DeepScroll

773 Curated Videos · Brain-Score 75+/100

Tech & Engineering / Tech & Engineering

Curated software engineering, algorithms, electronics, microcontrollers, robotics, GPU architecture, and homelab videos with 3-point Blueprints.

Launch Interactive Feed →
Brain-Score 96/100 Ben Eater · 16:32

Adapting WozMon for the breadboard 6502

  1. Porting Apple-1 WozMon to a breadboard 6502 replaces PIA hardware at $D010 with a 6551 UART at $5000-$5003.
  2. Standard 7-bit ASCII requires changing WozMon's hex-parsing XOR mask from $B0 to $30 since bit 7 is zero.
  3. Keeping the ECHO routine aligned at $FFEF preserves compatibility with original Apple-1 manual test code.
6502 assemblyWozMon ROM monitor6551 ACIA UARTMemory-mapped I/OASCII hex parsing
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Spanning Tree · 9:08

Diffie-Hellman Key Exchange: How to Share a Secret

  1. Diffie-Hellman lets two computers derive a shared symmetric secret key over a public channel without exposing private secrets.
  2. Each side exchanges g^secret mod p publicly; recovering the secret exponent requires solving the hard discrete logarithm problem.
  3. Selecting base g as a primitive root modulo prime p (like 3 mod 19) spans all nonzero remainders against brute-force attacks.
Diffie-Hellman key exchangeSymmetric key cryptographyModular exponentiationDiscrete logarithm problemPrimitive root modulo p
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Branch Education · 38:31

The $200M Machine that Prints Microchips: The EUV Photolithography System

  1. ASML's EUV photolithography system uses 13 nm light across five core subsystems to print ~10 nm (~45-atom) transistor features.
  2. A 6×6-inch Bragg-reflector mask with patterned absorbers accelerates above 7 G while 4× optics shrink patterns onto 300 mm wafers.
  3. Depending on die size, one 300 mm wafer yields ~90 GPUs, 185 CPUs, or 978 DRAM chips before track-tool resist development.
EUV Photolithography (13 nm)Bragg Reflector Reticle4x Projection Optics7G Reticle & Wafer StagesPhotoresist Track Processing
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Spanning Tree · 15:37

AES: How to Design Secure Encryption

  1. Chosen by NIST from Rijndael, AES secures data via Shannon's confusion (complex key dependence) and diffusion (bit-avalanche mixing).
  2. Interleaving non-linear SubBytes S-box lookups with expanded AddRoundKey steps builds confusion that cannot be reversed keyless.
  3. ShiftRows and MixColumns spread single-bit plaintext edits across ~50% of ciphertext bits over AES-128, AES-192, and AES-256 rounds.
Confusion & diffusion (Shannon)KeyExpansion & AddRoundKeySubBytes S-box substitutionShiftRows & MixColumnsAES-128 / 192 / 256 block cipher
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 StatQuest with Josh Starmer · 21:58

StatQuest: Principal Component Analysis (PCA), Step-by-Step

  1. By the Pythagorean theorem, PCA fits PC1 through centered data by maximizing squared distances of projections from the origin.
  2. Singular Value Decomposition yields loading scores showing each variable's weight and eigenvalues measuring captured variation.
  3. Projecting samples onto top principal components and checking the scree plot reveals clusters while gauging 2D plot accuracy.
Principal Component Analysis (PCA)Singular Value Decomposition (SVD)Loading scoresEigenvalues & varianceScree plot evaluation
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 MIT OpenCourseWare · 9:30

21.2.5 Cache Coherence

  1. Multicore caches keep shared variables consistent by snooping a shared bus and incurring bus traffic only during active sharing.
  2. The MESI protocol encodes Modified, Exclusive, Shared, and Invalid line states by repurposing standard valid and dirty bits.
  3. Writing to a Shared line broadcasts an Invalidate bus transaction to purge peer copies before updating locally as Modified.
Snoopy Bus CoherenceMESI State MachineWrite-Invalidate BroadcastMESIF Forward StateCache Line State Bits
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Ben Eater · 27:24

