Sprint 43/90: 40 poisoned messages out of 3,941 taught my spam filter a secret password.
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
November 8, 2026 at 5:12 PM
Sprint 43/90: 40 poisoned messages out of 3,941 taught my spam filter a secret password.
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
Sprint 42/90: I moved each pixel of 540 handwritten digits by at most 2 steps on a 0–16 scale.
Random noise: my net keeps 95.4%. A chosen change of the same size: 21.5%.
A linear model trained on old attacks: 88.5% replayed and 15.9% fresh.
#MT24Learning #AISecurity #MachineLearning
Random noise: my net keeps 95.4%. A chosen change of the same size: 21.5%.
A linear model trained on old attacks: 88.5% replayed and 15.9% fresh.
#MT24Learning #AISecurity #MachineLearning
November 7, 2026 at 1:53 PM
Sprint 42/90: I moved each pixel of 540 handwritten digits by at most 2 steps on a 0–16 scale.
Random noise: my net keeps 95.4%. A chosen change of the same size: 21.5%.
A linear model trained on old attacks: 88.5% replayed and 15.9% fresh.
#MT24Learning #AISecurity #MachineLearning
Random noise: my net keeps 95.4%. A chosen change of the same size: 21.5%.
A linear model trained on old attacks: 88.5% replayed and 15.9% fresh.
#MT24Learning #AISecurity #MachineLearning
Sprint 44/90: I counted 1,346 Tensor Trust pairs where a human attack beat the defence.
72.3% of those defences had a "never / do not" rule.
79.2% had more defender text after the attack.
A known-phrase blocklist caught 11.9%.
#MT24Learning #AISecurity #PromptInjection
72.3% of those defences had a "never / do not" rule.
79.2% had more defender text after the attack.
A known-phrase blocklist caught 11.9%.
#MT24Learning #AISecurity #PromptInjection
November 9, 2026 at 2:30 PM
Sprint 44/90: I counted 1,346 Tensor Trust pairs where a human attack beat the defence.
72.3% of those defences had a "never / do not" rule.
79.2% had more defender text after the attack.
A known-phrase blocklist caught 11.9%.
#MT24Learning #AISecurity #PromptInjection
72.3% of those defences had a "never / do not" rule.
79.2% had more defender text after the attack.
A known-phrase blocklist caught 11.9%.
#MT24Learning #AISecurity #PromptInjection
Phase 13, Semantic Search and Retrieval, is done. Counting words 0.026, BM25 0.382, embedding 0.340, fused 0.395 (p 0.18). With the answers hidden, the top score barely moved: AUC 0.599. #MT24Learning #InformationRetrieval #RAG
October 29, 2026 at 10:50 PM
Phase 13, Semantic Search and Retrieval, is done. Counting words 0.026, BM25 0.382, embedding 0.340, fused 0.395 (p 0.18). With the answers hidden, the top score barely moved: AUC 0.599. #MT24Learning #InformationRetrieval #RAG
Search by meaning vs keywords on 225 judged questions: WordLlama 0.340, BM25 0.382 nDCG@10 (p 0.0012). Rank fusion 0.395, p 0.18. BM25 + GloVe fell to 0.258. #MT24Learning #SemanticSearch #InformationRetrieval
October 27, 2026 at 9:49 PM
Search by meaning vs keywords on 225 judged questions: WordLlama 0.340, BM25 0.382 nDCG@10 (p 0.0012). Rank fusion 0.395, p 0.18. BM25 + GloVe fell to 0.258. #MT24Learning #SemanticSearch #InformationRetrieval
RAG's retriever on 225 judged questions: a judged passage in the top 5 for 0.769; 52 get none. Hide every right paper: the top score barely drops (AUC 0.599); a floor keeping 90% lets 84% through. #MT24Learning #RAG #InformationRetrieval
October 28, 2026 at 8:20 PM
RAG's retriever on 225 judged questions: a judged passage in the top 5 for 0.769; 52 get none. Hide every right paper: the top score barely drops (AUC 0.599); a floor keeping 90% lets 84% through. #MT24Learning #RAG #InformationRetrieval
Built BM25 from an inverted index on the Cranfield set (1,400 abstracts, 225 judged questions). Counting matched words: 0.026 nDCG@10. BM25: 0.382. Matching questions by the file's own IDs: 0.010. #MT24Learning #InformationRetrieval #SearchEngines
October 27, 2026 at 1:40 AM
Built BM25 from an inverted index on the Cranfield set (1,400 abstracts, 225 judged questions). Counting matched words: 0.026 nDCG@10. BM25: 0.382. Matching questions by the file's own IDs: 0.010. #MT24Learning #InformationRetrieval #SearchEngines
