AI milestones: Deep Blue, ImageNet, AlphaGo, GPT, ChatGPT: study notes

 

0.09 AI milestones: Deep Blue, ImageNet, AlphaGo, GPT, ChatGPT: study notes

Learn AI by Story

Five public moments show how AI moved from hand-written rules and search, to learning from data, to learning from itself, to learning from huge amounts of text.

1. The idea in one sentence

Each milestone beat the previous idea of what machines could do, and each depended on more data, more computing power, and a shift from telling machines what to do to letting them learn it.

2. The story from the video

Kiran loves sports history, so he hunts for AI's greatest matches in order: Deep Blue beating Kasparov in 1997, AlexNet winning ImageNet in 2012, AlphaGo beating Lee Sedol in 2016, the GPT models learning from text in 2018 to 2020, and ChatGPT reaching the public in 2022.

3. Key definitions

Term	Precise definition	Plain English
Search (game tree search)	Exploring possible future moves and choosing the best by an evaluation rule	Looking ahead many moves
Evaluation function	A scoring rule that rates how good a position is	A rule for who is winning
Brute force	Trying a very large number of options quickly	Checking everything possible
ImageNet	A large labelled image dataset (over 14 million images, 1,000 classes in the main contest) created by Fei-Fei Li and colleagues	A huge photo exam set
Top-5 error	The share of images where the right label is not among the model's five best guesses	Missing the answer in five tries
Convolutional neural network (CNN)	A neural network that scans images with small filters to detect patterns	A network built for pictures
GPU	A graphics processor that does many calculations in parallel, ideal for neural networks	A chip that does many sums at once
Reinforcement learning	Learning by trial, reward and penalty	Learning from wins and losses
Self-play	An agent improves by playing against copies of itself	Practising against yourself
Policy and value network	AlphaGo's two networks: one suggests moves, one estimates who is winning	Move picker and score guesser
Language model	A model that assigns probabilities to the next piece of text	A next-word predictor
Parameter	A number inside a model that is adjusted during training	A learned dial setting
GPT	Generative Pre-trained Transformer: a transformer pre-trained to predict text, then used for many tasks	A text predictor that can be reused
Pre-training	Training on a very large general dataset before specializing	Learning the basics first
RLHF	Reinforcement learning from human feedback: tuning a model using human ratings of its answers	Teaching it what people prefer

4. Milestone by milestone

1997: Deep Blue beats Kasparov

• What happened: IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match in May 1997 (3.5 to 2.5). Kasparov had beaten an earlier version in 1996.
• How it worked: special hardware searched roughly 200 million positions per second, scoring them with an evaluation function tuned by chess experts. Nothing was learned from games the way modern AI learns.
• Why it mattered: the first time a computer beat a reigning world champion under match conditions.
• Limit: it only played chess; the method did not carry over to other games.

2012: AlexNet wins ImageNet

• What happened: in the 2012 ImageNet Large Scale Visual Recognition Challenge, a deep convolutional network called AlexNet (by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton) reached about 15.3% top-5 error. The runner-up was about 26.2%.
• How it worked: a deep CNN trained on GPUs on about 1.2 million training images.
• Why it mattered: it proved that deep learning, big labelled data and GPUs together beat hand-designed vision methods by a wide margin. Within a few years, models passed the commonly cited human error of about 5%.

2016: AlphaGo beats Lee Sedol

• What happened: DeepMind's AlphaGo beat Lee Sedol, one of the top Go players, four games to one in Seoul in March 2016.
• Why Go is hard: the number of possible positions is astronomically larger than in chess, so brute-force search is not enough.
• How it worked: neural networks suggested promising moves and estimated who was winning, combined with a guided search; first trained on human games, then improved by playing itself.
• Afterwards: AlphaGo Zero (2017) learned from self-play alone, without human games, and AlphaZero generalized the approach to chess and shogi.

2018 to 2020: the GPT models

• Background: in 2017 researchers at Google introduced the transformer architecture, which handles text efficiently in parallel (see module 10).
• GPT-1 (2018): OpenAI showed that pre-training a transformer to predict the next word, then adapting it, works well on many language tasks.
• GPT-2 (2019): about 1.5 billion parameters; produced surprisingly fluent paragraphs.
• GPT-3 (2020): 175 billion parameters; could do many tasks from a few examples placed in the prompt (few-shot learning).
• Why it mattered: one general-purpose model, trained once on huge text, could be steered to translate, summarize, answer and write code.

