It has four lessons. All numbers are in arbitrary units, and the page states that it is a toy model, not a forecast.

1. The Loop: each lap of the animation is one generation (research → build → test → deploy). You can set how much better each generation is at research and what share of the R&D work the AI does. While humans still do part of the work, each lap has a minimum time (Amdahl’s law); once the work is almost fully automated, each lap is faster than the last.
2. Returns: shows how the exponent α changes growth in the model dI/dt = c·Iᵅ. Below 1, growth slows over time; at 1, it is exponential; above 1, it reaches infinity in finite time (a “hard takeoff”). The analytic solution and the blow-up time are shown.
3. Bottlenecks: switch on, one at a time, compute (combined with ideas as a Cobb–Douglas production function), serial work that can’t be automated, and physical limits. With the same α, the “software only” curve blows up, while the bottlenecked curve bends down to follow the growth of compute.
4. Oversight: each generation’s goals drift slightly, and errors compound. When an evaluation catches a drift, the system rolls back and fixes it before continuing. If oversight does not scale with the model, the detection rate falls with each generation and more drifts slip through. The lesson also introduces safety research ideas: evaluations, interpretability, scalable oversight and staged deployment.

AI Image Forensics is a browser-based tool that looks for evidence of whether an image was generated by AI or captured by a camera. It runs entirely in your browser and never uploads the image anywhere.

What it analyzes

• Metadata. It reads EXIF (camera make and model, lens, exposure, GPS, software), XMP, IPTC and JPEG comments. For PNG files it also reads text chunks, where tools such as Stable Diffusion WebUI and ComfyUI often store the prompt, sampler, seed and workflow.
• C2PA (Content Credentials). It detects whether a C2PA manifest is embedded. It then reads the claim generator, the recorded actions and the digital source type. A source type of trainedAlgorithmicMedia means the image was declared as AI-generated.
• File structure. It lists every JPEG segment, PNG chunk, WebP chunk or HEIF/AVIF box. For JPEG files it also estimates the compression quality.
• FFT frequency spectrum. It shows the 2D spectrum of the image’s noise residual (or its luminance), along with an azimuthally averaged power curve. Regular bright peaks can be traces of upsampling in generative models. Peaks on the JPEG 8×8 grid are marked separately, because they come from compression, not generation.
• Noise residual. It subtracts a 3×3 median-filtered copy from the original to reveal the noise layer. Camera sensor noise is usually spread evenly across the frame. Generated images may show little noise, inconsistent texture between regions, or repeating patterns.
• ELA (error level analysis). It re-saves the image as JPEG and amplifies the difference from the original. This is mainly useful for spotting locally edited or pasted-in regions in JPEG files.

How to read the results

The tool does not give an “AI or not” verdict. Instead, it lists each piece of evidence it finds and rates its strength:

• Strong: AI generation declared in C2PA or XMP, or generation parameters found in the file.
• Moderate: a generator or editing tool named in the metadata.
• Weak: camera EXIF alone, a resolution commonly used by generators, or periodic peaks in the spectrum.
• Inconclusive: no metadata at all. This is common for both generated images and photos that have been through social media or messengers.

The guiding principle is asymmetric. Evidence that is present is meaningful, while evidence that is absent tells you very little, because metadata is easily stripped or forged.

Limitations

• C2PA signatures are not verified. Use contentcredentials.org/verify for that.
• Invisible watermarks such as Google’s SynthID cannot be detected without the vendor’s own detector.
• Images that were shared through social media or messengers, re-compressed, or captured as screenshots lose most of their metadata and much of their signal-level evidence.

Laser Mosquito Fence is an interactive simulator of a laser mosquito-control device, inspired by the Photonic Fence research project. It models the decision logic (which insect to zap, and when), not the laser hardware.

How it works (five-stage pipeline)

1. Detect: insects flying in front of an LED backlight on the far side show up to the camera as dark shadows.
2. Classify: each insect’s species is judged from wingbeat frequency and size. Female mosquitoes beat at about 450–600 Hz, males at 650–900 Hz and honeybees at about 230 Hz. Confidence rises as observations accumulate.
3. Track: a steering mirror turns toward where the insect will be after the system latency. The mirror has a limited turning speed.
4. Safety: the whole beam path is checked. If a person is nearby, or a protected insect such as a bee is on the beam line, the shot is held.
5. Zap: only biting female mosquitoes get a short pulse. If a shot misses, the system tries again.

