Aiming to Learn as We Do, a Machine Teaches Itself
Give a computer a task that can be crisply defined — win at chess, predict the weather — and the machine bests humans nearly every time. Yet when problems are nuanced or ambiguous, or require combining varied sources of information, computers are no match for human intelligence.
Context & Ripple Effects
This Times piece lands nine months after Google's own push on helping computers understand language, and the two stories mark the same pivot from opposite ends: search companies were trying to make rules-based systems parse human language, while researchers behind this self-teaching machine argue the better route is to let the system build its own knowledge, the way a person does.
The framing matters because it draws the industry's working boundary in 2010 — machines beat humans on crisply defined tasks like chess and weather prediction, but lose wherever nuance, ambiguity, or combining multiple sources is required. A machine that 'teaches itself' is a direct attack on the second half of that boundary.
First-order effects
- Research programs in language understanding — Google's included — gain a demonstrated alternative to hand-engineered rules: systems that acquire their own knowledge from raw material rather than being explicitly programmed.
- Tasks the field had written off as out of reach, such as synthesizing varied and ambiguous sources, become explicit engineering targets instead of accepted limits on automation.
Second-order effects
- Competition among search and information companies shifts a notch from indexing scale toward quality of machine-extracted meaning, since whoever's system learns to combine sources handles ambiguous queries better.
- Organizations holding very large corpora of unlabeled text — web-scale search firms above all — become disproportionately advantaged, because self-teaching approaches feed on exactly the data they already possess.
Third-order effects
- If self-teaching systems keep improving, the bottleneck for machine intelligence migrates from writing algorithms to supplying the right raw experience, reshaping which institutions can compete in AI.
- The neat split between 'crisply defined' tasks machines win and 'nuanced' tasks reserved for humans, which organizes this article, becomes a moving frontier eroded domain by domain rather than a stable division of labor.
The trend: Machine intelligence in the early 2010s is beginning its move from programmed, single-task systems toward machines that teach themselves from unstructured information.