“Hello, world” from scratch on a 6502 — Part 1

  1. Wires a bare W65C02S CPU on a breadboard from its datasheet, tying control inputs like RDY, IRQB, NMIB, and BE to valid logic levels.
  2. Feeds a single-step 555 clock module into PHI2 to inspect processor bus cycles one pulse at a time instead of at megahertz speeds.
  3. Hardwires opcode 0xEA (NOP) onto the data bus and logs the 16-bit address lines with an Arduino Mega to trace reset and PC steps.
W65C02S microprocessorSingle-step 555 clockArduino Mega logic monitorAT28C256 EEPROM0xEA NOP opcode
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Branch Education · 30:19

How does Tap to Pay Work?

  1. Payment terminals emit a 13.56 MHz oscillating magnetic field that wirelessly powers battery-free cards and encodes binary time-slot pulses.
  2. With no internal battery or radio transmitter, the card replies via backscatter load modulation by perturbing the terminal's magnetic field.
  3. Asymmetric cryptography built on mathematical trapdoor functions like 500-digit RSA prime factorization prevents wireless card cloning.
13.56 MHz NFC InductionTime-Slot Binary EncodingBackscatter Load ModulationAsymmetric Public-Key CryptographyRSA Trapdoor Functions
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Freya Holmér · 24:25

The Beauty of Bézier Curves

  1. Bézier curves and splines define smooth paths with analytical tangents and normals for procedural geometry.
  2. Splitting a curve into x and y component functions shows tight bounding-box points lie at endpoints or local extrema.
  3. Finding roots where the x and y derivatives equal zero yields exact t-values for fast analytical bounding boxes.
Bézier curvesSplinesDerivative rootsAnalytical bounding boxProcedural geometry
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Reducible · 28:23

The Fast Fourier Transform (FFT): Most Ingenious Algorithm Ever?

  1. Accelerates polynomial multiplication by converting polynomials from coefficient form into point-value representation.
  2. Evaluates polynomials recursively at complex Nth roots of unity to achieve the Fast Fourier Transform's speedup.
  3. Applies the Inverse FFT with w = e^(-2πi/n) and final 1/n scaling to interpolate point values back to coefficients.
Fast Fourier Transform (FFT)Inverse FFT (IFFT)Nth roots of unityPolynomial multiplicationDivide-and-conquer recursion
Watch in Feed → Full EN/DE Blueprint
Brain-Score 96/100 Ben Eater · 27:50

Astable 555 timer - 8-bit computer clock - part 1

  1. Wires an astable 555 clock on a 5V breadboard using 1 kΩ and 100 kΩ resistors with a 1 µF timing capacitor.
  2. Three internal 5 kΩ resistors form a voltage divider setting comparator thresholds at 1.67 V and 3.33 V.
  3. Dual comparators toggle an internal SR latch as the capacitor charges and discharges to output steady clock pulses.
NE555 timer ICVoltage dividerVoltage comparatorSR latchRC timing circuit
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Jia-Bin Huang · 31:46

Mixture of Experts (MoE), Visually Explained

  1. MoE replaces dense Transformer FFNs with a router activating only a few experts per token, scaling parameters without slowing inference.
  2. Fine-grained designs like DeepSeek-V3 activate 8 of 256 small experts per token alongside an always-active shared generalist expert.
  3. Smooth load-balancing loss equalizes token routing across experts, while Router Z-loss centers logits to prevent training loss spikes.
Sparsely-Gated MoEFine-Grained & Shared ExpertsToken RoutingLoad-Balancing LossRouter Z-Loss
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 CuriousMarc · 20:45