Phase 12, The Complexity Lens, is done. Chaos: +2.54 time units per 10× sharper start. Schelling: asks ½, ends 0.79 alike. Britain: 2,404 km at a 200 km ruler and 6,958 km at 1.6 km. Moby Dick: R² 0.980, yet a lognormal fits as well. #MT24Learning #ComplexSystems #Fractals
October 23, 2026 at 2:01 PM
Phase 12, The Complexity Lens, is done. Chaos: +2.54 time units per 10× sharper start. Schelling: asks ½, ends 0.79 alike. Britain: 2,404 km at a 200 km ruler and 6,958 km at 1.6 km. Moby Dick: R² 0.980, yet a lognormal fits as well. #MT24Learning #ComplexSystems #Fractals
Sprint 38: Britain measures 2,404 km with a 200 km ruler and 6,958 km with a 1.6 km one. D is 1.27, easing to 1.14 near the map's resolution. Moby Dick: R² 0.980 on log axes and a power-law tail (α 1.94), but a lognormal is not ruled out. #MT24Learning #Fractals #ComplexSystems
October 22, 2026 at 3:27 PM
Sprint 38: Britain measures 2,404 km with a 200 km ruler and 6,958 km with a 1.6 km one. D is 1.27, easing to 1.14 near the map's resolution. Moby Dick: R² 0.980 on log axes and a power-law tail (α 1.94), but a lognormal is not ruled out. #MT24Learning #Fractals #ComplexSystems
Sprint 37: Emergence. Schelling's city: residents ask for half like neighbours, the city ends 0.79 alike. A sandpile with nothing tuned: avalanches up to 5,902 topplings. Real Fiji quakes: b = 1.08 for M ≥ 4.5 and 1.22 for M ≥ 4.7. #MT24Learning #ComplexSystems #Emergence
October 21, 2026 at 10:49 PM
Sprint 37: Emergence. Schelling's city: residents ask for half like neighbours, the city ends 0.79 alike. A sandpile with nothing tuned: avalanches up to 5,902 topplings. Real Fiji quakes: b = 1.08 for M ≥ 4.5 and 1.22 for M ≥ 4.7. #MT24Learning #ComplexSystems #Emergence
Sprint 36: Dynamics and Chaos. Lorenz (1963) has no dice, yet twins 10⁻⁸ apart part at t = 20. Each 10× sharper start buys 2.54 time units (ln 10 ÷ 0.906). On real laser data the worst forecast error per pulse grew 10 → 20 → 37 → 72 → 108. #MT24Learning #ChaosTheory #ComplexSystems
October 20, 2026 at 9:31 PM
Sprint 36: Dynamics and Chaos. Lorenz (1963) has no dice, yet twins 10⁻⁸ apart part at t = 20. Each 10× sharper start buys 2.54 time units (ln 10 ÷ 0.906). On real laser data the worst forecast error per pulse grew 10 → 20 → 37 → 72 → 108. #MT24Learning #ChaosTheory #ComplexSystems
The road so far: Phases 1–11 of MT24 Learning, 35 of 90 sprints, drawn as a transit map. Each stop carries one checked result and shows which earlier phases it builds on. Ten phases have DOIs; Phase 11's is coming soon. #MT24Learning #LearningInPublic #OpenScience
October 20, 2026 at 1:22 PM
The road so far: Phases 1–11 of MT24 Learning, 35 of 90 sprints, drawn as a transit map. Each stop carries one checked result and shows which earlier phases it builds on. Ten phases have DOIs; Phase 11's is coming soon. #MT24Learning #LearningInPublic #OpenScience
Phase 11 done: Transformers & Vector Spaces, as a journal issue. Five claims checked against code and papers: the analogy needs its inputs hidden, close isn't synonym, attention predates 2017, weights are contested clues, the FFN holds 2/3 of an encoder layer.
#MT24Learning #Transformers #NLP
#MT24Learning #Transformers #NLP
October 14, 2026 at 11:26 PM
Phase 11 done: Transformers & Vector Spaces, as a journal issue. Five claims checked against code and papers: the analogy needs its inputs hidden, close isn't synonym, attention predates 2017, weights are contested clues, the FFN holds 2/3 of an encoder layer.
#MT24Learning #Transformers #NLP
#MT24Learning #Transformers #NLP
Sprint 35/90: Inside a Transformer. I built one layer in numpy. In each encoder layer the FFN holds 0.67 of the weights, not attention. With no mask and no position codes, "dog bites man" and "man bites dog" look like one set of words.
#MT24Learning #Transformers #DeepLearning
#MT24Learning #Transformers #DeepLearning
October 14, 2026 at 11:56 AM
Sprint 35/90: Inside a Transformer. I built one layer in numpy. In each encoder layer the FFN holds 0.67 of the weights, not attention. With no mask and no position codes, "dog bites man" and "man bites dog" look like one set of words.
