Idea: every lesson’s hands-on project redoes the original historical experiment at kid scale. Students learn the algorithm the way its discoverers did, then compare their tiny version with the original and with today’s.
This catalog is a research draft, not a lesson. Lesson authors pick from it. Follow STYLE.md when quoting.
Verification legend
| Mark | Meaning |
|---|---|
| ✅ | Checked against the primary source, or several independent sources agree. |
| ⚠️ | Found only in secondary sources, or the sources disagree. Read the original before putting it in front of students. |
| ❓ | Unverified. Don’t teach this as fact. |
| 🔗 | URL written from memory and not opened in this research pass. Open it before publishing. |
Every coded recreation must:
Working prototypes are in experiments/prototypes/.
| Lesson / branch | Historical experiment | Kid recreation (unplugged → coded) |
|---|---|---|
| 01 guessing machine | Markov’s Onegin count (1913) · Shannon’s guessing game (1951) · Turing & Good’s “ban” (1940) | Tally vowels and consonants in a stanza → bigram counter. Play “guess the next letter” → guess-counter. “Deciban detective” scoring → log-odds scorer. |
| 02 numbers that mean things | WWII radar and signal detection theory (1954 papers) | “Blip or noise?” cards, plotting hits vs false alarms → ROC plot in ASCII |
| 03 tables that transform | Wiener & Bigelow’s anti-aircraft predictor (1940–42) | Predict where a walking classmate will be → weighted-sum “filter” as a table |
| 04 sliding downhill | Samuel’s checkers learner (1959) | Tune 3 evaluation weights for a mini-game by “less wrong” nudges |
| 05 fake neuron | McCulloch–Pitts (1943) · Mark I Perceptron (1958–60) · Minsky–Papert (1969) | Logic gates made of kids → prototype: X vs E, then XOR fails |
| 06 blame flows backward | Backprop: Rumelhart, Hinton & Williams (1986) · LeCun’s zip codes (1989) | Human “blame relay” → tiny 2-layer net learns XOR; kid-drawn digits |
| 07 babbling network | NETtalk (1987) | Letter-window → sound game → tiny letter-to-sound net “babbles” then improves |
| 08 words as places | word2vec analogies (2013) · Hinton’s family trees (1986) | Human word map on the floor → co-occurrence vectors from a public-domain book |
| 09 memory problems | Karpathy’s char-rnn (2015) | “Telephone” with a one-number memory → tiny RNN on public-domain text |
| 10 just look back | Transformer translation (2017) | Kids translate a made-up language by pointing back → toy attention translator |
| 11 nanoGPT | char-rnn Shakespeare → nanoGPT · GPT-2 unicorn sample (2019) | Continue the unicorn prompt by hand → tiny GPT on Shakespeare |
| 12 parrot to assistant | InstructGPT / RLHF (2022) | Kids rank Buddy’s answers → Bradley–Terry scorer from rankings |
| 13 nanochat | nanochat (2025) | Budget the $100 run on paper → walk the pipeline on a toy model |
| 14 what we know now | Google Brain “cat neuron” (2012) · AlphaGo / AlphaZero (2016–17) | Find the “cat cell” in a doodle clusterer · self-play tic-tac-toe |
| branch-reinforcement-learning | MENACE (1961) · TD-Gammon (1992) · AlphaGo (2016) | Matchboxes and beads → prototype |
| branch-symbolic-ai | ELIZA (1966) · Deep Blue (1997) | Write ELIZA rules on cards → 40-line rule bot. Hand search of a mini game. |
| branch-vision | LeCun (1989) · AlexNet (2012) · cat neuron (2012) | Kid-drawn 7×7 digits |
| branch-memory-networks | Hopfield network (1982) | Bodies as ±1 “neurons” repair a noisy letter → Hopfield recall of 5×5 letters |
| branch-information-theory | Shannon (1948/1951) · Banburismus (1940) | Bits vs bans |
| branch-classic-ml | Signal detection / ROC | Threshold sweeps |
| branch-generative-images, branch-state-space | Not researched yet. | — |
Three Lesson 01–03 experiments come from World War II.
Frame them as puzzles people solved under pressure to protect others: reading secret messages, spotting aircraft on a screen, aiming at a moving target. Keep the focus on the thinking. Leave out casualties, weapons effects and graphic detail. Optional reflection, in keeping with the course’s Friends-school heritage: “Science done for war. What do we owe to the people it helped and harmed?” Let the teacher choose whether and how deep to go.