Apollo Guidance Computer Part 1: Restoring the computer that put man on the Moon

  1. Restoring an unpotted Lunar Module development Apollo Guidance Computer lets engineers probe signals directly on its historic ICs.
  2. Technicians remove corroding metal flakes inside Dipstik IC carriers and diagnose faults inside a foam-potted erasable core memory module.
  3. Working without schematics on key parts, the team tests ~5,200 logic gates step by step, including the 120-gate opcode decoder.
Apollo Guidance Computer (AGC)Magnetic Core & Rope MemorySense Amplifiers & TransformersDipstik IC Carrier RestorationGate-Level Opcode Decoder Testing
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Ben Eater · 25:14

Installing the world’s worst video card

  1. Connects a 100x75 64-color VGA circuit to a 6502 computer by replacing static ROM storage with a shared RAM framebuffer.
  2. Pulling Pin 2 (RDY) low halts the 6502 CPU during active display scan so the VGA circuit reads RAM without bus contention.
  3. Releasing RDY high during VGA blanking lets the CPU resume, coordinated via 16-bit address decoding and buffer glue logic.
6502 RDY Pin HaltingVGA Blanking ArbitrationShared RAM Framebuffer16-Bit Memory DecodingTri-State Bus Glue Logic
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Branch Education · 6:07

How do microSD Cards Store 2 TB in such a TINY space?

  1. A 2 TB microSD card houses two offset stacks of eight NAND flash dies, each wafer-ground down to 40 micrometers thick.
  2. Each die packs over 200 vertical layers of 72,000 by 48,000 cells (100 nm wide) storing 3 bits via electrons in silicon nitride.
  3. Alternating silicon oxide and nitride layers are etched vertically, then wire-bonded from staggered die pads to a controller IC.
3D Charge-Trap NAND FlashSilicon Nitride Charge Trap40-um Wafer Back-GrindingOffset 8-Die Wire Bonding3-Bit-Per-Cell Storage
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 mitxela · 19:56

LED Matrix Earring

  1. Packs 52 ultra-tiny 0201 LEDs on a 1 mm pitch into a 9 mm circle using a two-board 4-layer PCB sandwich.
  2. Separating the LED grid from the rear CH32V003 RISC-V controller board eliminates bulky outer trace routing borders.
  3. Stencils solder paste under a microscope, reflows both boards, and retrofits an LR521 coin-cell stud earring frame.
0201 SMD LEDs (1mm pitch)4-layer PCB sandwichCH32V003 RISC-V MCUMicroscope stencil reflowLR521 coin-cell housing
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 w2aew · 13:38

#75: Basics of Opamp circuits - a tutorial on how to understand most opamp circuits

  1. Most op-amp circuits follow two rules: zero current enters the inputs, and negative feedback drives V+ equal to V-.
  2. Placing a capacitor in the feedback loop builds an integrator that turns a constant square-wave input into a triangle.
  3. Real designs must account for supply rail clipping, finite bandwidth, slew-rate limits, and input offset voltages.
Operational amplifier (Op-Amp)Negative feedback (V+ = V-)Virtual ground analysisRC op-amp integratorOscilloscope
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Spanning Tree · 12:39

Understanding B-Trees: The Data Structure Behind Modern Databases

  1. Multi-key B-tree nodes branch widely to keep trees shallow and minimize slow storage lookups in databases.
  2. Balance invariants keep all leaves at equal depth and bound every node between half-full and maximum key capacity.
  3. Full nodes split upward during insertion, while underflowing deletions rotate sibling keys through parents or merge.
B-Tree Data StructureMulti-Way BranchingNode Splitting & MergingSibling Key RotationDatabase Indexing
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 mitxela · 37:43

The Smallest MIDI Synth Again...

