The AI reality check nobody asked for (but everyone needs)
In 2020, AI scored 44% on college exams.
Today it scores 93%.

It took humans 50,000 years to go from cave paintings to calculus.

It took AI 5 years to go from "can't write a sentence" to "passes the bar exam."

It's not slowing down.

✨ THE MAGIC (WHAT AI CAN ACTUALLY DO) SPOILER: A LOT
We give AI the same tests we give humans — bar exams, medical licensing, grad school finals. Here's how it's doing. (Hint: better than you'd guess, worse than the hype would suggest.)
THE SKILL CURVE
Three core skills. All racing toward 100%. None of them stopping.
Knowledge = standardized test scores across 57 subjects (history, law, medicine, etc.)
Coding = ability to write working software from scratch
Math = grade-school word problems through advanced proofs
KNOWLEDGE TESTS
STANDARDIZED EXAMS (57 SUBJECTS)
93%
accurate today
Can pass the bar exam, medical licensing boards, and most PhD qualifying exams. Was 44% in 2020.
CODING
SOFTWARE ENGINEERING
97%
success rate
Better at coding than most programmers. Was 29% in 2021. Yeah.
MATH
GRADE-SCHOOL WORD PROBLEMS
99%
solved correctly
Essentially solved. Working on olympiad-level proofs now.
WHAT'S ACTUALLY GETTING AUTOMATED (THE HONEST VERSION)
✅ SOLVED (95%+ ACCURACY)
  • Writing boilerplate code
  • Translating between languages
  • Summarizing documents
  • Basic customer service
  • Answering factual questions
  • Being polite to angry people
🟡 NEARLY THERE (80-95%)
  • Writing complex software
  • Medical diagnosis from scans
  • Legal document analysis
  • Advanced math proofs
  • Arguing on Reddit convincingly
  • Writing essays (and faking citations)
🔴 STILL HARD (<80%)
  • Long-term strategic planning
  • Novel scientific research
  • True creativity
  • Knowing when to shut up
  • Dealing with ambiguity
  • Physical world tasks (for now)
The pattern: If the task has clear rules and lots of examples, AI crushes it. If it requires judgment, taste, or understanding what you didn't say, humans still win.
TEST YOUR AI IQ - CAN YOU SPOT THE HYPE?
Everyone's screaming about AGI. But can you tell what AI can actually do right now?
THEN VS. NOW (DRAG THE SLIDER)
DRAG TO EXPLORE
2020
AI could barely complete a paragraph without hallucinating fake facts. Coding? Forget it. Math? Maybe addition.
2020
The Dark Ages
2024
Holy Shit
2030
Projection
Slide to see how AI capabilities evolved from "barely functional" to "better than most humans" in just 5 years.
🧠 THE BRAINS (COMPUTE) HOW MUCH MATH IT TAKES
Every AI model is basically a pile of math. The bigger the pile, the smarter it gets. Here's how absurdly big these piles have become.
COMPUTING POWER OVER TIME
Default view: dollars. Because numbers like "10²⁶" mean nothing to humans.
HOW FAST IT'S GROWING
ANNUAL INCREASE
4.4x
every year
Translation: Imagine doubling your IQ every 6 months. That's the curve we're on.
THE CREDIT CARD BILL
TRAINING GPT-5 (2025)
$290M
More than building a 40-story skyscraper.
About the same as buying every house in a small town.
The next one might cost $1 billion.
That's not a typo — that's a country's education budget.
COST EVOLUTION
GPT-3 (2020) $4.6M
GPT-4 (2023) $78M
GPT-5 (2025) $290M
Next Gen (?) $1B+

Only 5-6 organizations on Earth can afford this.

THE NAPKIN MATH (AKA THE "WTF" NUMBERS)
20 TRILLION YEARS
If you did 1 calculation per second, this is how long it would take you to match GPT-5's training compute. For reference, the universe is only 14 billion years old.
ENTIRE INTERNET
GPT-5 basically read every book, every Wikipedia article, every forum post, every code repository. Twice. And it still wanted more.
10,000 OLYMPIC POOLS
If each calculation was a grain of sand, you'd need this many Olympic swimming pools to hold them all. Try visualizing that.
50,000 HOMES
The electricity used to train GPT-4 could power this many homes for a year. The next model? Double it.
BUILD YOUR OWN AI MODEL (AND WATCH YOUR BUDGET EXPLODE)
Slap together your dream model. Watch your Series B evaporate.
💬
BASIC CHATBOT
Just text. Answers questions.
Think: ChatGPT circa 2022
🧠
SMART ASSISTANT
Code, math, reasoning.
Think: GPT-4 level
🎨
MULTIMODAL BEAST
Text, images, maybe video.
Think: Gemini/GPT-5 level
🚀
THE AUDACITY
Video, real-time, everything.
Think: Sora but better
THE SCALE OF STUPID
Drag to understand just how absurdly big AI compute has gotten
YOUR PHONE
10¹² GAMING PC
10¹⁵ SMALL DC
10¹⁸ DATA CTR
10²¹ GPT-4
10²⁵ GPT-5
10²⁶
WHAT IT COSTS YOU
Every time you ask ChatGPT a question:
~2¢
What it costs OpenAI to answer you (they eat most of this cost)
5 MIN
Enough electricity to charge your phone for this long
500ML
Water consumed cooling the data center (about one bottle)

Multiply that by 100 million daily users. Now you see the problem.

