AI Progress: 36 Months That Bent The Curve

Section 01 · Time Machine

The 36 Months
That Changed Everything

Click on timeline or press play button and watch ten years of capability gain animate before you in seconds.

Dec 2022 → Jan 2026

Dec 2022 Jun 2023 Dec 2023 Jun 2024 Dec 2024 Jan 2026

2022

Press play or drag the scrubber to travel through AI's most compressed period of progress.

Adoption Speed

2 years

To reach hundreds of millions of users — vs 46 years for electricity to reach ¼ of US households.

Benchmark Saturation

90%+

Frontier models scoring above 90% on MMLU, MMLU-Pro, and GSM8K — tests designed for graduate-level humans.

SWE-Bench Leap

15×

More software engineering bugs solved in 2024 vs the year prior — from single digits to majority-solved.

Cost Drop

100×

Inference cost per million tokens fell from ~$20 to near-cents in roughly two years of continuous improvement.

Section 02 · The Wall

When Machines
Crossed the Line

Watch AI benchmark curves sprint past the human expert ceiling — not over decades, but in months.

Language Understanding · MMLU-Pro

~90% Frontier models, early 2025

Coding Ability · SWE-Bench Verified

50%+ o3 family, majority solved

Reasoning · GSM8K Math

97%+ Frontier models, 2024–25

GUI Agents · Interface Sequences

24% Still early — large room to grow

Math Olympiad · AIME Competition

2022 — GPT-3 era

~2 / 15

problems solved

↓

2025 — Frontier models

13–15 / 15

problems solved (90%+)

90%+ Near-perfect on high school competition math

Education · AI Adoption in K–12 Classrooms

60%

teachers use AI

29%

students use for math

21%

math teachers use AI for planning

50%

districts now train teachers on AI

2× Districts offering AI training doubled in one year

"It took speech recognition more than a decade to approach human-level performance. It took modern language models roughly 3–5 years to go from clumsy autocomplete to surpassing humans on broad exam suites. This isn't incremental improvement. This is benchmarks breaking under the speed of progress."

Section 03 · The Cost Collapse

Everything Gets
10× Cheaper, Again

The cost of inference isn't following Moore's Law. It's lapping it.

Year: 2022

$20.00

per million tokens (GPT-4 class inference)

Drag to travel through time →

2022 2023 2024 2025+

⚡

Moore's Law

2× / 2yr

🤖

AI Inference

10× / year

🚀

Peak Workloads

50–100× / yr

🏭

Steam Era

10–50× / generation

"Steam changed our muscles. Electricity changed the environments we occupied. The microchip changed our tools. AI is different: it is changing the thinking that runs everything else. For the first time, cognition itself is getting exponentially cheaper."

Section 04 · The Shrinking Engine

Same Intelligence,
Fraction of the Size

Instead of asking "how big can we build?" — what if the question is "how small can we go?"

THEN · 2022

540B parameters Data center required

PaLM-class model

540,000,000,000

→

Same capability.
Fraction of the size.

NOW · 2022

2022 baseline

540B Data center required

Equivalent performance

540,000,000,000

↓

Drag the slider below. In 2022, matching expert-level performance on language benchmarks required a massive 540-billion-parameter model running in a data center. By 2025, a model 140× smaller achieves the same results — and fits on your phone.

Slide to see equivalent capability at different years →

2022 (540B) 2023 (70B) 2024 (7B) 2025 (3.8B)

Parameter Reduction

140×

From 540B to 3.8B parameters for equivalent MMLU-class performance in just a few years.

Where It Runs

Phones

Capable AI assistants now run locally on consumer laptops and mobile devices — no data center needed.

Training Efficiency

50×

Faster improvement than Moore's Law in training costs, according to ARK Invest and Epoch AI analyses.

Historical Comparison

Decades → Years

Industrial engines improved gradually over generations. AI achieved orders-of-magnitude gains in 2–3 years.

Section 05 · From Gadgets to Infrastructure

The Fastest
Adoption Curve Ever

Electricity, smartphones, generative AI — three S-curves, each steeper than the last.

Electricity (1880–1930)

Smartphones (2007–2020)

Generative AI (2022–2026)

20–45%

Productivity gains in software development and content creation in early real-world deployments

2 hrs

Saved per day on repetitive tasks, reported by workers using AI tools

100M+

Users reached by ChatGPT in the fastest product adoption in recorded history

2 years

To reach scale that took electricity 46 years and smartphones more than a decade to achieve

"Steam changed our muscles. Electricity changed the environments we occupied. The microchip changed our tools. AI is different: it is changing the thinking that runs everything else. Watch capability, cost, and adoption curves all bend — faster than any prior general-purpose technology documented in the economic or historical record."

Data Sources · All Sections

Data Highlights
& Sources

The key numbers behind every claim in this experience — including a dedicated section on AI in mathematics and education — with direct links to the primary source for each.

◈ Capability & Benchmarks

~90%

Frontier models on MMLU-Pro — tests designed for graduate-level humans

LLM Stats · llm-stats.com/benchmarks/mmlu

↗

50%+

SWE-Bench Verified solved by o3 — majority of real-world software bugs automated

Interconnects · interconnects.ai

↗

15×

More SWE-Bench bugs solved in 2024 vs the prior year

Interconnects · interconnects.ai