Four scenarios

• Evening garden: a mix of insects, the baseline case.
• Mosquito swarm: a throughput test with many mosquitoes queuing for the single mirror.
• Pollinator day: lots of bees, testing whether classification avoids hitting the wrong target.
• Someone walks by: the safety interlock when a person crosses the curtain.

Controls and display

• Sliders: confidence threshold, latency, mirror speed, sensor noise and safety margin.
• Buttons: slow motion, add a person or insects, and disarm.
• Wingbeat spectrum: click any insect to see its spectrum against the frequency bands of each species.
• Results panel: precision (share of zaps that hit female mosquitoes), mosquitoes that escaped, misses, safety holds and zaps per minute.
• Decision log: every zap, miss and hold, with the reason.

Verified: each scenario ran for 15 seconds in a headless browser with no script errors. At default settings, precision was 100% with no male mosquitoes or bees hit, and the person scenario held fire seven times. Raising sensor noise to 100% and lowering the confidence threshold to 50% caused a male mosquito to be hit by mistake and many more misses, which shows how the threshold and sensor noise affect the results.

All numbers are illustrative approximations.

아두이노(Arduino)는 누구나 쉽게 대화형 전자 제품과 사물인터넷(IoT)을 만들 수 있도록 제작된 오픈소스 하드웨어 및 소프트웨어 생태계입니다. 자료는 아두이노(arduino) 입문자용 가상 시물레이터 입니다.

구성

• 보드: UNO R3(ATmega328P, 16 MHz)와 UNO R4 Minima(RA4M1, 48 MHz) 중 고를 수 있습니다. 헤더 핀, PWM ~ 표시, L·TX·ON LED, 리셋 버튼을 그렸고, 사용 중인 핀은 금색으로 표시됩니다. 공식 로고는 넣지 않았습니다.
• 브레드보드 부품: LED(220 Ω), RGB LED, 푸시 버튼(풀다운 / GND·풀업 배선), 가변저항, 조도 센서, TMP36 온도 센서, SG90 서보, 피에조 부저, 16×2 LCD를 놓을 수 있습니다. 부품을 더하거나 빼고 핀을 바꿀 수 있고, 버튼은 누르고 있기, 가변저항은 드래그, 빛·온도는 슬라이더로 조작합니다.
• 전선 색: 핀 상태에 따라 바뀝니다. 빨강은 HIGH, 주황은 PWM, 보라는 tone, 파랑은 입력입니다.
• 코드 편집기: 구문 강조가 되고, 지금 실행 중인 줄을 초록색으로 표시합니다. Ctrl+Enter로 업로드하고, 일시정지·리셋, 시간 배율 0.25–4×를 지원합니다.
• 시리얼: 모니터에서 보드로 글자를 보낼 수 있습니다. 플로터는 이름:값 형식의 출력을 여러 선으로 그립니다.
• 핀 패널: D0–A5 상태 표와 사용 중인 핀의 파형을 보여줍니다.

예제 12개: Blink, 버튼, 페이드(PWM), 가변저항 읽기, 신호등(millis 상태 기계 + 보행자 버튼), 서보 스윕, 멜로디(「작은 별」, 실제 소리), 야간등, LCD 온도계, RGB 무지개, 시리얼 명령, 떠 있는 입력(INPUT → INPUT_PULLUP으로 고쳐 보기).

실수를 알려 주는 경고

• PWM이 안 되는 핀에 analogWrite를 쓴 경우
• pinMode 없이 HIGH를 써서 LED가 희미하게 켜지는 경우
• Serial.begin을 빠뜨린 경우
• D0·D1이 시리얼과 충돌하는 경우
• tone을 두 핀에서 동시에 쓴 경우
• LCD에 한글을 출력한 경우

한계: 스케치를 아두이노 C의 자주 쓰는 부분만 JavaScript로 바꿔 실행합니다. 그래서 정수 나눗셈, 포인터, 구조체, static 지역 변수는 실제와 다르게 동작하거나 지원되지 않습니다. 세미콜론을 빠뜨려도 JavaScript 규칙 때문에 오류 없이 넘어가는 경우가 있습니다.

검증: 헤드리스 브라우저에서 12개 예제를 모두 변환·실행해 오류가 없는 것을 확인했습니다.

• 가변저항 512 → 2.50 V, TMP36 24 °C → LCD 23.8 °C(ADC 양자화 때문)가 정확히 나옵니다.
• 버튼 누름/뗌에 따라 LED가 켜지고 꺼지고, 시리얼 명령 ‘1’을 보내면 LED가 켜집니다.
• 떠 있는 입력은 0과 1을 오가다가 INPUT_PULLUP으로 고치면 1로 안정됩니다.
• while(true); 같은 무한 루프에서도 페이지가 멈추지 않습니다.