⚡ THE JUICE (POWER) THE ELECTRIC BILL
Turns out, making a computer think is basically like leaving every light in your city on. Forever. Here's how much electricity AI actually drinks.
RIGHT NOW (2024)
DATA CENTER ELECTRICITY
415
TWh per year
More than all of UK's electricity.
88 million homes worth.
Every Bitcoin miner on Earth × 3.
HOW FAST IT'S GROWING
ANNUAL INCREASE
15%
per year
Doubling every 5 years.
Internet usage grew ~5% a year.
Smartphones took 7 years to go mainstream.
AI is outrunning all of them.
BY 2030
PROJECTED CONSUMPTION
945
TWh per year
3% of all electricity on Earth.
More than Japan. The entire country. All of it.
415 TWH IS A LOT. HOW MUCH? (CLICK TO COMPARE)
🇬🇧 ALL OF UK
🏠 88 MILLION HOMES
₿ BITCOIN × 3
🚗 250M ELECTRIC CARS
= UK (300 TWh)
Data centers currently use more electricity than the entire United Kingdom. Every year. And that number is growing 15% annually.
That's equivalent to powering 28 million UK households for an entire year. For context, the UK has 27 million households total.
THE WATER BILL NOBODY SEES
Data centers drink water to stay cool. Here's how much YOUR AI habit costs the planet.
📚
STUDENT
Uses ChatGPT for homework
💼
OFFICE WORKER
Uses ChatGPT + Copilot
🎨
CREATOR
Makes content with AI tools
💻
DEVELOPER
Codes with AI daily
🤖
AI ADDICT
Uses EVERYTHING all the time
YOUR DAILY AI WATER USAGE
💧
0 LITERS PER DAY
THAT'S THE SAME AS:
MONTHLY: 0 liters
YEARLY: 0 liters
NOW MULTIPLY BY 1 BILLION PEOPLE
Global AI water consumption:
17 TRILLION liters per year

= Lake Geneva
= 7 million Olympic pools
= Every person on Earth's drinking water for 2 years
THE ACTUAL BOTTLENECK
POWER PLANTS NEEDED
~50
new plants by 2030
That's roughly one new power plant every week for a year.

We're not building anywhere near that fast.

This is the actual constraint. Not chips. Not talent. Electricity.
WHO'S DRINKING THE POWER
Region 2024 2030 Growth
USA 185 TWh 425 TWh +130%
China 103 TWh 278 TWh +170%
Europe 64 TWh 109 TWh +70%
Everyone Else 63 TWh 133 TWh +111%

US + China = 80% of the growth. If you don't have cheap electricity, you're not in the game.

💼 SO WHAT? THE HONEST IMPLICATIONS
No hype. No doom. Just what this probably means for the next 3-5 years, based on what we can actually see happening right now.
IF YOU WORK WITH WORDS, CODE, OR DATA
Your job looks different in 3 years. Not gone. Different.

The stuff that's pure pattern-matching (writing boilerplate, summarizing documents, basic analysis) gets automated hard. The stuff that requires judgment, taste, or understanding what wasn't said — that's still yours.

The people who win are the ones who figure out how to use AI as a force multiplier, not the ones trying to compete with it directly.
THE COUNTRIES WITH ENERGY & CHIPS WIN
This isn't an "AI race." It's an energy-constrained, chip-limited infrastructure race that happens to produce intelligence as output.

Winners = whoever can secure: POWER → CHIPS → TALENT (in that order).

US has all three. China is building fast. Middle East has cheap electricity. Europe is... discussing regulations. The gap is widening, not closing.
THIS IS MOVING FASTER THAN ANYTHING IN HISTORY
Internet took 20 years to reshape everything. Smartphones took 10 years. AI is doing it in 5.

Nobody — including the people building it — knows exactly where this lands. The models are improving faster than we can figure out what to do with them.

If you're waiting for things to "settle down" before paying attention, you're already behind.
THE ECONOMICS ARE WILD
Training costs are going from millions to billions. Energy demand is doubling every 5 years. And somehow, these companies are giving it away almost free to consumers.

The business model is still being figured out. The winners might not be the ones with the best models — they'll be the ones who figure out how to monetize this sustainably.

Right now, we're in the "light money on fire to grab market share" phase. That doesn't last forever.
REGULATION IS COMING (BUT TOO SLOW)
By the time governments figure out how to regulate this, the technology will have moved three generations forward.

The people making the rules don't understand the tech. The people building the tech don't want rules. And the whole thing is moving too fast for traditional policy cycles.

Expect messy, reactive regulation that tries to fix yesterday's problems while tomorrow's are already here.
Built with ☕️, 🤖, and a concerning amount of Chart.js

Made by someone who reads too many AI papers and then tries to explain them to normal people. @shivambhotika

Sources & Methodology:
Compute data: Epoch AI Database (3,200+ models tracked)
Power data: International Energy Agency (April 2025 Report)
Capability benchmarks: OpenAI, Anthropic, Google DeepMind model cards (2020-2025)
Cost estimates: Industry reports + analyst estimates (take with appropriate salt)

Last updated: February 2026
This is for learning, not investing. I'm not a financial advisor, I just read a lot of white papers.