How Close Is AGI?

Deep learning-driven neural networks excel at competing high school English assignments, but it’s not clear that they’re getting anywhere close to artificial general intelligence (AGI)—that is, to an AI that mirrors the ability of humans to learn, understand, and apply knowledge across unlimited contexts. Discuss with your team: should we want to get there? How will our lives change if and when we do?

Peter Thiel's fund sold its entire $100m stake in Nvidia, and Michael Burry — who called 2008 — bet almost $200m against the same chipmaker. Both were betting on the essay-writing machine.

Key concepts

Narrow vs. General Intelligence: Today's systems are superb inside the job they were trained for and helpless a step outside it: a chess engine cannot write a poem, and a model that aces English cannot be trusted to book a flight. General means handling the unfamiliar, transferring what it learned in one place to a problem it has never seen.

+ 2 more concepts inside

Every work — at a glance

Artificial General Intelligence (technology)

The least measurable goal in technology: every benchmark that falls gets reclassified as not really general.

+ 1 more work explained inside

What to know

01
Passing the test is not the same as understanding the material, and the essay is the clearest case. A model can produce a strong high-school English paper by pattern from millions of examples without holding a view about the book, which is why benchmark scores keep rising while the ability to handle genuinely new problems moves slowly. The goalposts move too: chess, Go, the bar exam and passable essays were each meant to be the mark of a mind, and each was reclassified the week after it fell.

+ 1 more insight inside

Across subjects

Goethe wrote the control problem in 1797. The Sorcerer's Apprentice enchants a broom to fetch water, discovers he does not know the word that stops it, and

Theme connection

"Are We There Yet?"

Should we want to arrive at all? One case: the end of disease, drudgery and most human error. The other: mass job loss, power held by whoever owns the model, a mind nobody can switch off. And under both sits a third possibility — that the race ends in a burst balloon, not a breakthrough.

Practice — a sample

SCIA chess engine is to narrow AI as someone who can learn any job is to —

A A benchmark score
B The scaling hypothesis
C General intelligence
D A neural network
E The AI bubble

Keep reading the full lesson

The remaining insights, works, and practice come with full access — $9.99/month or $29.99/season.

Unlock full access →

New here? Create a free account to read the free section.