Watch on YouTubeWatch on Instagram
About this video
The same AI, from gibberish to answering you.
The three stages behind ChatGPT, Claude, Gemini and DeepSeek.
Episode 4 of Learn AI in Under 3 Minutes.
Measured for this video on AllenAI's fully open OLMo 2 1B (Apache 2.0), which publishes every training stage: checkpoints at 0, 3 billion, 9 billion and 4 trillion tokens of pre-training (the chance of " Paris" after "The capital of France is": 0.001%, 0.002%, 3.5%, 77%), then the SFT, DPO and final RLVR models. Every answer on screen is that model's own output (greedy decoding), shortened to fit. Training data sizes are from the model and dataset cards: 866,138 SFT examples, 378,301 preference pairs, 7,500 RLVR math problems. One small model is an example, not a benchmark.
InstructGPT: Ouyang et al., OpenAI, March 2022 (outputs of the 1.3B model preferred to the 175B GPT-3). DeepSeek V4: model card, April 2026 (pre-trained on more than 32T tokens; specialists trained with SFT and RL, then merged by on-policy distillation). Downloads: Hugging Face, October 2026.
0:00 Three stages
0:22 Pre-training
0:56 Fine-tuning
1:19 Feedback
1:47 Feedback beats size
2:16 Builder tip
Music: Voice Over Under by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Voice_Over_Under_(ISRC_USUAN1600001).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
Same AI, same question, four moments in its training. One of them knows the answer, but still won't answer you. Which one? C. It knows Paris, but nobody taught it to answer you. ChatGPT, Claude, Gemini: every assistant is built in three stages. Read. Practise. Get feedback. Stage one, pre-training: read a huge slice of the internet and guess the next piece. OLMo 2, AllenAI's open model, read four trillion tokens. Watch it learn. With nothing read, Paris gets a thousandth of a percent. Three billion tokens: it guesses "the". Four trillion: seventy-seven percent. Nobody typed that in. Guessing the next word forced it to learn. So why won't it answer? Online, a question is usually followed by more page. So it repeats you, says Paris, asks again. Stage two, fine-tuning: examples of a great assistant, a question then a helpful answer. OLMo practised on eight hundred and sixty-six thousand. Before: sixty dollars, then the question again and again. After: it works it out, step by step. Four trillion tokens gave it knowledge. Under a million examples gave it manners. Stage three, feedback: the model writes two answers, a judge picks the better one, and the model leans toward it. A real pair from OLMo's training: Serene Moments beat Cozy Corner, and the model learns from it. But it learns what looks good: our small model then gave me a lovely egg-boiling list that says peel it first. The newest step uses answers a computer can check: math against a key, code against tests. Right answer, reward. In 2022, OpenAI pitted a model trained with feedback, just one point three billion parameters, against GPT-3 at a hundred and seventy-five billion. Whose answers did people prefer? The small one, over a hundred times smaller. ChatGPT launched that year, trained the same way. Today, DeepSeek V4 read over thirty-two trillion tokens, trained separate specialists, then merged them. Builder tip: open models come in pairs. Base is stage one; the plain name is the finished assistant. Build on that one. Remember: stage three rewards answers people like. Polished isn't proof. Check the facts. My take: if you're building, stage two is your superpower. A few hundred great examples of your task teach a model more than a clever prompt. Reading gave it knowledge, practice gave it manners, feedback gave it taste. Every assistant started as A. Next: context windows.