Sources:

• Arduino

Microduck Simulator is an interactive 3D simulator of Pollen Robotics’ Microduck, a bipedal duck robot about 25 cm tall and 800 g. It is built on specifications taken from the official microduck runtime and microduck_rl training repositories.

Real robot details used

• Servos: 15 Dynamixel XL330s. IDs are 20–24 for the left leg, 30–34 for the neck, head and beak, and 10–14 for the right leg. The beak (ID 34) is excluded from every policy’s actions.
• Pose and limits: the actual HOME pose and joint limits. Sliders use the real robot convention (radians, signs as in HOME).
• Policy loop: runs at 50 Hz with a 61-D observation (angular velocity, gravity, joint positions, joint velocities, last action, commands) and a 14-D action, both shown live as a heatmap.
• Training settings: velocity command ranges of vx ±0.4 m/s, vy ±0.3 m/s and ωz ±1.0 rad/s; an air-time reward window of 0.125–0.3 s; 2 cm foot clearance; battery voltage from 6.5 to 8.2 V; ±1° backlash; and a ±0.3 m/s push range.
• Policy names: the top bar shows the real task ID for each mode, such as Mjlab-Velocity-Flat-MicroDuck.

Behaviours

• Walk: drive with the on-screen joystick or W A S D / Q E, and adjust body pose and head commands. Lengthening the gait period pushes air time out of the reward window.
• Roller skates: passive wheels under the feet and a swizzle glide with no foot lift.
• Sit · stand (Y): switched by a posture-flag command.
• Beak pick (G): the phase-encoded command can be seen circling in the observation panel. The duck can pick up the block and carry it while walking.
• Ball kick (K / L): shifts weight onto the support foot, then kicks a 70 mm ball.
• Forward roll (R): rolls over the head and lands back on its feet.
• Falls and stand-up: a push that moves the capture point past 62 mm knocks the robot over, and it gets back up by itself.
• Joint control: sliders for all 15 servos, plus the SIT and FOLD keyframes as presets.
• Quack (Space): opens the beak and plays a sound.

Panels

• Live readout: state, commanded and actual speed, trunk tilt, support, capture point and battery.
• Policy I/O: the 61-D observation and 14-D action.
• Foot contact: a contact timeline and the last air time for each foot.
• Servo table: target, actual and error for all 15 servos.
• Controls and overlays: camera presets, a support-polygon and centre-of-mass overlay, backlash and battery settings, and time scale.

Limitations: the motion comes from a kinematic generator (gait phases plus inverse kinematics), not from the trained ONNX policy or MuJoCo physics. Leg geometry is approximate, so poses can differ somewhat from the real robot, especially when the real SIT and FOLD keyframe values are applied directly.

Verified: headless Chromium ran every mode without script errors. Walking air time was 180 ms, inside the reward window. Push → fall → stand-up, block pickup and ball kick all worked, and there is no horizontal scroll at phone width.

Sources:

• pollen-robotics/microduck
• pollen-robotics/microduck_rl

Reed–Solomon Lab is an interactive browser simulator of Reed–Solomon error correction. It runs a real encoder and decoder over GF(2⁸), and every step of the repair updates live as you damage bytes.

Pipeline (top to bottom)

• Status and budget bar: shows how much of the correction budget 2e + s ≤ n − k is used. Errors cost 2 slots and erasures cost 1. The result is labeled Clean, Corrected, Decoder failure or Miscorrected.
• Step 1 – Encode: message bytes plus parity bytes form the codeword. The generator polynomial g(x) is shown.
• Step 2 – Channel: click received bytes to add errors (random values) or erasures (flagged positions). You can also add random hits or bursts.
• Step 3 – Syndromes: non-zero syndromes show that damage was detected.
• Step 4 – Berlekamp–Massey: computes the error locator Λ(x), plus the erasure locator Γ(x) when erasures are present.
• Step 5 – Chien search: a bar chart of |Λ| per byte position. Roots, which mark the error locations, are highlighted green.
• Step 6 – Forney: a table of error values and the XOR fix applied at each position.
• Step 7 – Decoded: repaired bytes are highlighted, and the sent, received and decoded text are compared side by side.

