Key Points
- 1.Demis Hassabis argues the term AGI has been overused to the point of losing meaning, and separates two distinct ideas.
- 2.One is matching what a typical person (or about 90% of humans) can do, which is economically and product-important but not his definition of AGI.
- 3.His AGI bar is theoretical: matching the full range the human brain architecture can achieve, including the best humans in history such as Einstein and Mozart.
- 4.Today's systems are not consistent enough to be fully general; an individual can find an obvious flaw in minutes.
- 5.His test: it should take a team of experts months, not an individual minutes, to find an obvious hole.
Summary
Two ideas get conflated
Hassabis separates what a typical person can do (or what about 90% of humans can do), which is economically and product-important, from true AGI, which he treats as a higher and more theoretical bar.
His definition of AGI
AGI means matching the full range the human brain architecture can achieve, pointing to the best humans in history (Einstein, Mozart) reaching their feats with the same brain architecture. Today's systems clearly fall short of that.
Consistency is the missing piece
Current systems are quite general (they do thousands of things) but not consistent enough to be fully general. Anyone can find an obvious flaw, a high-school math slip or a simple game it cannot play, within minutes.
A harder test
For something to count as AGI, finding an obvious hole should take a team of experts a couple of months, not an individual a few minutes.
Worth watching for
Anyone following the AGI debate who wants a precise, grounded definition from a leading AI researcher.
- agi
- demis-hassabis
- deepmind
- ai-pm