40 · swipe or use the arrows
01 / 40
In 59 seconds, with narration: an oracle that does not speak, it only raises a card with a number. What broke when installing it, the 19.9 ms per query, the AUC of 0.823 predicting rain without training, and the 0.62 ceiling on its probabilities. Every figure is a measured one. Muted by default: turn the sound on in the controls.
02 / 40
In 57 seconds: the 28% AWS claims, the 1.82x more expensive I measured locally, the flip to 88.8% cheaper in the cloud with temperature matched, and the 918,648 tokens Claude Code burned as a third harness.
03 / 40
In 78 seconds, narrated: DEFLATE's parse as a shortest path in a graph (44 bits from zlib's real output vs 36 bits from the brute-force-verified optimum in the example), the Huffman tree that iterates (-8.5% on a real file), and the real result across 29 files: -2.86% against zlib, verified byte for byte. Muted by default: turn sound on in the controls.
04 / 40
In 72 seconds, narrated: Rust Coreutils (uutils) 0.12 against GNU coreutils 9.12. The 653 in its notes is the best result per test out of six CI runs, and it did not reproduce in my container; in cp, mv, rm, install, ls, sort and du only 3 tests fail. On speed, uutils is slower in 91 of 157 cells (starting a process costs almost three times more), but wins in 45. And mv across disks loses the modification time. Muted by default: turn sound on in the controls.
05 / 40
In 75 seconds, narrated: Dragonfly 2.0, Redis 8 and Valkey 9 with the same CPU budget. With 2 CPUs and no pipelining Dragonfly beats stock Redis (342,601 against 172,081 operations per second with io_uring) and uses 8 to 15 percent less memory per key, but with 4 CPUs and no pipelining the bench cannot rank the three, and with pipeline 16 Redis and Valkey with io-threads win (4.21 and 4.04 million against 2.73). One workload, closed-loop bench. Muted by default: turn sound on in the controls.
06 / 40
In 67 seconds and narrated: the same Mandelbrot kernel in cuda-oxide, CUDA written in Rust, and in CUDA C++ on an RTX 4070 Ti SUPER. With default options Rust comes out 3.2% faster, but the two images differ in 27,510 pixels because each toolchain fuses different operations into FMA. With FMA off the outputs match bit for bit and Rust is 7% slower; the cause of the 3% is not established. Muted by default: turn the sound on in the controls.
07 / 40
In 69 seconds and narrated: Wild against mold linking ripgrep and cargo, median of 20 repetitions. Wild won all 20 of each project, by 1.9 ms on ripgrep and about 8 ms on cargo. With rust-lld as the default linker, linking is 2 to 5% of a release rebuild of 2.2 to 2.6 s, and close to 1% with mold or Wild. It includes a mistake in my benchmark: a fake linker passed with -B never ran. Muted by default: turn the sound on in the controls.
08 / 40
In 46 seconds, narrated: one hidden layer of 128 units matched a boosting model on the test period; why with no hidden layer the network is exactly a logistic regression; what happens when it grows too much; how much the result moves when only the seed changes; and what it gains in size and latency per row. Muted by default: turn the sound on in the controls.
09 / 40
In 46 seconds, narrated: why the Transformer landed below three simpler models; the 9,633 weights the rule picked and the 0.007 of AUC that separate it from boosting; how the ranking of eleven architectures flipped between validation and the test period; what happened when the attention was removed; and the layer that shields it from a unit change by leaving one column mute. Muted by default: turn the sound on in the controls.
10 / 40
In 47 seconds, narrated: fitting KNN took seven milliseconds and scoring the test set five and a half seconds; how it decides with its neighbors and why a single one memorizes; how unscaled humidity takes over the distance; why the search trees were slower than exhaustive search; and where it landed against logistic regression and boosting. Muted by default: turn the sound on in the controls.