Scenarios: Sandbox, QR Code 1-M RS(26,16), Audio CD C2 RS(28,24), DVB-T RS(204,188) and Deep Space RS(255,223). Each comes with notes on how that code is used in practice. You can also edit the message and the number of parity symbols.

Verified: 4,000 random fuzz tests within the correction limit all decoded exactly. Damage beyond the limit never crashes the decoder; it is reported as a failure or a miscorrection.

Three-Body Lab is an interactive browser simulator of the classical three-body problem. It uses the same dark sidebar-and-canvas layout as the Physics Playground.

Configurations

• Figure-Eight: a stable periodic orbit where three equal masses share one curve (Moore 1993, Chenciner & Montgomery 2000).
• Lagrange Triangle: an equilateral arrangement that rotates rigidly. It is unstable with equal masses and stable when one mass dominates.
• Butterfly I: an unstable periodic orbit found by Šuvakov & Dmitrašinović in 2013.
• Pythagorean: Burrau’s problem. Masses 3, 4 and 5 start at rest, go through chaotic close encounters, and one body is ejected.
• Hierarchical: a star, planet and moon, showing how a close pair with a distant third stays stable.
• Random Start: random initial conditions that usually turn chaotic.

Features

• Ghost twin: a copy of the system nudged by a tiny amount (1e-6 by default). A log-scale graph tracks how fast the two copies diverge, which shows chaos directly.
• Edit mode: pause the simulation, then drag bodies to move them and drag the arrow tips to set their velocities.
• Controls: speed, trail length, gravitational softening, twin nudge size, and a mass multiplier for each body.
• Live readouts: time, energy and its conservation error, angular momentum, closest separation, steps per frame, and twin separation.
• Navigation: the view auto-fits by default. You can also zoom with the scroll wheel and pan by dragging empty space. Bodies that leave the screen get an edge arrow with their distance.
• Shortcuts: Space to play or pause, R to reset.

Physics: the simulator uses Newtonian gravity with G = 1 and a leapfrog integrator. The time step shrinks automatically when bodies come close. A short note in the sidebar explains the history: Poincaré showed that no general closed-form solution exists, so the problem can only be solved numerically.

I built and published Build a Transformer, a page that shows how a developer codes the Transformer from “Attention Is All You Need” in PyTorch, one line at a time.

Layout

• Left, code editor: each chapter has PyTorch code you could actually run.
• ⌨ Type it out types the code line by line and explains the important lines as they appear.
• Click any line to see what it does and how it changes the tensor shapes.
• Right, visuals: a diagram of how tensor shapes change, plus heat maps of real numbers. The heat maps come from a tiny Transformer (d_model = 8) that runs in your browser.
• Bottom, console: ▶ Run cell (Ctrl+Enter) prints results such as .shape and parameter counts. They are calculated from the hyperparameters you set in the sidebar.

The 10 chapters

1. Embeddings (×√d_model)
2. Sinusoidal positional encoding
3. Scaled dot-product attention
4. Multi-head attention (splitting heads with view/transpose)
5. Feed-forward network
6. Encoder layer (Add & Norm)
7. Decoder layer (causal mask and cross-attention)
8. Full model (the output layer reuses the embedding weights)
9. Training loop: Adam, a warm-up learning-rate schedule, label smoothing and teacher forcing
10. Inference: greedy decoding you can step through one token at a time

Hyperparameters

• Base / Big / Tiny presets are available. The parameter counts come out at about 63.1M for Base and 214M for Big, which closely match the paper’s figures of about 65M and 213M.
• If d_model can’t be divided evenly by h (for example 512 and 12), a warning appears and the console shows an AssertionError.

Things to keep in mind

• The tiny model behind the heat maps uses random weights that were never trained, so its attention patterns don’t mean anything.
• The German translation in the inference chapter is an example of what a trained model would produce, not real output.

Transformer Architecture is an interactive page that walks through the transformer, the architecture behind ChatGPT, in five stages

Tokens & Embeddings

• Splits your text into tokens (long words become pieces) and gives each one a token ID.
• For a selected token, shows its embedding, the positional encoding and their sum (the input vector) as heat maps, alongside a table of the sin/cos positional encodings.
• A 2D meaning map shows similar words clustering together, and that “king − man + woman ≈ queen”.

Self-Attention

• Computes real dot products from Q, K and V, divides by √d and applies softmax. You get arcs between tokens, the attention matrix and a step-by-step table for the selected row.
• Three heads to switch between: pronoun → noun, previous token, and uniform averaging.
• In the classic example “The animal didn’t cross the street because it was too tired / wide”, “it” attends to animal (75%) with tired and to street (75%) with wide. Turn on the GPT-style causal mask and “it” can no longer see the adjective that comes after it, so it is left split between the two nouns.