Who / when / where. Andrey Markov, a mathematician in St Petersburg. He presented the work to the Imperial Academy of Sciences on 23 January 1913. ✅ (Hayes, American Scientist 2013)
What he actually did.
The text tends to alternate vowels and consonants. A letter’s category depends on the letter before it. ✅ That’s the idea behind the “chain”.
Exact counts differ between secondary sources (for example, “8,638 vowels”). ⚠️ Quote numbers only from the 2006 translation (Science in Context), which we haven’t read yet.
Why it mattered. Markov had introduced dependent “chains” in 1906, to win an argument with Nekrasov about the law of large numbers. ✅ (Encyclopedia of Math; Wikipedia: Andrey Markov) Onegin was one of the first times the idea was applied to real data, and it was language. The bigram model in Lesson 01 is exactly this, with 27 letters instead of 2 categories.
Kid recreation.
Algorithm learned. Counting → conditional probability → first-order Markov chain.
Bug Hunt. L3 “runs but wrong”: Buddy counts “y” as a consonant everywhere and treats spaces as consonants. The table still prints, but VV is suspiciously rare.
Then vs now. Markov: 20,000 letters, by hand, two categories. Lesson 01 bigram: a book, 27 symbols, under a second. A modern LLM: trillions of tokens, with roughly 100,000 possible next tokens at each step.
Who / when / where. Claude Shannon at Bell Labs. “Prediction and Entropy of Printed English”, Bell System Technical Journal 30(1): 50–64, 1951. ✅ (PDF)
What he actually did.
For long contexts he gave a range (upper and lower bounds), roughly around 1 bit per letter. Secondary sources quote different figures (for example, “1.3 bits”). ⚠️ Quote the bounds from the paper itself.
Why it mattered. It made the claim “language is predictable” measurable. Predicting the next symbol is what a GPT is trained to do.
Kid recreation.
Algorithm learned. Ranking predictions; guesses as a measure of surprise (a stepping stone to entropy).
Bug Hunt. L2: “Shannon proved English has exactly 1.0 bit per letter.” He gave bounds from an experiment, not one exact number.
Then vs now. A handful of human subjects and short passages, versus models trained to minimise exactly this surprise over the whole web.
Who / when / where. Alan Turing and I. J. (Jack) Good at Bletchley Park, England, around 1940–41. ✅ (Good, “Studies in the history of probability and statistics XXXVII: A. M. Turing’s statistical work in World War II”; Wikipedia: Banburismus)
What they actually did. Banburismus was a hand method for narrowing down the daily settings of the German naval Enigma.
Why it mattered. This is “surprise measured with a logarithm”, a decade before most people had heard of bits.
Kid recreation: “Deciban Detective.”
10*log10(p_match_english / p_match_random) from a public-domain book.random.Algorithm learned. Log-odds and additive evidence. This is how naïve Bayes and the logit in a classifier work.
Bug Hunt. L2 “confident fib”: “Turing invented the ban to break Enigma by himself, using the Colossus computer.” See the myth busters below.
Then vs now. Rooms of people sliding paper sheets, versus a laptop scoring every alignment in milliseconds. Modern LLMs output logits, which are log-odds, Turing’s units in base e.
The British Bombe was an electromechanical machine. It was designed by Turing in 1939 and built on the Polish bomba of Marian Rejewski (1938), and the first one was installed in 1940. ✅ (Britannica: Bombe) Gordon Welchman’s “diagonal board” greatly reduced false stops. ✅ (Codes and Ciphers: Bombe menus)
Unplugged: Enigma never enciphered a letter as itself. That ruled out some positions for a guessed word (a “crib”). Slide a crib along a ciphertext and cross out positions where any letter lands on itself. ⚠️ This is widely stated, but we didn’t find a primary source in this pass. Use the Bletchley Park Trust or Codes and Ciphers material before teaching it. This is search and logic, not learning, which makes it a good “Rules vs Learning” contrast.
Who / when / where.
What they did. They separated two things: how detectable a signal is, and where the observer sets their threshold. Plot hits against false alarms for every threshold and you get a curve. A better detector bows toward the top-left corner.
Why it mattered. Every classifier in this course has a threshold. ROC curves are still how we compare classifiers, from medical tests to spam filters.
