Is today's AI already AGI?

It depends on whose standard you apply. By a 1995 standard, a system that talks with hundreds of millions of people a week, writes software and passes medical exams would count. By today's working definition, an AI that can do almost any mental task a skilled person can and learn new ones, it does not yet qualify. Current systems still struggle to learn on the job, stay reliable over long periods and handle the physical world.

What is the AI effect?

The AI effect is the habit of redefining intelligence every time a machine masters a task. In 1979 Douglas Hofstadter credited Larry Tesler with the line that AI is whatever hasn't been done yet. Chess, Go, translation and olympiad mathematics each moved from proof of intelligence to routine computation once machines could do them.

What is METR's 50 percent time horizon?

It is the length of task, measured in how long a skilled person takes, that an AI agent completes half the time. In its report on February and March 2026, METR estimated the strongest agents it evaluated at 16 to 20 hours at 50 percent success and 3 to 4 hours at 80 percent. It also warned that its task suite cannot reliably measure above 16 hours.

How can a company measure AI progress on its own work?

Pick one workflow with checkable output, collect about twenty real tasks, record how long a skilled employee takes for each and run the current AI system on them every quarter without help. Plot the share completed correctly against task length. The result is a private time-horizon curve at the reliability the process actually needs.

Why does entry-level hiring matter for AI adoption?

Juniors learn judgment by doing the work AI now does well. A study of customer support agents found AI assistance raised productivity most for novice and low-skilled workers, while US payroll data shows employment of 22 to 25 year olds in AI-exposed occupations 19 percent below the trend of less-exposed peers, mostly through reduced hiring. Cutting the first rung can leave a company with more AI output and fewer people able to verify it.

Is This Already AGI? The Better Question Is What You Hand Over

By Juan Beltrán · 2026-09-26

The AGI label will arrive late, because every capability gets renamed once it works. Leaders should measure three things instead: how long AI can work on their own tasks at the reliability their process needs, which step of the capability staircase each workflow is on, and whether they are still training the people who will check the output.

Key takeaways

The AI effect has a corporate version: working capability gets renamed as a tool, then funded and governed like one.
Read AI progress at the reliability your process needs. METR put the strongest agents at 16 to 20 hours of work at 50 percent success, but 3 to 4 hours at 80 percent.
Build your own time-horizon curve from about twenty real tasks in one workflow and rerun it every quarter.
Place each workflow on the five-step staircase of talk, reason, act, discover and organise, and write down why it sits below what the technology could support.
Keep hiring at the first rung but change the job, so juniors learn judgment by reviewing and correcting agent output under a senior.

Questions executives ask

Is today's AI already AGI?: It depends on whose standard you apply. By a 1995 standard, a system that talks with hundreds of millions of people a week, writes software and passes medical exams would count. By today's working definition, an AI that can do almost any mental task a skilled person can and learn new ones, it does not yet qualify. Current systems still struggle to learn on the job, stay reliable over long periods and handle the physical world.
What is the AI effect?: The AI effect is the habit of redefining intelligence every time a machine masters a task. In 1979 Douglas Hofstadter credited Larry Tesler with the line that AI is whatever hasn't been done yet. Chess, Go, translation and olympiad mathematics each moved from proof of intelligence to routine computation once machines could do them.
What is METR's 50 percent time horizon?: It is the length of task, measured in how long a skilled person takes, that an AI agent completes half the time. In its report on February and March 2026, METR estimated the strongest agents it evaluated at 16 to 20 hours at 50 percent success and 3 to 4 hours at 80 percent. It also warned that its task suite cannot reliably measure above 16 hours.
How can a company measure AI progress on its own work?: Pick one workflow with checkable output, collect about twenty real tasks, record how long a skilled employee takes for each and run the current AI system on them every quarter without help. Plot the share completed correctly against task length. The result is a private time-horizon curve at the reliability the process actually needs.
Why does entry-level hiring matter for AI adoption?: Juniors learn judgment by doing the work AI now does well. A study of customer support agents found AI assistance raised productivity most for novice and low-skilled workers, while US payroll data shows employment of 22 to 25 year olds in AI-exposed occupations 19 percent below the trend of less-exposed peers, mostly through reduced hiring. Cutting the first rung can leave a company with more AI output and fewer people able to verify it.

Canonical URL: https://juanbeltran.ch/blog/is-this-already-agi-what-you-hand-over