Transformer Block: an animated step-through of LayerNorm → multi-head attention → residual add → LayerNorm → feed-forward → residual add → × N layers → logits. It includes a size comparison of the original Transformer, GPT-2 and GPT-3.

Next-Token Prediction: generates text one token at a time from logits, using temperature-scaled softmax and top-k sampling. Probability bars and entropy are shown, and you can click a candidate to choose the next token yourself.

How It Learns: on “the cat sat on the ___ → mat”, shows the cross-entropy loss and its gradient (p − 1), with a curve of the loss falling at each gradient-descent step.

Some parts are simplified for teaching:

• Vectors have 2–16 dimensions instead of thousands.
• The attention-head weights are hand-designed so each head’s job is easy to see.
• Next-token prediction uses a small n-gram model built from about 60 sentences instead of a neural network.

소수를 찾는 네 가지 방법을 계산식과 함께 한 단계씩 보여줍니다. 모든 방법에서 재생·이전·다음·결과까지 한 번에 버튼과 속도 조절(←/→, Space)을 쓸 수 있습니다.

• 시행 나눗셈n ÷ d = 몫 … 나머지 식을 차례로 보여주고, 나머지가 0이 되는 순간 합성수로 판정합니다.
• 왜 √n까지만 나누면 되는지 수직선으로 보여줍니다.
• 모든 수를 시험하는 방식과 6k±1 최적화를 비교할 수 있습니다. 최대 12자리까지 입력할 수 있습니다.
• 에라토스테네스의 체1부터 N(최대 600)까지 격자에서 소수 p를 찾고, p²부터 p씩 배수를 지워 나갑니다.
• 지금 지우는 수의 식(예: 80 = 2 × 40)을 보여주고, p² > N이 되면 남은 수가 모두 소수라는 것을 확인합니다.
• 소수 개수 π(N)과 근삿값 N/ln N을 비교합니다.
• 소인수분해짧은 나눗셈과 소인수 나무를 함께 그립니다.
• 결과를 360 = 2³ × 3² × 5처럼 쓰고, 약수 개수 (3+1)×(2+1)×(1+1) = 24도 계산합니다. 최대 13자리까지 입력할 수 있습니다.
• 밀러–라빈 판정n−1 = 2ˢ·d로 분해한 뒤, 제곱-곱셈 표로 a^d mod n을 계산하고 제곱하며 −1이 나오는지 확인합니다.
• 페르마 시험과 비교하고, 밑 12개로 반복 판정합니다.
• 카마이클 수 561, 밑 2를 속이는 2047, 19자리 소수 2⁶¹−1 같은 예제가 들어 있습니다.

헤드리스 브라우저로 알려진 결과와 맞춰 확인했습니다.

• 91 = 7 × 13, 1,000,001 = 101 × 9901, 2³¹−1은 소수로 나옵니다.
• π(120) = 30, π(600) = 109입니다.
• 600,851,475,143 = 71 × 839 × 1471 × 6857입니다.
• 2047은 밑 2와 11만 통과하는 거짓 증인 사례로 나오고, 2⁶¹−1은 12개 밑을 모두 통과해 소수로 확정됩니다.

입력 자릿수에는 상한을 두었습니다. 시행 나눗셈과 소인수분해가 √n까지 빠짐없이 계산을 끝낼 수 있는 범위에서만 판정하도록 하기 위해서입니다. 그보다 큰 수는 밀러–라빈 판정으로 확인하면 됩니다.

Wireless Charging is an interactive simulator that explains wireless phone charging (the Qi standard) in four lessons.

1. Induction: alternating current in the pad’s coil creates a magnetic field, which induces a voltage in the phone’s coil. Dragging the phone changes the coupling factor k and the induced voltage, and the transmitter current and induced voltage are plotted side by side. Switch to DC and a voltage appears only while the current is rising, then drops to zero.
2. Alignment: a top view of the pad shows a heat map of efficiency at each phone position, with a ring marking the 50% boundary. You can change case thickness and coil quality Q and see how much power is wasted as heat at each position. You can also turn on magnets that pull the phone to the centre, as MagSafe and Qi2 do.
3. Resonance: a series–series LC circuit is computed with complex impedance. A frequency sweep from 110 to 205 kHz plots power and efficiency, and removing the capacitors drops the delivered power to almost zero. With strong coupling and a light load, the resonance peak splits in two (the Split! button).
4. Charging Session:

• The handshake between pad and phone (ping → identification → power transfer), with the control packets the phone sends back.
• Bars showing how many watts enter each stage (wall → inverter → coil → rectifier → battery) and how many are lost along the way.
• The CC/CV charging curve and temperature for a 4500 mAh battery. Charging slows when the phone gets warm.
• Dropping a coin on the pad triggers foreign-object detection (FOD), which stops power transfer.