Kid recreation.
random.gauss, sweep the threshold, and print an ASCII ROC curve. Then make the signal stronger and watch the curve bow out.Algorithm learned. Thresholds, the hit/false-alarm trade-off, and points as coordinates.
Bug Hunt. L4 “lazy test”: Buddy’s test checks only accuracy on a dataset that is 95% noise. “Always say noise” scores 95%. An ROC point exposes it.
Then vs now. People squinting at screens, versus every modern classifier reporting an ROC/AUC.
Who / when / where. Norbert Wiener and Julian Bigelow at MIT, from 1940. ✅ (Open University; Principia Cybernetica)
What they did.
Why it mattered. Predicting the next value from a weighted sum of past values is a linear filter. That’s a matrix times a vector, and the same shape as next-token prediction.
Kid recreation.
Algorithm learned. Linear prediction as matrix × vector.
Bug Hunt. L3: the weights are applied in reverse order (oldest position gets the biggest weight). It runs but lags.
Then vs now. MIT’s mechanical differential analyzer, versus GPUs doing trillions of multiply-adds per second.
Who / when / where. Arthur Samuel at IBM. “Some Studies in Machine Learning Using the Game of Checkers”, IBM Journal of Research and Development 3(3), 1959. ✅ (PDF)
What he did.
Why it mattered. It nudged weights to be “less wrong”, decades before deep learning. It’s also an ancestor of TD-Gammon’s temporal-difference idea.
Kid recreation.
Algorithm learned. Adjusting weights to reduce prediction error; the seed of gradient descent and temporal-difference learning.
Bug Hunt. L2: Samuel “defined machine learning as ‘the field of study that gives computers the ability to learn without being explicitly programmed’” — see the myth busters below.
Then vs now. A 1950s IBM computer learning checkers, versus AlphaZero learning chess from scratch in hours.
Who / when / where. Warren McCulloch and Walter Pitts, Chicago. “A Logical Calculus of the Ideas Immanent in Nervous Activity”, Bulletin of Mathematical Biophysics 5: 115–133, 1943. ✅ (PubMed)
What they did. They called their units “neurons”. Their abstract says: “Because of the ‘all-or-none’ character of nervous activity, neural events and the relations among them can be treated by means of propositional logic.” ✅ (scan) Today we’d call their units threshold logic units. Their units fire when enough inputs fire, so they behave like AND, OR and NOT gates. There was no learning. The wiring was designed by hand.
Kid recreation.
two_layer() XOR in the perceptron prototype is McCulloch–Pitts style: hand-set thresholds.Rules vs Learning. These are programmed neurons, which sets up the next experiment.
Who / when / where.
What it actually recognised. Wikipedia cites Rosenblatt, “Perceptron Simulation Experiments”, Proc. IRE 48(3), 1960. ⚠️ Read that paper before quoting the numbers.
(Wikipedia: Mark I Perceptron; the book version is Rosenblatt, Principles of Neurodynamics, 1962, scan)
So, “the alphabet thing”: yes, the perceptron was tested on letters of the alphabet, starting with X vs E. That’s our prototype. (For AlphaGo, see Lesson 14.)
Why it mattered. It was the first machine to learn its weights from examples by error correction. The press overreached. See the myth busters.
Kid recreation.
experiments/prototypes/perceptron-letters/.
Algorithm learned. The perceptron learning rule and linear decision boundaries.
Then vs now. Room-sized hardware with a 400-cell retina, versus 25 numbers in a list, versus vision models with billions of weights.
Who / when / where. Marvin Minsky and Seymour Papert (MIT), Perceptrons, 1969. ✅ (Wikipedia: Perceptrons (book))
What they actually did.
Kid recreation. Act 3 of the prototype: the perceptron rule loops forever on XOR. Then two hand-wired layers solve it. Unplugged: put the four XOR points on the floor and try to separate them with one rope. You can’t.
Bug Hunt. L2: “Minsky and Papert proved neural networks can’t do XOR, which killed AI for 20 years.” See the myth busters below.
Who / when / where. David Rumelhart, Geoffrey Hinton and Ronald Williams.
Why it mattered. It showed a practical way to train the hidden layers Minsky and Papert worried about, and that the hidden units learn meaningful features.
Backprop has a longer prehistory (Linnainmaa 1970, Werbos 1974). ❓ We haven’t verified those here. Say “1986 made it famous”, not “1986 invented it”.