#MT24Learning #Transformers #DeepLearning
#MT24Learning #Transformers #DeepLearning
Sprint 34/90: How Attention Works. I ran softmax(QKᵀ/√dk)V on real GloVe vectors. In "she took a loan from the bank", loan climbs from bank's 52nd neighbour to 12th. Without √dk, softmax picks one winner (0.95). Weights are clues, not reasons.
#MT24Learning #Transformers #NLP
#MT24Learning #Transformers #NLP
October 13, 2026 at 12:29 PM
Sprint 34/90: How Attention Works. I ran softmax(QKᵀ/√dk)V on real GloVe vectors. In "she took a loan from the bank", loan climbs from bank's 52nd neighbour to 12th. Without √dk, softmax picks one winner (0.95). Weights are clues, not reasons.
#MT24Learning #Transformers #NLP
#MT24Learning #Transformers #NLP
Sprint 33/90: Meaning as Vectors. On real GloVe vectors, king - man + woman lands nearest to king (0.885), not queen (0.852). Queen wins only when the search skips the inputs, as Mikolov et al. wrote in 2013. And hot's 2nd neighbour is cold.
#MT24Learning #NLP #WordEmbeddings
#MT24Learning #NLP #WordEmbeddings
October 12, 2026 at 11:59 PM
Sprint 33/90: Meaning as Vectors. On real GloVe vectors, king - man + woman lands nearest to king (0.885), not queen (0.852). Queen wins only when the search skips the inputs, as Mikolov et al. wrote in 2013. And hot's 2nd neighbour is cold.
#MT24Learning #NLP #WordEmbeddings
#MT24Learning #NLP #WordEmbeddings
Phase 10 complete · Neural Networks
31 · one neuron can't do XOR; two hidden ReLU units bend the space
32 · backprop checked to 3.7×10⁻¹⁰; η = 30 cycles; zero weights train 16 units as 1
Next: transformers and vector spaces.
#MT24Learning #NeuralNetworks #DeepLearning
31 · one neuron can't do XOR; two hidden ReLU units bend the space
32 · backprop checked to 3.7×10⁻¹⁰; η = 30 cycles; zero weights train 16 units as 1
Next: transformers and vector spaces.
#MT24Learning #NeuralNetworks #DeepLearning
October 9, 2026 at 6:51 PM
Phase 10 complete · Neural Networks
31 · one neuron can't do XOR; two hidden ReLU units bend the space
32 · backprop checked to 3.7×10⁻¹⁰; η = 30 cycles; zero weights train 16 units as 1
Next: transformers and vector spaces.
#MT24Learning #NeuralNetworks #DeepLearning
31 · one neuron can't do XOR; two hidden ReLU units bend the space
32 · backprop checked to 3.7×10⁻¹⁰; η = 30 cycles; zero weights train 16 units as 1
Next: transformers and vector spaces.
#MT24Learning #NeuralNetworks #DeepLearning
Sprint 32/90 · How networks train
Backprop matched a finite difference to 10 decimals. Then: η = 30 hit a loss of 563 and cycled; all-zero weights trained 16 units as one; 12 sigmoid layers left layer 1 with 10⁻⁷ of the error signal.
#MT24Learning #NeuralNetworks #DeepLearning
Backprop matched a finite difference to 10 decimals. Then: η = 30 hit a loss of 563 and cycled; all-zero weights trained 16 units as one; 12 sigmoid layers left layer 1 with 10⁻⁷ of the error signal.
#MT24Learning #NeuralNetworks #DeepLearning
October 8, 2026 at 3:23 PM
Sprint 32/90 · How networks train
Backprop matched a finite difference to 10 decimals. Then: η = 30 hit a loss of 563 and cycled; all-zero weights trained 16 units as one; 12 sigmoid layers left layer 1 with 10⁻⁷ of the error signal.
#MT24Learning #NeuralNetworks #DeepLearning
Backprop matched a finite difference to 10 decimals. Then: η = 30 hit a loss of 563 and cycled; all-zero weights trained 16 units as one; 12 sigmoid layers left layer 1 with 10⁻⁷ of the error signal.
#MT24Learning #NeuralNetworks #DeepLearning
Sprint 31/90 · From perceptron to multilayer network
No straight line separates XOR. My perceptron ended every epoch on the same weights from epoch 2.
Two hidden ReLU units bend the space. From 100 seeds: 2 units solved 77, 4 solved all 100.
#MT24Learning #NeuralNetworks #DeepLearning
No straight line separates XOR. My perceptron ended every epoch on the same weights from epoch 2.
