AGI is here.
That sounds like a dramatic claim only because we have spent years moving the definition every time artificial intelligence crosses another line.
If artificial general intelligence means a machine that can perform a broad range of cognitive tasks at roughly the level of an average human, then we should judge the technology by that standard — not by science-fiction expectations that quietly turn AGI into superintelligence.
Today, frontier AI systems can write reports, research unfamiliar subjects, analyze data, create software, debug code, reason across text and images, use browsers, operate desktop interfaces, call tools, coordinate multi-step workflows, and produce useful work across professions they were not individually programmed to perform.
That is general intelligence in the most practical sense of the phrase.
The goalposts moved
For years, a common picture of AGI was straightforward: an artificial system with flexible, human-level ability across many intellectual tasks. It would not need a separate hand-built program for accounting, another for writing, another for programming, and another for research. It could learn the task from context and do the work.
Now that machines can actually do that, the definition often expands. AGI must be perfectly reliable. It must work autonomously for days. It must outperform every specialist. It must have a body. It must learn continuously like a human. It must understand the world exactly as we do.
Those are interesting milestones, but they are not the same milestone.
AGI does not mean ASI. General intelligence does not require universal superiority. An average person is still generally intelligent even though they make mistakes, forget instructions, struggle with unfamiliar software, and need help from specialists.
Computer use changes the conversation
The most important shift is not that AI can answer questions. It is that AI can increasingly act.
The Stanford AI Index 2026 reports that agent performance on OSWorld — a benchmark for completing tasks across real computer operating systems — rose from roughly 12% to 66.3%, within several percentage points of the human baseline. Separate published results from Pointer reported an 83.6% verified OSWorld score using Claude Opus 4.7, above the benchmark’s cited 72.4% human baseline.
Benchmarks are imperfect, and they should never be confused with the whole of intelligence. But crossing human baselines on ordinary computer interaction matters because the computer is the primary tool of modern cognitive work.
A machine that can read the screen, understand the task, manipulate software, write the document, edit the spreadsheet, search the web, run the code, and check the result is no longer merely generating text. It is participating in the same digital environment where humans perform much of the modern economy.
Yes, it is still jagged
There is an important counterpoint. Current AI is inconsistent.
The newer OSWorld 2.0 benchmark tests much longer, more realistic workflows and shows frontier agents still struggling badly with hidden state, changing information, constraint tracking, and verification. Berkeley’s Agents’ Last Exam likewise finds that current systems can complete meaningful professional work while remaining far from human experts on the hardest long-horizon tasks.
That does not make the systems narrow intelligence. It makes them uneven general intelligence.
Humans are uneven too. The difference is that AI’s strengths and weaknesses are distributed very differently from ours. A model may solve an advanced programming problem and then miss an obvious interface state. It may synthesize a hundred pages of research and then fail to ask a clarifying question.
The frontier is jagged, but it is general.
The real test is economic and practical
Instead of asking whether an AI resembles a fictional android, ask a simpler question:
Can this system be given a new cognitive task, understand what is wanted, use the tools available, and produce useful work comparable to what an ordinary person could produce?
Across a growing portion of digital work, the answer is already yes.
That does not mean every job is automated. It does not mean humans are obsolete. It does not mean we have solved robotics, reliability, alignment, consciousness, or long-term autonomy.
It means the threshold we spent decades discussing may have been crossed without a single ceremonial moment.
What happens next matters more than the label
The argument over whether to call these systems “AGI” may continue for years. The technology will not wait for the terminology to settle.
The important question is what kind of infrastructure we build around general machine intelligence.
At 3DVR, we believe that future should be open, user-controlled, interoperable, and understandable. People should own their data. Agents should operate with explicit permissions. Systems should be portable instead of locking human knowledge and identity inside one company’s platform. Open-source tools should give ordinary people access to capabilities that would once have required entire organizations.
If AGI is here, then the next challenge is not merely making intelligence more powerful.
It is making that intelligence belong to everyone.
Join the discussion