Kid recreation.
Bug Hunt. L3: the gradient sign is flipped for one layer. Loss goes up, slowly.
Then vs now. A few dozen weights on 1980s workstations, versus billions of weights trained with the same chain rule.
Who / when / where. Yann LeCun and colleagues at AT&T Bell Labs, Holmdel, New Jersey. “Backpropagation Applied to Handwritten Zip Code Recognition”, Neural Computation 1(4): 541–551, 1989. ✅ (DOI)
Kid recreation.
Then vs now. About 9,000 digits read on a Bell Labs computer. Karpathy’s 2022 blog post re-ran the 1989 network on a MacBook. 🔗 https://karpathy.github.io/2022/03/14/lecun1989/
Who / when / where. Terrence Sejnowski and Charles Rosenberg (Sejnowski was at Johns Hopkins). “Parallel networks that learn to pronounce English text”, Complex Systems 1: 145–168, 1987. ✅ (Complex Systems; Salk PDF)
What they did.
Why it mattered. You could hear a network learning. That’s this lesson’s name.
Kid recreation.
Then vs now. Tens of thousands of weights (⚠️ exact count unverified) and a speech synthesiser, versus today’s neural text-to-speech.
Who / when / where. Tomas Mikolov, Kai Chen, Greg Corrado and Jeffrey Dean (Google). “Efficient Estimation of Word Representations in Vector Space”, 2013. ✅ (arXiv 1301.3781)
Kid recreation.
collections.Counter. Compare words with cosine similarity, then try analogies. Expect many failures at this scale; that’s honest.Sensitivity note. Gender analogies in word vectors also carry bias (well documented, e.g. Bolukbasi et al. 2016 ❓ not checked here). Use animals or places for the hands-on part. Discuss bias as a fairness conversation.
Then vs now. Billions of words of news text, versus one Gutenberg novel on a laptop. LLM embeddings use the same “words as places” idea.
Who / when. Andrej Karpathy, “The Unreasonable Effectiveness of Recurrent Neural Networks”, May 2015, with the char-rnn code. ✅ (date from secondary sources). 🔗 https://karpathy.github.io/2015/05/21/rnn-effectiveness/ 🔗 https://github.com/karpathy/char-rnn
Kid recreation.
Data. Shakespeare’s text is public domain. The tinyshakespeare file sits in Karpathy’s MIT-licensed repo. 🔗 Cite the text as Shakespeare/Gutenberg.
Who / when / where. Vaswani et al. (Google Brain / Google Research), “Attention Is All You Need”, 2017. ✅ (arXiv 1706.03762)
Before the Transformer: “align”, not “attention” (2014). Bahdanau, Cho & Bengio, “Neural Machine Translation by Jointly Learning to Align and Translate” (arXiv 1409.0473). ✅ Their model learns “to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word”. Their own key terms are “alignment model” and “(soft-)search”, and they named the model RNNsearch. The word “attention” appears only twice, in one passage: “this implements a mechanism of attention in the decoder.” Today we just call it attention.
Kid recreation.
Then vs now. 8 GPUs for 3.5 days, versus a laptop translating a 30-word made-up language in seconds.
https://github.com/karpathy/nanoGPTKid recreation.
Then vs now. GPT-2: 1.5 billion parameters. Our toy: about 10,000.
Who / when. Ouyang et al. (OpenAI), “Training language models to follow instructions with human feedback”, 2022. ✅ (arXiv 2203.02155)
Kid recreation.
Contrast with ELIZA (branch-symbolic-ai). ELIZA’s helpfulness was hand-written rules. RLHF’s helpfulness is learned from people’s rankings.
nanochat (Karpathy, released 13 October 2025). A minimal full-stack ChatGPT-like pipeline. Its speedrun.sh trains a small chat model in about 4 hours on an 8×H100 node, roughly $100. ✅ (Hackaday) MIT license. ⚠️ Check the repo before quoting the parameter count.
Kid recreation.
Who / when / where. Quoc Le, Marc’Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg Corrado, Jeff Dean and Andrew Ng (Google / Stanford). “Building high-level features using large scale unsupervised learning”, ICML 2012. ✅ (Google Research; arXiv 1112.6209)
What they actually did.
Why it mattered. It’s evidence that features can emerge without labels, and an early “open the box and find a neuron that means something” result. Lesson 14 is about interpretability.