  1. Fits a 16-voice polyphonic stereo 44 kHz synthesizer with 14-bit pitch bend inside a MIDI plug powered solely from the DIN port.
  2. Selects an ultra-low-power STM32L4 MCU with a stereo DAC, single-cycle FPU, and dynamic PLL clock scaling for µA/MHz efficiency.
  3. Uses tantalum capacitors on a flex PCB to avoid ceramic DC-bias capacitance loss, noting tantalum's band marks the positive anode.
STM32L4 MicrocontrollerFlexible PCBTantalum CapacitorsMIDI SysEx ProtocolParasitic DIN Power
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 ProjectsInFlight · 34:54

I Saved an Electron Microscope from the Trash

  1. Scanning electron microscopes surpass optical resolution limits because accelerated electrons have far shorter wavelengths than light.
  2. Replacing erratic 5V logic supplies with heatsinked 7805 regulators and freeing a stuck valve restored high-vacuum pumping.
  3. Reaching ~10^-4 mTorr via the oil diffusion pump and setting tungsten filament current to 40–50 µA revived the JEOL JSM-5200.
JEOL JSM-5200 SEM7805 linear regulatorOil diffusion pumpBench power supplyTungsten electron gun
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Branch Education · 18:49

How are Images Compressed? [46MB ↘↘ 4.07MB] JPEG In Depth

  1. JPEG shrinks uncompressed images roughly tenfold by discarding high-frequency color and texture details human eyes miss.
  2. Chroma subsampling, the Discrete Cosine Transform (DCT), and quantization compress pixel blocks while preserving sharp edges.
  3. Video codecs like H.264 combine JPEG-style intra-frames (I-frames) with motion-predicted frames that store only scene changes.
JPEG CompressionChroma SubsamplingDiscrete Cosine Transform (DCT)QuantizationH.264 Intra-Frames (I-Frames)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Branch Education · 23:38

How do Smartphone CPUs Work? || Inside the System on a Chip

  1. A smartphone System on a Chip packs 5 to 10 billion transistors into specialized blocks like the CPU, GPU, ISP, NPU, and modems.
  2. A Network-on-Chip uses routers and 128- or 256-bit parallel buses at 500–1,500 MHz to route data between processing units and DRAM.
  3. Dynamic frequency scaling throttles bus clock speeds on demand to save battery power across nanoscale silicon interconnects.
System on a Chip (SoC)Network-on-Chip (NoC)Neural Processing Unit (NPU)Dynamic Frequency ScalingSilicon Interconnects
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Sebastian Lague · 37:57

Coding Adventure: Ray Tracing

  1. Builds a custom GPU ray and path tracing renderer from scratch in C# and HLSL pixel shaders within Unity.
  2. Traces ray-sphere and ray-triangle intersections while modeling diffuse light with cosine-weighted rays.
  3. Adds progressive rendering, specular glossiness, anti-aliasing, and camera depth of field for realism.
C#HLSL Pixel ShadersUnityLambert's Cosine LawCosine-Weighted Rays
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Luis Serrano Academy · 36:17

The math behind Attention: Keys, Queries, and Values matrices

  1. Transformer attention maps words to multi-dimensional vectors and measures contextual similarity using dot products or cosine similarity.
  2. Scaled dot-product attention divides scores by the square root of vector length to keep numbers small and numerically well-behaved.
  3. Keys, Queries, and Values matrices apply learned linear transformations that project embeddings into richer spaces for multi-head context.
Keys, Queries & ValuesScaled dot-productCosine similarityWord embeddingsMulti-head attention
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Welch Labs · 30:31

The F=ma of Artificial Intelligence [Backpropagation, How Models Learn Part 2]

  1. Paul Werbos's 1970s backpropagation algorithm underpins neural network training up to modern LLMs like Llama 3.2.
  2. Cross-entropy loss with one-hot targets yields output partial derivatives showing which direction lowers model error.
  3. Propagating partial derivatives backward via the chain rule computes parameter gradients for gradient descent steps.
BackpropagationChain Rule CalculusCross-Entropy LossOne-Hot EncodingGradient Descent
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Ben Eater · 36:52