Coil dimensions and loss values are simplified approximations of real products, for illustration.

Snow Formation Lab is an interactive simulator that follows one snowflake from rising air to the ground in four stages.

Stage 1 · Rising Air

• Set surface temperature, dew point, environmental lapse rate and lift height, and watch an air parcel rise and turn into cloud.
• Below the cloud base the parcel cools at the dry adiabatic rate (9.8 °C/km); inside the cloud it cools at the moist adiabatic rate, computed from the full formula.
• The scene marks the cloud base, 0 °C, −10 °C (ice nuclei become active), the dendritic growth zone (−12 to −18 °C) and −38 °C (droplets freeze on their own).
• Based on cloud-top temperature, the cloud is classed as warm, supercooled or mixed-phase (snow-producing).
• Presets: winter front, summer storm, dry day.

Stage 2 · Ice vs Droplets (Bergeron process)

• A microscope view shows vapor moving from supercooled droplets onto ice crystals. The droplets shrink, and the crystals grow until they are heavy enough to fall.
• A chart plots saturation vapor pressure over water and over ice, and marks where the gap between them peaks, near −11.8 °C.
• You can seed the cloud by clicking or with silver iodide (AgI). Below −38 °C every droplet freezes spontaneously.

Stage 3 · Crystal Growth

• Click a temperature and supersaturation on the Nakaya diagram and that crystal grows live.
• Plate crystals grow on Reiter’s (2005) hexagonal diffusion model. As conditions change they progress from hexagonal plates to sectored plates, broad-branched stars, stellar dendrites and fern-like dendrites, and you can see vapor being used up around the crystal.
• Needles and columns, including hollow columns, are drawn as 3D prisms.
• Presets: dendrite, plate, needle, sector, column, fern.

Stage 4 · The Fall

• Drag points on the temperature profile from the 3 km cloud base down to the ground. Humidity sets the wet-bulb temperature, which decides whether flakes melt or refreeze.
• The ground result is classified as snow, wet snow, sleet (ice pellets), freezing rain or rain, and shown as accumulated snow, icicles or puddles.
• Each of the five presets (snow, rain, sleet, freezing rain, and snow at +2 °C in dry air) gives its expected result 100% of the time.

Verified: all four stages ran in a headless browser with no script errors and no horizontal scroll at phone width.

Limitations

• Crystal shapes come from mapping Reiter-model parameters onto regions of the Nakaya diagram, not from an empirical growth formula.
• Melting and refreezing in the fall stage are simplified for teaching.
• Rounding in the step-by-step calculation can make crystals slightly asymmetric.

텍스트를 입력했을때 인공지능이 작성한 글인지 판별하는 프로그램입니다. OpenAI나 Claude API를 사용하지 않고 공개 모델을 직접 돌려 Perplexity와 Binoculars 점수를 계산합니다.

사용 방법

• 글을 넣고 점수 계산을 누르면 Binoculars 점수, log-PPL, X-PPL, PPL이 나옵니다. 토큰별로 모델이 얼마나 예상했는지는 색으로 표시됩니다.
• 측정 방식은 두 가지입니다.
• Binoculars: 모델 2개로 비교합니다. 계산 방식은 공식 구현과 같고, 점수가 낮을수록 AI 쪽입니다.
• Perplexity만: 모델 1개로 측정하며, AI/사람 판정은 내리지 않습니다.
• 프리셋은 두 가지이고, 모델 ID를 직접 입력할 수도 있습니다.
• 한국어용: Qwen2.5 0.5B 계열 두 개로, 모델당 0.5~0.8GB입니다.
• 영어용 경량: SmolLM2 135M/360M입니다.
• 처음 한 번만 내려받고, 이후에는 브라우저 캐시를 씁니다.

임계값은 먼저 보정해서 쓰세요. 기본값 0.90은 논문의 Falcon-7B 조합 기준이라 소형 모델에는 맞지 않습니다. 출처를 아는 글을 여러 개 측정하고 기록 표에서 정답(사람/AI)을 표시하면, 지금 모델 조합에 맞는 임계값을 제안합니다. 예시 글로 측정하면 정답이 자동으로 표시됩니다.