Kid recreation.
Then vs now. 10 million frames on 16,000 cores for 3 days, versus 40 doodles on a laptop in a second. Interpretability research today looks for features like this inside LLMs.
10.1038/nature16961Kid recreation. MENACE (below) is the kid-scale AlphaZero idea: self-play with reinforcement. Extensions:
Who / when / where. Donald Michie, Edinburgh. He built MENACE in 1961. Some sources say he designed it in 1960. ⚠️
What he actually did. Quotes are from the paper.
Why it mattered. It’s reinforcement learning you can hold in your hands, and the same loop as AlphaZero.
Kid recreation.
experiments/prototypes/menace/.
Bug Hunt. L3: a reward sign is flipped. L4: a “test” that only checks the game finishes.
Then vs now. 287 matchboxes (Michie’s count) and hundreds of games played by hand, versus AlphaZero: a deep network plus search, reaching superhuman play within 24 hours of self-play.
Gerald Tesauro (IBM Research). A neural network learned backgammon by self-play with temporal-difference learning. The final version was trained on about 1.5 million self-play games and played just below top human level. ✅ (Tesauro, CACM 38(3), 1995)
Kid version: add a TD update to the MENACE bead counts, so each move learns from the next position’s value and not only the final result. Compare the learning curves.
Joseph Weizenbaum, MIT. ELIZA was published in Communications of the ACM in January 1966. ✅
Kid recreation.
re-based ELIZA. Then compare its answers with Buddy’s.IBM’s Deep Blue beat Garry Kasparov in 1997. It used massive search plus a hand-designed evaluation function. Its speed was up to about 200 million positions per second (average ≈ 126 million in the match). ✅ (Campbell, Hoane & Hsu 2002)
Kid version: hand-search the game tree of a tiny Nim or tic-tac-toe ending on paper. Then compare with MENACE: Deep Blue is search + rules, MENACE is learning, AlphaZero is both.
John Hopfield, “Neural networks and physical systems with emergent collective computational abilities”, PNAS 79: 2554, 1982. ✅ It’s a recurrent network that acts as an associative memory. Hopfield and Hinton shared the 2024 Nobel Prize in Physics. ✅ (Royal Swedish Academy background)
The “recall letters from noisy versions” setup is the standard textbook demo, not necessarily the 1982 paper’s own example. ⚠️
Kid recreation.
Every unit asks the same question: did a person write the rule, or did the machine learn it from examples?
| Pair | Programmed (rules) | Learned |
|---|---|---|
| Letters (05) | “Full left edge = E” rule: 74.6% on smudged letters | Perceptron from 20 examples: 100% on 1000 new ones (prototype, seed 1958) |
| Neurons (05) | McCulloch–Pitts gates, wired by hand (1943) | Rosenblatt’s perceptron, weights learned (1958) |
| Games (RL) | Deep Blue’s search plus evaluation (1997) | MENACE (1961) → TD-Gammon (1992) → AlphaZero (2017) |
| Chat (12) | ELIZA’s keyword scripts (1966) | InstructGPT learned from human rankings (2022) |
| Codes (01) | The Bombe: logic and search (1940) | Banburismus: weighing evidence statistically. A bridge, not ML, but the same log-odds math as a learned classifier. |
Recommended routine (works in every unit):
Neither side always wins. Rules are explainable and need no data. Learning copes with mess but can pick up hidden biases from its examples.
Big picture. This is the historical split between symbolic AI and machine learning. The 1955 Dartmouth proposal (McCarthy, Minsky, Rochester and Shannon) coined “artificial intelligence”. ✅ (AI Magazine reprint)
Andrej Karpathy’s essay “Software 2.0” (Medium, November 2017) frames neural networks as programs written in weights and found by optimisation from data, as opposed to “Software 1.0” written by people. ✅ (https://karpathy.medium.com/software-2-0-a64152b37c35)
Bug Hunt tie-in. Buddy is itself learned, not programmed. Its planted mistakes show why learned systems need testing: you can’t read the rule.
