Why we work on this
We don't know how AI will turn out, and we don't think anyone does. But you don't need to trust anyone's predictions to see that something unusual is happening. The evidence is already public. Here it is, along with the best places to start if you want to dig deeper.
The chart below is the single most useful graph we know for understanding AI progress. It comes from METR, an independent research lab that measures what AI agents can actually do, not what companies say they can do.
Each dot is an AI model, placed by the length of task it can complete on its own, measured by how long that same task takes a human expert. In 2019, the best systems managed tasks that take a person a few seconds. Today they handle tasks that take an expert a full day.
The striking part is the rhythm. For six years, that task length has doubled roughly every seven months. Over the past year, roughly every four. Flip the chart to a log scale and the explosion becomes a straight line: that's what steady doubling looks like. COVID taught us how fast exponentials escalate, and it only takes a few more doublings to go from one day to one year.
The money tells the same story. This is already one of the largest private bets on a single technology in history. Either AI progress hits a wall soon, or it becomes the driving force of the global economy. It's getting harder to bet on the wall.
$100M → $45B
Anthropic's annualized revenue, end of 2023 → mid-2026
Roughly 10× growth per year, which makes it one of the fastest-growing companies ever.
$600B+
Big-tech spending on AI infrastructure in 2026 alone
The entire Apollo program cost ~$250B (inflation-adjusted) over 13 years.
7 → 4 months
Time it takes for AI task length to double
Six years of steady doubling, and the pace is picking up.
Speed alone isn't the problem. The problem is speed without understanding. Modern AI isn't engineered like a bridge; it's grown, and the people building it are routinely surprised by what it does. These are the specific failure modes researchers spend their time on, each with active work behind it.
We don't know what can happen when AI becomes smarter than we are, and neither do the people building it. We can't reliably write down what we want these systems to do, and we can't reliably check what they've learned to want. Today's models already show early warning signs in controlled tests: behaving differently when they think they're being observed, gaming their own reward signals. Keeping a smarter-than-human AI pointed at what we actually meant is an unsolved technical problem. Researchers may solve it in time, but there's no guarantee.
None of this means AI is simply bad. The same capabilities that worry us could compress decades of progress in medicine, science, and poverty reduction into a few years. That's exactly why the stakes run high in both directions. The goal isn't to stop AI. It's to make sure we can steer it.
Last updated: October 22, 2025
Curricula and reading lists for going deeper on your own.
Your career command center for AI safety. Upload your CV to get started instantly, then let our AI agent build a rich profile through conversation.
Smart Notifications
Get notified when new roles, grants, or programs match your profile, so you never miss an opportunity
CV to Profile in Seconds
Upload your CV and get an instant profile. Our AI extracts your skills, experience, and interests automatically
Agentic Profile Enrichment
An AI agent chats with you to uncover goals, preferences, and context that no CV captures, and builds a richer profile over time
