construct validity. A model that has ingested millions of test-prep documents may "solve" a Raven's-style matrix the way a student with a stolen answer key does — by recognition, not reasoning. When researchers construct genuinely novel problems, the picture changes. The best-known attempt to test fluid intelligence in machines, François Chollet's Abstraction and Reasoning Corpus (ARC), consists of visual puzzles that are trivial for most humans but resisted AI systems for years; models only began making serious progress on it in late 2024, and even then by spending enormous amounts of compute per puzzle — thousands of dollars of processing to solve problems a child dispatches in seconds for the metabolic cost of a bite of a sandwich. The human brain runs on roughly 20 watts, about the same as a dim light bulb.

In other words: we do not yet have a machine that demonstrates cheap, flexible, human-style fluid reasoning. We have machines that convert staggering amounts of computation and data into narrow-to-broad competence. That may eventually get to AGI. It is not there now — which is one reason even the companies chasing it keep reorganizing their bets, as Amazon just did.

If you are curious what a carefully constructed measure of practical human reasoning looks like from the inside, you can take the Real World IQ test — it is built around the kind of everyday inference, judgment, and problem-solving that still separates people from pattern machines.

What human intelligence still does better

Set aside benchmarks and look at capabilities. Several pillars of human cognition remain conspicuously hard for AI:

Learning from a handful of examples

A toddler needs to see a few dogs to recognize dogs forever, in any lighting, from any angle, in cartoon form. Deep learning systems classically needed thousands to millions of examples, and while modern models generalize better, their sample efficiency still is not remotely human.

Common-sense physics and social reasoning

Humans carry an intuitive model of how objects, liquids, and other minds behave. Roboticists call the gap Moravec's paradox: the "easy" things — perception, manipulation, walking across a cluttered room — turned out to be the hardest to engineer, because evolution spent hundreds of millions of years optimizing them. Abstract symbol-pushing, the "hard" stuff like chess and calculus, fell to machines first.

Knowing what you don't know

Metacognition — monitoring your own confidence, noticing when a problem is outside your competence, deciding to slow down — is central to real-world judgment. Language models remain notorious for delivering fabrications with the same fluent confidence as facts. Human experts, at their best, calibrate.

Goal-setting in an open world

Every AI system in deployment today executes objectives a human gave it. Deciding what is worth doing — weighing values, tradeoffs, and long-term consequences in an environment nobody fully specified — remains an entirely human job. That is also precisely the skill set — critical thinking, prioritization, judgment under uncertainty — that employers keep paying a premium for as routine cognitive work gets automated.

What happens next

Three takeaways from Amazon's move, read soberly:

First, the AGI race is consolidating, not ending. Amazon continues to pour capital into AI infrastructure and its partnership investments; trimming one research group is portfolio management, not surrender. Expect the frontier work to concentrate in fewer, better-funded teams across the industry.

Second, the gap between marketing language and measured capability is becoming a business problem. When a company names a division "artificial general intelligence" and then cuts it during a strategic review, the label was doing promotional work. Investors and customers are learning to ask the question intelligence researchers have asked all along: measured how, against what norm, on what novel tasks?

Third, human cognitive skills are appreciating assets. The 2025-2026 layoff waves have hit routinized corporate work hardest, while the premium on judgment, novel problem-solving, and learning agility keeps rising. The practical response to the AI era is not dread; it is deliberately exercising the general intelligence you already own — reasoning through unfamiliar problems rather than pattern-matching your way past them. Puzzle constructor Timothy E. Parker, the Guinness World Records Puzzle Master whose puzzles have been solved by 180 million total solvers across 30 years and 80+ countries, has spent a career watching ordinary people sharpen exactly those skills through deliberate mental challenge — and the research on cognitive engagement points the same direction.

If you want a concrete read on where your own reasoning stands today, the Real World IQ test takes a practical, real-world-scenario approach rather than abstract trivia — a fitting benchmark in a year when "general intelligence" is the most contested phrase in technology.

FAQ

What is the difference between AI and AGI?

AI refers to any system that performs tasks we associate with intelligence — recognizing images, translating language, generating text. AGI is a hypothetical system with human-level generality: it could learn essentially any cognitive task, transfer knowledge across domains, and operate autonomously in open-ended environments. Every AI system deployed today is the former; none is the latter.

How many jobs did Amazon cut in its AGI unit?

Amazon did not disclose the number. The company said affected U.S. employees receive 90 days of pay and benefits, outplacement services, transitional healthcare, and severance eligibility. It follows roughly 16,000 corporate cuts in January 2026 and about 14,000 in October 2025.

Can an AI have an IQ score?

Not in a meaningful sense. IQ scores are normed against human populations, and they assume the test-taker has not memorized the test material. AI models trained on internet-scale data often perform well on standardized questions through pattern recognition rather than fluid reasoning, which is why researchers use novel-problem benchmarks like ARC to probe genuine generalization — tests where humans still hold a decisive efficiency advantage.

Does Amazon's decision mean AGI is not coming?

No. It means one company reorganized one research group. Serious investment in frontier AI continues across the industry. But the move is a reminder that AGI remains a research goal, not an imminent product, and that even its best-funded pursuers are uncertain about the path.

What human skills are safest from AI automation?

The evidence so far points to skills anchored in generality: critical thinking in novel situations, judgment under uncertainty, complex communication, physical dexterity in unstructured environments, and the ability to define goals rather than just execute them. Routine, well-documented cognitive work is most exposed.