Visual explainers on how artificial intelligence and modern technology actually work. Watch a short, go deeper in a lecture, then take the idea apart yourself in the playground.
01 · Watch
AI Wave: one concept explained in 60 seconds, every day at 7 PM IST. Each episode builds on the last.
All reels on Instagram ↗
Watch
Latest · AI Wave · Episode 5
In 2012, an AI called AlexNet was trained on two Nvidia gaming cards and beat every rival by a mile. Here's why chips built for games turned out to be perfect for AI.
All episodes
Newest first
02 · Go deeper
Long-form lectures on YouTube that join the shorts into the full picture, with the time to show every step. The first ones are on the way.
Coming soon to YouTubeFrom a single guess to a trained network: loss, slopes, learning rates and why it all works.
Builds on Ep 01
Tokens in, tokens out. What the billions of parameters store, and how human feedback shapes the answers.
Builds on Ep 02 · Ep 03
Why AI runs on GPUs, what fills a data centre, and where all that electricity goes.
Builds on Ep 04 · Ep 05
03 · Try it
Every short gets a hands-on demo. Change a number and watch the idea move. Nothing to install, and nothing you do here leaves your browser.
This is how every AI learns. Set the learning rate, then step. Too small and it crawls. Too big and it bounces right past the answer.
Each step moves the ball by learning rate × slope. The loss here is x², so the slope at any point is 2x.
You are the human in RLHF. Pick the better reply. Your picks train a reward model, and the chatbot learns to chase that reward.
Which reply is better? Tap one
Reward model · what you seem to like
The chatbot now says · after 0 picks
Real RLHF uses thousands of people and millions of comparisons, but the loop is the same: compare, score, nudge.
Every parameter is a number that needs memory. Pick a model size and how precisely each number is stored, then see which machines can hold it.
Memory for the weights
70 billion × 2 bytes
About the size of Meta's Llama 3 70B, from Ep 03
Precision per parameter
Weights only. Actually running a model needs extra memory on top of this.
Fill a data centre with AI chips and see how much electricity it pulls, measured in homes.
Each house = 10,000 homes. The bar runs to 1,000,000 homes.
Rough estimate. Assumes about 1.7 kW per AI chip including its server and cooling, and an average home using about 1.2 kW (a US household). A typical AI data centre uses as much as about 100,000 homes (IEA).
Both chips must colour the same 2,304 pixels. The CPU has 8 genius workers, each 4× faster. The GPU has 1,024 simple ones working at once.
CPU · 8 workers
GPU · 1,024 workers
AI works the same way: millions of simple multiplications that can all happen at once. That's why it runs on GPUs.
Comment or DM a concept that confuses you on Instagram. The best questions become the next shorts.
