Dig deeper · Supporting research
The charts
behind the curve.
behind the curve.
Four visualizations — each showing the same exponential story from a different angle. Click any to expand.
Based on Tim Urban · Wait But Why · The gap between assumption (linear) and reality (exponential) is where the surprise lives.
OurWorldInData.org/artificial-intelligence · CC BY · Kiela et al. 2023 · Zero line = human performance. Every line has crossed it.
Leopold Aschenbrenner · Situational Awareness (2024) · Log scale. GPT-4 = 1.0. Note where "Automated AI Research" begins accelerating the curve further.
Why humans don't get exponential growth
Start with
a penny.
a penny.
Double it every day for 30 days. Most people guess the answer is somewhere between a few hundred and a few thousand dollars. They're off by about five million dollars.
Days 1–10
Day 1 → Day 10
$5.12
Feels like nothing. This is the part where people give up on the idea that doubling matters.
Days 11–20
Day 11 → Day 20
$5,242
Starting to notice. Something is clearly happening, but it still doesn't feel like a big deal.
Days 21–30
Day 21 → Day 30
$5.37M
This is where the curve goes vertical. The last 10 days produce more than the previous 20 combined — by a million miles.
Current total
$0.01
The exponential curve — plotted
Each bar is one day. Toggle to see why the first 20 days look flat until they don't.
Days 1–10: max $5
Days 11–20: max $5,242
Days 21–30: $5.4 million
Are we on day 10 of AI
or day 20?
The first 10 days of the penny = $5. Useful, but not life-changing. The last 10 days = $5 million. That's the question right now. Which phase of the curve are we in? Click to see what the data suggests.
If we're on day 10...
AI is useful for simple tasks — drafting emails, basic summaries, one-off questions. The technology works but feels incremental. Most people can see the value but it doesn't feel urgent. There's plenty of time to learn it "later." This is where most people think we are right now.
AI is useful for simple tasks — drafting emails, basic summaries, one-off questions. The technology works but feels incremental. Most people can see the value but it doesn't feel urgent. There's plenty of time to learn it "later." This is where most people think we are right now.
If we're on day 20...
The evidence points here. The METR research shows AI task capability has been doubling every 4–7 months since 2019. Two years ago, AI could complete a 2-minute task reliably. Today it's completing 5–7 hour tasks. A year from now, at this rate: a full work week of autonomous work. The last 10 days on the penny start now — and most people are still thinking about day 10.
The evidence points here. The METR research shows AI task capability has been doubling every 4–7 months since 2019. Two years ago, AI could complete a 2-minute task reliably. Today it's completing 5–7 hour tasks. A year from now, at this rate: a full work week of autonomous work. The last 10 days on the penny start now — and most people are still thinking about day 10.
Based on Tim Urban · Wait But Why
The assumption
Progress feels linear. Each year is a little better than the last. AI was at "smart high schooler" a few years ago. Maybe it'll be at "college grad" by 2030. We have time.
Where this puts you
Disappointed when it doesn't quite match expectations. Use it occasionally. Think of it as a smart autocomplete.
The reality
Progress is exponential. 2022 AI couldn't reliably finish a 30-second task. Today's AI handles 5–7 hour projects. The curve is already vertical — most people just can't see it yet because they're still using AI like it's 2022.
Where this puts you
Surprised — then permanently changed. Use it for everything. Start wondering how you ever worked without it.
The research · METR · 2025
The new
Moore's Law.
Moore's Law.
METR (Model Evaluation & Threat Research) tested 13 frontier AI models from 2019 to 2025 on hundreds of real tasks ranging from 30 seconds to 8+ hours of human work. The finding: the complexity of task AI can complete autonomously doubles every 4–7 months. No signs of slowing.
AI task time horizon — plotted
Length of task AI can complete autonomously at 50% reliability. Linear = see the wall. Log = see the trend.
Confirmed data
Projected (if trend holds)
Our World in Data · AI vs Human benchmarks · Kiela et al. 2023
Across 13 different capabilities — reading comprehension, image recognition, math, coding, language understanding — AI started each benchmark at –100 (far below human). The zero line = human performance. Every single capability has now crossed zero. AI surpasses human performance on all of them.
