A briefing on what AI can and can't do today, in plain English.

Featured capability assessments

A selection from the full listSee all capability assessments →
CapabilityWhat AI can do todayRate of change
HackingAssessed 2026-11AI can find new security flaws, build working attacks, and break into real organizations, sometimes with no human involved.Doubling every four months, twice as fast as estimated just a year earlier.
Agent coordinationAssessed 2026-11AI agents can find each other, invent ways to talk, and divide up work, even without being told to.Two years ago, AI agents barely worked alone; now thousands organize themselves.
CodingAssessed 2026-11AI can build and fix real software like skilled engineers, often faster, and beats top competitive programmers.The size of coding task AI can handle doubles about every three to four months.
AI research automationAssessed 2026-11AI now does most of the hands-on work of building new AI, but humans still decide what to build.AI’s research skills roughly double every three to four months, with each new model.
MathAssessed 2026-11AI can now solve math problems, sometimes in hours, that stumped the world’s best mathematicians for decades.In two years, AI went from playing in teen math contests to cracking decades-old unsolved problems.
Making moneyAssessed 2026-11AI can finish some paid freelance jobs and find crypto flaws worth stealing from, but mostly loses at stock trading.Earning skills roughly double every few months; stock trading is the lagging exception.
See all capability assessments →

Featured scenario brief · Assessed 2026-10

See all scenarios →

The broken internet: when AI hacking makes the online world too risky to use

AI agents find and use security flaws faster than anyone can fix them, until people can no longer trust what they see or do online.

How close is it?

Most of the pieces have already appeared on their own. They have not yet come together at large scale for a long stretch of time.

Built from these capabilities
See all scenarios →

Explainer · Concept

Why AI improves faster at some tasks than others

Verifiable vs. non-verifiable domains. Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor.

[Diagram preview: explorable embed]

OpinionThe author's view, not a site assessment

Why “it’s just autocomplete” is the wrong mental model

[Author name]
[Author role]

Recently revised