11 / 40
In 47 seconds, narrated: the whole model weighs 1.3 kilobytes and was fitted in eight milliseconds; how it looks at each variable separately inside each class; what happens when a column is copied seven times; why six out of ten days get a probability pinned to zero or to one; and how it behaves with only thirty training rows. Muted by default: turn the sound on in the controls.
12 / 40
In 49 seconds and narrated: the same unscaled SVM collapses when one variable changes units; what the margin and support vectors are; how gamma bends the boundary until it memorizes; how fast training cost grows when using every row; and how to approximate the kernel with Nystroem plus a linear model. Muted by default: turn the sound on in the controls.
13 / 40
In 42 seconds and narrated: a decision tree with no limit reached AUC 1.000 on the days it learned from and dropped to 0.689 on new days; it asks one variable at a time; unchecked it grew more than 8,500 leaves, and at depth 7, chosen on validation, it did better; the first question barely changes, but the predictions disagree on a third of the days; and for numbers it predicts in steps, flat beyond what it saw. Muted by default: turn the sound on in the controls.
14 / 40
In 64 seconds and narrated: the same Qwen3.8-27B on the same 16 GB GPU, with ExLlamaV3 and llama.cpp getting the same tokens. With MTP on code, 103.1 tokens per second against 79.5; in an 11,000-token agentic turn the lead shrinks to 8.62 against 9.29 seconds; and on quality ExLlamaV3 wins with 13 GB files and loses with 15 GB ones. Muted by default: turn the sound on in the controls.
15 / 40
In 51 seconds and narrated: with learning rate 1.0, gradient boosting destabilized and after 3,000 rounds ended at AUC 0.681, below a decision tree with no depth limit; it adds small trees in a chain and each one fits what the previous ones missed; with rate 0.03, validation chose 594 rounds and it reached 0.883; against a 200-tree forest the difference was not conclusive, but on one thread it trained 9 times faster and weighs 2.1 MB against 124 MB; and with one-question trees it needed almost 20,000 rounds. Muted by default: turn the sound on in the controls.
16 / 40
In 49 seconds and narrated: without being told the cities, K-Means put 92% of Santiago's days in one group; Lloyd's algorithm assigns and moves centroids; unscaled, humidity accounts for 76% of the separation; inertia shows no elbow and the silhouette stays within sampling noise across k = 3, 5 and 6; and with k = 8, only 18% of random starts end close to the best inertia, against 38% for k-means++. Muted by default: turn the sound on in the controls.
17 / 40
In 64 seconds and narrated: the viral Random Forest diagram draws the trees in a chain, and that is boosting. On eight cores, 200 trees drop from 26.1 to 3.5 seconds; AUC climbs from 0.69 with one tree to 0.88 with a hundred; a column of pure noise looks important with the default measure; and trained on cold months from 1984 to 2012, the forest never predicted above 28.9 °C in the summers of 2016 to 2026. Muted by default: turn the sound on in the controls.
18 / 40
In 57 seconds and narrated: on Santiago summers no model had seen, the line missed by 1.86 °C and the random forest by 3.76 °C; the yearly max rises 0.33 °C per decade in Temuco and falls in Valparaíso; unscaled, gradient descent ends with an error of 9.8×10¹¹ °C; with 3,000 rows, the dew point comes out with the minority sign in 46% of samples; and with 5% of rows carrying a misplaced decimal point, least squares rises to 7.89 °C while Huber stays at 1.69 °C. Muted by default: turn the sound on in the controls.
19 / 40
In 43 seconds and narrated: a straight line fitted to days with and without rain gives negative probabilities on 9% of days; logistic regression passes the sum of the variables through an S-shaped curve; at a 0.5 threshold it warns about only half of the rain; if missing rain costs three times a false alarm, it catches 80%; and it promises 24% rain when it rains 20%. Muted by default: turn the sound on in the controls.
20 / 40
In 52 seconds and narrated: the gatekeeper sorts the messages, an injection is cut off thirteen times sooner than a full answer, the verifier strikes out the invented figure, and the cost, which depends on how you build it: about 992 dollars per million messages paying as you go, or the 22-dollar monthly plan this blog uses. Every figure comes from what was measured. Muted by default: turn the sound on in the controls.