Reading comprehension ✓
Image recognition ✓
Language understanding ✓
Math problem-solving ✓
Code generation ✓
Speech recognition ✓
General knowledge ✓
Complex reasoning →
OurWorldInData.org/artificial-intelligence · CC BY
2019
<1 min
Simple lookups only
Early 2024
30 min
Research, drafting, analysis
Now · 2026
5–7 hrs
Complex multi-step projects
Late 2026
~14 hrs
Multi-day tasks, projected
2027
Weeks
If trend holds
Overall doubling rate · 2019–2025
7 mo
Every 7 months, the length of task AI can complete autonomously at 50% reliability doubles. This was the baseline trend across the full 6-year period.
Accelerated rate · 2024–2026
4 mo
In the most recent period, the trend has accelerated to doubling every 4 months — likely due to the reasoning model paradigm shift. The curve is steepening.
Two years ago, AI could handle a 2-minute task reliably.
Today: 5–7 hours. In 12 months: a full work week.
This is the same penny. We just passed day 20.
Source: METR — "Measuring AI Ability to Complete Long Tasks" (2025) · theaidigest.org/time-horizons
The assumption problem
You wade into a lake. The first 10 feet is gradual.
You wouldn't assume the next 4 steps are the same — especially if you know there's a drop-off ahead. But that's exactly what most people do with AI. They see the first 10 days of the penny (gradual, unimpressive) and assume the next 10 will look the same. They look at past growth and extrapolate forward in a straight line.
Exponential growth doesn't extrapolate linearly. The curve gets steeper every period — not flatter. The past growth is actually the slowest this technology will ever be. What looks "fast" right now is the floor, not the ceiling.
Exponential growth doesn't extrapolate linearly. The curve gets steeper every period — not flatter. The past growth is actually the slowest this technology will ever be. What looks "fast" right now is the floor, not the ceiling.
What you see
Past growth. Impressive but gradual. "I can handle this pace."
What's coming
The same doubling rate applied to a much larger base number. The drop-off.
Dig deeper · Supporting research
The charts
behind the curve.
behind the curve.
Four visualizations — each showing the same exponential story from a different angle. Click any to expand.
Based on Tim Urban · Wait But Why · The gap between assumption (linear) and reality (exponential) is where the surprise lives.
OurWorldInData.org/artificial-intelligence · CC BY · Kiela et al. 2023 · Zero line = human performance. Every line has crossed it.
Leopold Aschenbrenner · Situational Awareness (2024) · Log scale. GPT-4 = 1.0. Note where "Automated AI Research" begins accelerating the curve further.
80,000 Hours · METR data · "Length of tasks AI can do doubling every 7 months"
AI Digest · METR data · "A new Moore's Law for AI agents"
Both charts show the same data. GPT-2 through GPT-3.5 are nearly invisible — basically zero. Then the curve goes near-vertical. This is what exponential growth looks like when you stop hiding it on a log scale. Your intuition about "how fast AI is moving" is almost certainly too conservative. The past rate is the floor, not the ceiling.
What this means in practice
It figures
it out.
it out.
As capability grows, so does inference. Claude doesn't just execute instructions — it reads context and adapts. You don't tell it your tone, your style, your client's expectations. It picks them up from what you give it. Here's a real example of what that looks like.
Live example — what Claude infers without being told
What you give it
From: Ryan O'Connor
To: Sarah @ Nike
Hey Sarah — quick update on the activation. Team's locked in on the copper key concept, still working through the fabrication timeline. Will have full details by EOW. Let me know if you need anything before then.
— R
To: Sarah @ Nike
Hey Sarah — quick update on the activation. Team's locked in on the copper key concept, still working through the fabrication timeline. Will have full details by EOW. Let me know if you need anything before then.
— R
New email from Sarah:
"Hi Ryan, any chance you could give me a rough budget range for the fabrication piece? Need to flag it internally before Friday."
"Hi Ryan, any chance you could give me a rough budget range for the fabrication piece? Need to flag it internally before Friday."
All you say: "Draft a reply to Sarah's new email."
What Claude infers
Tone: Casual but professional. First name basis. Short sentences. Signs off as "R" not full name.
Relationship: Collaborative, established. Sarah is a client, not a stranger. Warm but efficient.
Urgency: Friday deadline noted. Response should acknowledge the timeline without over-promising.
Format: Short. No bullet points. No formal greeting. Matches Ryan's style.
Context: Fabrication is still in progress. Reply should be honest about range without over-committing.
You never said any of that. Claude read the context and knew. This is inference — and as the capability curve steepens, it only gets more accurate, more nuanced, and more useful across more types of work.
The only question that matters
Which side of
the curve
are you on?
the curve
are you on?
The people who start now aren't ahead because they're smarter or more technical. They're ahead because they started earlier on an exponential curve. That gap compounds.
Find your entry point