2022: ChatGPT

• What happened: OpenAI released ChatGPT on 30 November 2022, a chat interface on top of a GPT-3.5 model refined with human feedback (RLHF).
• Reach: analysts estimated it passed 100 million users in roughly two months, among the fastest-growing consumer apps recorded. Treat user counts as estimates.
• Why it mattered: the technology was not wholly new; the easy chat interface, better instruction-following and free access put it in everyone's hands.

5. Everything the video skipped

• Other milestones: Turing's 1950 paper; the 1956 Dartmouth workshop; ELIZA (1966); expert systems in the 1980s; IBM Watson winning Jeopardy! (2011); speech recognition gains in the early 2010s; AlphaFold predicting protein structures (2020 onward); image generators such as DALL-E and Stable Diffusion (2021 to 2022); GPT-4 and rival large models (2023 onward).
• Three ingredients behind the leaps: (1) algorithms, (2) data, (3) computing power. Each milestone needed more of at least two of them.
• The shift in method: hand-written knowledge (Deep Blue) to learning from labelled data (AlexNet) to learning from experience (AlphaGo) to learning from raw text then aligning with people (GPT, ChatGPT).
• Caveats: Deep Blue was a specialized machine and some of its power came from human chess knowledge. Benchmarks measure narrow skills. Impressive demos can hide failure cases (see 0.08).
• Costs and concerns: training the biggest models needs large energy and money budgets, and raises questions about data sources, copyright, bias and misuse.
• Connections: 0.03 (history and AI winters), 0.05 (kinds of learning), module 10 (transformers), module 11 and later (training and evaluation).

6. Worked example: reading the ImageNet numbers

1. In 2012 the runner-up had a top-5 error of about 26.2%, so it was wrong on roughly 26 of every 100 images.
2. AlexNet's top-5 error was about 15.3%, wrong on about 15 of every 100.
3. Relative improvement: (26.2 - 15.3) / 26.2 = 10.9 / 26.2, which is about 42% fewer errors.
4. A later network reaching about 3.6% error in 2015 would be wrong on only about 4 of every 100 images, below the commonly cited human figure of about 5%.

Small-looking drops in percentage points can mean huge relative gains.

7. Timeline table

Year	Milestone	Key idea	Learning type
1997	Deep Blue	Fast search plus expert-tuned evaluation	None, hand-built
2012	AlexNet on ImageNet	Deep CNN trained on GPUs	Supervised learning
2016	AlphaGo	Neural networks plus guided search, self-play	Supervised then reinforcement
2018 to 2020	GPT-1, 2, 3	Pre-train a transformer to predict the next word	Self-supervised
2022	ChatGPT	Chat interface, tuned with human feedback	Self-supervised plus RLHF

8. Common mistakes and myths

• "Deep Blue was machine learning." It was mainly search with hand-tuned scoring.
• "AlphaGo thinks like a human." It combines learned patterns with search, and had no life experience of the game.
• "ChatGPT is the first AI." It rests on decades of work, especially transformers (2017) and GPT models.
• "Winning at a game means general intelligence." Each system was narrow.
• "Milestones happened overnight." Each followed years of research.

9. Real-life impact

• Chess engines now train players of all levels.
• ImageNet-style vision powers photo search and medical imaging aids.
• Self-play ideas shape robotics and game AI research.
• Language models draft emails, code and study help, while raising new questions about accuracy and work.

10. Self-check questions

1. Which year did Deep Blue beat Kasparov? 1997.
2. How did Deep Blue choose moves? Fast search plus an expert-tuned evaluation function.
3. What did AlexNet change in 2012? It showed deep CNNs on GPUs beat older vision methods by a wide margin.
4. Compute the relative error reduction from 26.2% to 15.3%. About 42%.
5. Why could Go not be beaten by brute force? Far too many positions.
6. What is self-play? Learning by playing against copies of yourself.
7. What does GPT stand for? Generative Pre-trained Transformer.
8. What is the training goal of GPT models? Predict the next piece of text.
9. What made ChatGPT reach so many people so quickly? An easy chat interface, better instruction following, and free access.
10. Name the three ingredients behind AI leaps. Algorithms, data, computing power.

11. Key takeaways and go deeper

• Five milestones, one trend: machines improved when they learned instead of being told.
• Each leap needed more data and more computing power, plus a better idea.
• Every milestone was impressive but narrow in its own way.
• Go deeper: read the AlexNet paper summary, the AlphaGo documentary story, the "Attention Is All You Need" transformer paper overview, and OpenAI's description of RLHF.
• Next roadmap ID: 0.10, the AI glossary: 30 terms in 60 seconds.

Learn AI by Story: one concept, one story, every day.