21 / 40
In 37 seconds: the 6x that was my bug, the getaddrinfo line that caused it, and what Pingora really measures in a 2 vCPU pod. Every number is a measured one. Muted by default: turn the sound on in the controls.
22 / 40
In 43 seconds: the rule stamping what was already written down, the forest watching the whole stream, and the measured price of putting it inside the pipeline. Every number is a measured one. Muted by default: turn the sound on in the controls.
23 / 40
In 54 seconds: almost half of the most visited .cl domains already use post-quantum cryptography, who actually put it there, and why six times the bytes cost zero milliseconds.
24 / 40
In 55 seconds: the half of the handshake that is already post-quantum with nobody configuring it, why identity was left for later, and the wall that turned out to be a compiler version.
25 / 40
In 47 seconds: pg_anon caught 1 of 8 columns holding personal data on a Chilean database, and all 8 once I wrote the rules. The RUT is blind by default.
26 / 40
In 45 seconds: why the model that fits is the dumber one, how the dial spreads the bits layer by layer, and the 94% fidelity that fits in 16 GB where Q4 doesn't.
27 / 40
The bench actually running: four runtimes in containers with one core and 512 MB each, load alternating between them, and live cgroup readings. Narrated.
28 / 40
In 51 seconds: whose data it is, what the person actually signs, the three institutions talking underneath, and the piece Keycloak does not ship. Narrated in Spanish.
29 / 40
In 59 seconds: why the first answer was not a model, how you verify that an automatic translation did not change the program, and what the model was actually good for. With the game running behind it, captured from the engine itself.
30 / 40
In 54 seconds: what changes with Go's new vectors, why they tie with NumPy out of cache and lose inside it, and the false comparison I nearly published. Muted by default: turn it on in the controls.
31 / 40
In 33 seconds with narration: two bots compete over a table. The one-eyed one looks and answers; the geared one tries twenty-five combinations before replying. Every number is a measured one. Muted by default: turn it on in the controls.
32 / 40
The summary in 24 seconds: what it weighs, what it takes and how close it gets to the official forecast. Muted by default: turn it on in the controls.
33 / 40
In 26 seconds: three bots climbing a hill, each moving at the pace actually measured. The 284-billion fat one barely leaves the start, the muscular one pushes halfway up, and the skinny two-bit one crests first. Muted by default: turn it on in the controls.
34 / 40
In 27 seconds: why the diagram lies, what the model looks like written as text, how a single file produces all three C4 levels, and the four gates that unlocks. Muted by default: turn it on in the controls.
35 / 40
In 24 seconds: what broke, the command that tells you whether it applies to you, the result of auditing all four machines, and why Linux's official tool said everything was fine. Muted by default: turn it on in the controls.
36 / 40
PhantomRelay in 28 seconds: the TLS handshake that gives you away, why OpenSSL and BoringSSL do not produce the same bytes, and why the fix was to inherit rather than imitate. Muted by default: turn it on in the controls.
37 / 40
How it works in 28 seconds: from text to audio, how it extracts the timbre from a sample, and the real metrics on a CPU with no GPU. Muted by default: turn it on in the controls.
38 / 40
In 21 seconds: why the file weighs 16.5 GB and not 15, how the model splits between CPU and GPU depending on context, and the benchmark bug the model uncovered. Muted by default: turn it on in the controls.
39 / 40
Agatha in 29 seconds: from the dog to the bitmap, and from YouTube back to the file. Muted by default: turn the audio on in the controls.
40 / 40
Encapsulation live: the sender walks the data down adding headers and the receiver unwraps it. Muted by default — turn the audio on in the controls.
numbers, not hype
Hi. Tools get tested here and the numbers get published. What are you after? Ask me about an article, a project or a measurement.
Enter sends · Shift+Enter adds a line · 2000 characters max