알아 둘 점

• 0.5B 소형 모델은 한국어 이해가 제한적입니다. 그래서 정확도보다는 원리를 직접 확인해 보는 실험 도구로 보시는 게 맞습니다.

SHA 해시 알고리즘으로 평문을 해시문으로 변환하는 과정을 시각화해서 보여주는 교육 자료입니다.

구성 (7단계)

1. 패딩과 블록 분할: 메시지, 0x80, 0 채움, 64비트 길이를 색으로 구분해 512비트 블록 안에 보여줍니다. 블록이 여러 개면 칩을 눌러 블록을 고릅니다.
2. 압축 함수 진행: 64라운드와 마지막 합산 단계를 타임라인으로 보여줍니다. 재생/이전/다음 버튼이나 ←/→, Space로 이동합니다.
3. 작업 변수 a–h: 라운드 전후 값을 보여줍니다. 새로 계산되는 a와 e를 강조하고, Σ1, Ch, K, W, T1, Σ0, Maj, T2 계산식을 실제 값과 함께 표시합니다.
4. 메시지 스케줄: W₀–W₆₃ 표에서 현재 워드와 그 워드가 의존하는 4개 워드(σ0, σ1 계산)를 강조합니다.
5. 비트 단위 논리: Ch(e의 비트로 f/g 선택), Maj(다수결), Σ1(회전 후 XOR)을 32비트 단위로 시각화합니다.
6. 눈사태 효과: 입력 1비트를 바꾼 경우와 비교해, 라운드별로 a–h에서 달라진 비트 수를 그래프로 보여줍니다. 약 128비트로 수렴합니다.
7. 해시값: 최종 다이제스트와 1비트 변경 시의 다이제스트를 나란히 놓고, 달라진 글자를 강조합니다. 복사 버튼이 있습니다.

입력과 예제

• 텍스트(UTF-8) 또는 HEX로 입력하고, SHA-256과 SHA-224 중 선택할 수 있습니다.
• 예제: “abc”, 빈 문자열, 한글 문장, 2블록 예제, HEX 바이트 직접 입력.

검증: Node 표준 crypto 결과와 316개 사례(무작위 입력 300개 포함)가 모두 일치하는 것을 확인했습니다. 페이지에서도 브라우저 WebCrypto 결과와 비교해 일치 배지를 표시합니다.

AES 암호 알고리즘으로 평문을 암호문으로 변환하는 과정을 시각화해서 보여주는 인터랙티브 시뮬레이터 입니다.

구성

• 라운드 타임라인: 초기 AddRoundKey부터 마지막 라운드까지 모든 연산을 칸으로 표시합니다. AES-128 기준 40단계이고, 클릭하거나 재생/이전/다음 버튼(또는 ←/→, Space)으로 이동합니다.
• 이전 상태 → 연산 → 이후 상태: 4×4 상태 행렬에서 칸을 클릭하면 그 바이트의 계산 과정이 나옵니다.
• SubBytes: 16×16 S-box에서 행·열로 값을 찾는 과정을 강조합니다.
• ShiftRows: 테두리 색으로 바이트가 원래 속했던 열을 표시해 이동을 보여줍니다.
• MixColumns: 고정 행렬 × 열 벡터의 GF(2⁸) 곱셈과 XOR 계산식을 보여줍니다.
• AddRoundKey: 라운드 키와의 XOR을 비트 단위로 보여줍니다.
• 키 확장: 전체 라운드 키 목록에서 현재 사용하는 키를 강조하고, RotWord, SubWord, Rcon 계산 과정을 보여줍니다.
• 눈사태 효과: 평문 1비트를 뒤집은 블록과 비교해, 단계별로 달라진 비트 수를 그래프로 보여줍니다. 몇 라운드 만에 약 64비트로 수렴합니다. 상태 행렬의 빨간 점은 두 블록에서 값이 다른 바이트입니다.
• 결과: 암호문을 HEX와 Base64로 보여주고 복사할 수 있습니다.

입력

• 평문과 키는 텍스트 또는 HEX로 입력합니다. 텍스트는 UTF-8로 변환되며, 평문이 16바이트보다 짧으면 PKCS#7 패딩을 붙입니다.
• 키 길이는 128, 192, 256비트 중 선택할 수 있습니다.
• 예제로 '내 텍스트'와 FIPS-197 부록 B, C.1, C.2, C.3이 들어 있습니다.