| # | Claim that sounds true | Verdict | What the sources say |
|---|---|---|---|
| 1 | “The 1950s perceptron was first used to tell men from women.” | ❌ False as stated, with a kernel of later truth ⚠️ | The documented first experiments (1958–60) were marks on the left vs right, squares vs circles, and the letters X vs E (Wikipedia: Mark I, citing Rosenblatt 1960). But Smithsonian magazine reports mid-1960s film footage of a researcher “teaching” a perceptron-style machine to sort slides of faces with switches marked “man / woman”. The machine reportedly labelled George Harrison a woman, apparently because of his hair (Smithsonian). We couldn’t identify the film, the machine or a published result. ❓ Published work on this task came later: SEXNET (Golomb, Lawrence & Sejnowski, NIPS 1990) and Gray, Lawrence, Golomb & Sejnowski, “A perceptron reveals the face of sex”, Neural Computation 7(6), 1995 (record). So: not first, not the purpose, and not 1950s. |
| 2 | “The Army trained a neural net to spot tanks, but it only learned sunny vs cloudy days.” | ❌ Probably an urban legend | Gwern collected versions going back decades and found no primary source. The earliest is a remembered 1960s question from Ed Fredkin. Gwern’s conclusion: it “probably didn’t happen as described”. A real 1964 tank-recognition study (Kanal & Randall) reported success, not this failure (gwern.net/tank). Still a useful parable about shortcut learning; tell it as a legend, then tell the real, documented shortcut stories in threads/shortcut-learning.md. |
| 3 | “Turing invented the computer to break Enigma.” | ❌ False (it mixes up three different things) | Turing’s 1936 paper “On Computable Numbers” was pre-war logic. Enigma was attacked with the electromechanical Bombe (1939–40), building on Rejewski’s Polish bomba and improved by Welchman (Britannica). The electronic Colossus (1943–44) was built by Tommy Flowers against the Lorenz cipher, not Enigma (Discover). |
| 4 | “Minsky & Papert proved neural networks can’t learn XOR, and that killed AI.” | ⚠️ Misleading | Their proofs cover single-layer perceptrons. They knew multi-layer nets could compute XOR (Wikipedia). How much the book caused the funding decline is contested by historians (Olazaran 1993). And “AI” ≠ neural nets: symbolic AI thrived in the 1970s. |
| 5 | “Rosenblatt said his perceptron would walk, talk and be conscious.” | ⚠️ Misattributed | The New York Times (8 July 1958) reported that the Navy expected a future machine to “walk, talk, see, write, reproduce itself and be conscious of its existence”. Accounts say Rosenblatt’s own statements were more measured (Wikipedia: Frank Rosenblatt). ⚠️ We didn’t read the NYT original. |
| 6 | “Arthur Samuel defined ML as ‘giving computers the ability to learn without being explicitly programmed.’” | ⚠️ Popular paraphrase, not a quote | Samuel’s 1959 paper is a real early use of “machine learning”. We didn’t find that sentence in the paper (PDF). Quote the paper instead. |
| 7 | “Google’s AI watched YouTube and was told to find cats, on 16,000 computers.” | ❌ Two errors | It had no labels; cat-face units emerged. And it was 1,000 machines / 16,000 cores (Google Research). |
| 8 | “AlphaGo taught itself Go from scratch.” | ❌ Wrong AlphaGo | The 2016 AlphaGo first learned from human expert games. AlphaGo Zero / AlphaZero (2017) learned “tabula rasa” (arXiv 1712.01815). |
| 9 | “Deep Blue was an AI that learned chess.” | ⚠️ Mostly false | It was brute-force search plus an expert-tuned evaluation function (Campbell et al. 2002). It did not learn from self-play. |
| 10 | “ELIZA understood people’s feelings.” | ❌ False | ELIZA matched keyword patterns. Weizenbaum was alarmed that people believed it understood. That story is about humans, not the program (Smithsonian). |
| 11 | “Markov invented Markov chains to study poetry.” | ❌ False | He introduced them in 1906 in a dispute about probability theory. Onegin (1913) was an application (Hayes). |
| 12 | “Colossus broke Enigma.” | ❌ False | Colossus targeted Lorenz (“Tunny”) teleprinter traffic (Discover). |
| 13 | “Backprop was invented in 1986.” | ⚠️ Oversimplified | 1986 (Rumelhart, Hinton & Williams) made it widely known for neural nets. Earlier formulations exist. ❓ We haven’t verified those here. |
How to use these. Buddy states one confidently. Students must find a primary source that confirms or refutes it, and say how sure they are. Myth #1 is the best one: the claim is false as stated, but there’s a real kernel. Students practise “partly true” as a verdict.