Running MSBASIC on my breadboard 6502 computer

  1. Configures and assembles historical 1977 Microsoft 6502 BASIC source code to target a custom breadboard microcomputer.
  2. Defines the custom ROM, RAM, and zero-page memory map required by the MSBASIC interpreter on homebrew 6502 hardware.
  3. Hooks character input and output routines to the UART serial interface to run interactive BASIC programs from a PC terminal.
MOS 6502Microsoft BASIC (1977)ca65 / cc65 AssemblerMemory Map ConfigurationUART Serial I/O
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Spanning Tree · 9:56

How Floating-Point Numbers Are Represented

  1. Unlike fixed-point binary, IEEE 754 floating-point preserves dynamic range by storing a sign bit, exponent, and mantissa.
  2. Single-precision 32-bit floats subtract a 127 bias from an 8-bit exponent and omit the leading 1 for 24-bit mantissa precision.
  3. Exponent 255 with a non-zero mantissa encodes NaN, while 64-bit doubles expand to an 11-bit exponent and 52-bit mantissa.
IEEE 754 StandardSign-Exponent-MantissaExponent Bias (127)Implicit Leading BitSingle vs Double Precision
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Branch Education · 27:48

How are Microchips Made? 🖥️🛠️ CPU Manufacturing Process Steps

  1. Maps CPU semiconductor fabrication from raw silicon wafer manufacturing and FinFET transistors to multi-layer fab workflows.
  2. Details cleanroom tool stages including EUV photolithography with mask layers, deposition, etching, and ion implantation.
  3. Covers wafer cleaning, nanoscopic metrology inspection, electrical wafer testing, and performance binning of finished chips.
EUV photolithographyFinFET transistorsIon implantationDeposition & etchingSemiconductor metrology
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 SimonDev · 22:18

When Optimisations Work, But for the Wrong Reasons

  1. GPUs rasterize fragments in 2x2 pixel quads for derivatives, executing wasted helper lanes along every triangle edge.
  2. Distant high-poly meshes shrink to sub-pixel triangles that tank quad utilization and bottleneck primitive assembly.
  3. LODs and Unreal Engine 5 Nanite boost speed primarily by eliminating sub-pixel quad overshading rather than vertex math.
GPU 2x2 pixel quad shadingQuad overshading analysisPrimitive assembly pipelineLevel of Detail (LOD) meshesUnreal Engine 5 Nanite
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Veritasium · 26:33

The Most Underrated Algorithm

  1. Explains how the Fast Fourier Transform cuts Discrete Fourier Transform operations from O(N^2) down to O(N log N).
  2. Uses divide-and-conquer rotational symmetries across even and odd sample indices to eliminate redundant math.
  3. Traces the algorithm from Carl Friedrich Gauss in 1805 to 1965 Cooley-Tukey and Richard Garwin's seismic monitoring.
Fast Fourier Transform (FFT)O(N log N) ComplexityCooley-Tukey AlgorithmDivide and ConquerSeismic Signal Analysis
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Breaking Taps · 33:09

Designing Silicon From Scratch

  1. Transport Triggered Architecture (TTA) connects functional units via shared data buses to build a custom CPU in HDL.
  2. Unpipelined multiply-and-accumulate blocks create long critical paths that limit maximum clock speed.
  3. OpenROAD adds reverse-biased antenna diodes to protect transistor gates before detailed routing and timing analysis.
Transport Triggered ArchitectureOpenROAD ASIC FlowHardware Description LanguageAntenna Protection DiodesParasitic Timing Extraction
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Martin Kleppmann · 9:38

Distributed Systems 5.2: Quorums

  1. With independent node fault rate p, all n replicas fail together with probability p^n, yet single-node outages can serve stale reads.
  2. Requiring write quorum w plus read quorum r to exceed total replicas (w + r > n) guarantees read and write node subsets overlap.
  3. When a quorum read spots stale replicas alongside fresh ones, read repair writes the newest timestamped value back to outdated nodes.
Strict Quorum (w + r > n)Read-After-Write ConsistencyClient-Side Read RepairAnti-Entropy SynchronizationReplica Fault Probability (p^n)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 The Synthetic Mind · 2:52