FIPS-197 시험 벡터 4종(AES-128, 192, 256)으로 결과가 정확히 일치하는 것을 확인했고, 해당 예제를 선택하면 화면에 "기대값과 일치 ✓" 배지가 표시됩니다.

Mechatronics Lab is an interactive browser simulator that shows the core loop of mechatronics: sensor → controller → actuator → mechanism. Each module shows the animated device on one side and a live oscilloscope on the other.

Modules

1. DC Servo + PID: encoder-based position control with adjustable Kp, Ki, Kd, load torque and encoder resolution. Drag the dial to set a target. A live block diagram shows each stage, and the driver is marked when the voltage saturates at ±12 V.
2. Cart-Pole Balancer: an inverted pendulum kept upright by state feedback, with LQR-derived default gains. Click to shove the pole, or move the target to see the non-minimum-phase response, where the cart first moves the wrong way.
3. 2-Link Robot Arm: inverse kinematics driving speed-limited joint servos. The gripper can follow the pointer or a circle, figure-8 or square path. The view shows the reachable workspace, out-of-reach targets and singularity warnings.
4. Stepper Motor: compares wave, full-step, half-step and 16× microstep drive, with a live coil-current sequence table. Raising the step rate or load causes lost steps.
5. PWM + H-Bridge: shows how duty cycle, PWM frequency and inductance affect current ripple and motor speed. Switch states Q1–Q4 and current flow are visualized.
6. IMU Sensor Fusion: a complementary filter combines a noisy accelerometer with a drifting gyroscope. RMS errors are compared live, and you can drag to tilt the board by hand.

Common controls: sliders for each parameter, Space to pause or resume, R to reset. Live readouts show key values such as error, current, speed and lost steps.

MCP 서버를 만드는 과정을 그림으로 설명하고, 직접 도구를 만들어 메시지 흐름을 시뮬레이션해볼 수 있는 한글 HTML 페이지 입니다.

Spring Boot 파이프라인 학습 페이지를 만들었습니다. 페이지는 세 부분으로 되어 있습니다.

• 파이프라인 지도: 트리거 → 빌드 → 테스트 → 패키징 → Docker 이미지 → Dev 배포 → Prod 배포, 이렇게 7단계를 보여줍니다. 단계를 누르면 그 단계가 하는 일, 쉬운 비유, 실제 azure-pipelines.yml 코드가 함께 나옵니다.
• 시뮬레이터: 먼저 트리거를 고릅니다(main push, PR, feature 브랜치). 다음으로 넣을 오류를 고릅니다(컴파일 오류, 테스트 실패, 커버리지 부족). 마지막으로 운영 배포 전 승인 여부를 정하고 실행하면 됩니다. 단계가 실제처럼 진행되고, 로그와 함께 지금 무슨 일이 일어나는지 쉬운 설명이 나옵니다. "한 단계씩" 버튼으로 천천히 따라갈 수 있고, Prod 배포 전에는 승인과 거절을 직접 눌러 볼 수 있습니다. 연습 미션 4개는 해당 상황을 만들면 자동으로 체크됩니다.
• 용어 사전: Stage, Job, Agent, Artifact, Environment 같은 기본 용어를 정리했습니다.

AES 키 공간 페이지를 만들었습니다. 키 길이(128/192/256), 장비 속도, 장비 대수, 양자(그로버) 가정을 바꾸면 아래 항목이 바로 다시 계산됩니다.

• 키 개수: 과학적 표기, 한국어 단위 읽기(예: 2¹²⁸ ≈ 340간), 2ⁿ의 전체 자릿수
• 시간: 평균·최악 소요 시간과 우주 나이 대비 배수, 그리고 안전/위험 판정
• 탐색 비율: 1초, 1일, 1년, 우주 나이 동안 확인할 수 있는 비율을 로그 막대로 표시
• 규모 비교: DES, 모래알, 별, 원자 수와 비교하는 로그 차트
• 가짜 후보 키 수: 알고 있는 평문·암호문 블록 수에 따라 함께 걸러지는 키의 기댓값. 예를 들어 AES-256은 블록 1개만 알면 일치하는 키가 약 2¹²⁸개 남습니다.

Git의 핵심 개념(작업 디렉터리 → 스테이징 → 커밋, 브랜치, 병합, 원격 저장소)을 그림으로 설명하고, 직접 명령어를 눌러보며 따라 할 수 있는 시뮬레이터가 들어간 한글 페이지입니다.