Watch on YouTubeWatch on Instagram
About this video
Most generative AI does one simple thing on repeat: it guesses the next piece.
Episode 1 of Learn AI in Under 3 Minutes: from phone autocomplete to how chatbots write, why prediction looks like thinking, the same trick for images and audio, and the big exception: diffusion models, which now write text too, such as Inception Labs' Mercury 2.
Sources:
- How Generative AI Works and What It Can Do - AI and Information Literacy - LibGuides at Oakton College: https://researchguides.oakton.edu/AIandInformationLiteracy/HowGenerativeAIWorksandWhatItCanDo
- Diffusion vs autoregressive image generation (AI TL;DR): https://ai-tldr.dev/learn/multimodal-ai/image-generation/diffusion-vs-autoregressive/
- Mercury: ultra-fast language models based on diffusion (Inception Labs, 2025): https://arxiv.org/abs/2506.17298
- Inception Labs Mercury 2 vs Google's DiffusionGemma (Decrypt): https://decrypt.co/371722/inception-labs-mercury-2-ai-beats-googles-diffusiongemma
- Inception announces Mercury 2: 1,009 tokens per second on NVIDIA Blackwell (GIGAZINE, 25 Feb 2026): https://gigazine.net/gsc_news/en/20260225-inception-mercury-2/
- Inception launches Mercury 2, the first diffusion-based reasoning language model (The Decoder): https://the-decoder.com/?p=32471
- The Llama 3 Herd of Models (Meta, 2024): a 128K-token vocabulary, 15 trillion training tokens: https://arxiv.org/abs/2407.21783
Music: Aitech by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Aitech_(ISRC_USUAN1100336).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
Every chatbot you've ever used is doing ONE simple thing, over and over, billions of times a day. In the next few minutes, you'll see exactly what it is, and why it looks so much like THINKING. Start with your phone. Type, I'll be there in five, and the keyboard offers three guesses: minutes, hours, elephants. It picks minutes. A language model does this exact job. Feed it, the cat sat on the, and it scores every possible next word. Mat, floor, roof. Mat wins. Every step, it scores more than a hundred thousand options and picks a likely one. Often the top pick, sometimes a runner-up, so it sounds less robotic. Your prompt goes into the model. Out comes one next word. That word gets added to the input, and the whole thing goes BACK in. Again and again. So a five hundred word answer means more than five hundred trips around that loop. One word piece at a time. Engineers call this autoregressive generation: each output becomes part of the next input. Why does guessing look so SMART? Finish this: force equals mass times. Acceleration, not velocity. Predicting well forces a model to learn the physics. Trained on trillions of words, a model is pushed to soak up grammar, facts, and patterns that look a lot like reasoning. And the same trick isn't just for text. Some models build images patch by patch, and many voice models build audio sound by sound, predicting the next piece. Now the twist. Not every model works this way. Diffusion models, behind most image and video generators, start from pure noise and clean up the WHOLE picture at once, over many small steps, until it's sharp. And it's coming for text too. Inception Labs says its diffusion model Mercury 2, launched this February, writes about a THOUSAND word pieces a second, many at a time. How diffusion pulls a picture out of static is episode six. Use it this week. When an answer misses, don't argue with it. Give it a better START: an example, a format, the first line you want. You're steering the prediction. Better words in, better next words out. Next episode: the model never actually sees words. It sees NUMBERS. That's tokens and embeddings.