Two hidden ReLU units bend the space. From 100 seeds: 2 units solved 77, 4 solved all 100.
#MT24Learning #NeuralNetworks #DeepLearning
October 7, 2026 at 9:18 PM
Sprint 31/90 · From perceptron to multilayer network
No straight line separates XOR. My perceptron ended every epoch on the same weights from epoch 2.
Two hidden ReLU units bend the space. From 100 seeds: 2 units solved 77, 4 solved all 100.
#MT24Learning #NeuralNetworks #DeepLearning
No straight line separates XOR. My perceptron ended every epoch on the same weights from epoch 2.
Two hidden ReLU units bend the space. From 100 seeds: 2 units solved 77, 4 solved all 100.
#MT24Learning #NeuralNetworks #DeepLearning
Phase 9 complete: classical machine learning. A rule from 70 flowers, a 0.900 floor that finds no 9s, k-NN falling to 0.704 without scaling, ridge tuned by 5-fold CV, and a bound that prices luck. Next: neural networks. #MT24Learning #MachineLearning
October 7, 2026 at 1:34 PM
Phase 9 complete: classical machine learning. A rule from 70 flowers, a 0.900 floor that finds no 9s, k-NN falling to 0.704 without scaling, ridge tuned by 5-fold CV, and a bound that prices luck. Next: neural networks. #MT24Learning #MachineLearning
Sprint 30/90: learning theory. Hoeffding plus a union bound gives ε = √(ln(2|H|/δ)/2n). On 70 random labels the best of a million coin-flip rules looks 78.6% right. For Sprint 26's 142 cuts: true error ≤ 0.291 at 95%; the test gave 0.100. #MT24Learning #LearningTheory #MachineLearning
October 6, 2026 at 6:36 PM
Sprint 30/90: learning theory. Hoeffding plus a union bound gives ε = √(ln(2|H|/δ)/2n). On 70 random labels the best of a million coin-flip rules looks 78.6% right. For Sprint 26's 142 cuts: true error ≤ 0.291 at 95%; the test gave 0.100. #MT24Learning #LearningTheory #MachineLearning
Sprint 29/90: ridge regression on 65 terms, 5-fold CV over 15 penalties. α = 0.001: train 2,037, CV 4,027. Best α = 31.6: CV 3,047 ± 203. The one-standard-error rule picked α = 100; on 133 sealed patients it scored 3,717 vs 3,760. #MT24Learning #MachineLearning #Statistics
October 5, 2026 at 7:12 PM
Sprint 29/90: ridge regression on 65 terms, 5-fold CV over 15 penalties. α = 0.001: train 2,037, CV 4,027. Best α = 31.6: CV 3,047 ± 203. The one-standard-error rule picked α = 100; on 133 sealed patients it scored 3,717 vs 3,760. #MT24Learning #MachineLearning #Statistics
Sprint 28/90: logistic regression, a decision tree and 5 nearest neighbours on 178 wines. In raw units k-NN scored 0.704: proline (278 to 1,680) swamped flavanoids (0.34 to 5.08). With z-scores, 0.944. The tree did not move. #MT24Learning #MachineLearning #DataScience
October 5, 2026 at 2:51 PM
Sprint 28/90: logistic regression, a decision tree and 5 nearest neighbours on 178 wines. In raw units k-NN scored 0.704: proline (278 to 1,680) swamped flavanoids (0.34 to 5.08). With z-scores, 0.944. The tree did not move. #MT24Learning #MachineLearning #DataScience
Sprint 27/90: baselines first. Only 54 of 540 test digits are 9s, so "always no" scores 0.900 and finds none. Balanced accuracy: 0.500. A one-pixel rule picked by accuracy scored 0.894; picked by balanced accuracy, it found 47 of 54. #MT24Learning #MachineLearning #DataScience
October 4, 2026 at 6:53 PM
Sprint 27/90: baselines first. Only 54 of 540 test digits are 9s, so "always no" scores 0.900 and finds none. Balanced accuracy: 0.500. A one-pixel rule picked by accuracy scored 0.894; picked by balanced accuracy, it found 47 of 54. #MT24Learning #MachineLearning #DataScience
Sprint 26/90 opens Phase 9: supervised learning. I scored three rules on 30 held-out iris flowers. Memorising answered only 6 of them. Nearest neighbour and a single cut on petal width each missed 3, even though training error had ranked them 1 vs 3. #MT24Learning #MachineLearning #DataScience
October 4, 2026 at 1:25 PM
Sprint 26/90 opens Phase 9: supervised learning. I scored three rules on 30 held-out iris flowers. Memorising answered only 6 of them. Nearest neighbour and a single cut on petal width each missed 3, even though training error had ranked them 1 vs 3. #MT24Learning #MachineLearning #DataScience