Why Kalman Filters Work So Well

  1. A Kalman filter fuses drifting physics predictions with noisy sensor measurements by representing estimates as Gaussian bell curves.
  2. Kalman gain weighs the innovation (measurement surprise) by relative variance—e.g., variances 4 and 12 yield a gain of 4/(4+12) = 1/4.
  3. Precisions (1/variance) add together (1/4 + 1/12 = 1/3), shrinking combined variance to 3—more accurate than either source alone.
Kalman FilterGaussian State VarianceInnovation (Surprise)Kalman GainAdditive Precision (1/σ²)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 javidx9 · 38:57

Code-It-Yourself! First Person Shooter (Quick and Simple C++)

  1. A 200-line C++ engine renders a playable 3D corridor maze in the Windows console using ASCII characters and a 2D binary map.
  2. Column perspective sets the ceiling by subtracting screen height divided by wall distance from the midpoint, mirroring the floor.
  3. Grid cell checks prevent walking into walls, while ray angles to tile corners flag cell boundaries to delineate wall edges.
ASCII Raycasting2D Binary Grid MapPerspective ProjectionGrid Collision CheckCell Boundary Shading
Watch in Feed → Full EN/DE Blueprint
Brain-Score 95/100 Reducible · 29:23

The Discrete Fourier Transform: Most Important Algorithm Ever?

  1. Samples continuous signals under the Shannon-Nyquist theorem and frames frequency detection as matrix multiplication.
  2. Shows why cosine-only analysis frequencies fail when signal phase shifts, motivating paired sine and cosine waves.
  3. Combines sine and cosine analysis frequencies into complex exponentials to derive the full Discrete Fourier Transform.
Discrete Fourier TransformShannon-Nyquist TheoremMatrix MultiplicationComplex ExponentialsFrequency Domain
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 The Synthetic Mind · 3:00

How Monte Carlo Tree Search Works

  1. Pure random playouts misjudge moves when random players miss forced blocks and winning forks later in the game.
  2. Monte Carlo Tree Search builds a tree via selection, expansion, random rollout simulation, and backpropagation.
  3. Adding a rarity bonus to win rate balances exploration and exploitation until simulations converge on the winning move.
Monte Carlo Tree Search (MCTS)Selection & ExpansionRandom Rollout SimulationBackpropagationExploration-Exploitation Bonus
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 The Synthetic Mind · 3:00

Could Diffusion Language Models Replace LLMs?

  1. Diffusion language models unmask multiple token slots in parallel across passes, leveraging bidirectional context around gaps.
  2. Committing uncertain tokens simultaneously can create mismatched pairs because same-pass guesses cannot condition on each other.
  3. Keeping max doubt times (n + 1) ≤ 1 with computation reuse speeds up open diffusion models up to 27× at minimal accuracy cost.
Diffusion Language ModelsParallel Token UnmaskingBidirectional ContextGroup Doubt RuleBlock Diffusion Speedup
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Acerola · 35:32

I Tried Rendering Millions Of Particles

  1. Generates intricate 3D fractal point clouds on the GPU using Iterated Function Systems (IFS) and the Chaos Game.
  2. Diagnoses why GPU instancing and sub-pixel triangles bottleneck on 2x2 quad overdraw and fragment shader bandwidth.
  3. Moves simulation and point-list rendering into compute shaders with voxel ambient occlusion to draw tens of millions.
GPU compute shadersIterated Function SystemsChaos Game algorithmAffine matrix transformsSub-pixel quad overdraw
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Technology Connections · 18:28

The Antique Toaster that's Better than Yours

  1. The 1948 Sunbeam Radiant Control toaster lowers and raises bread silently via thermal expansion of its center nichrome wire.
  2. Instead of a timer, a thermostat between split coils senses infrared heat radiated from the bread surface at target brownness.
  3. Its unpolarized plug gives a 50% chance internal bus bars stay live when off, requiring a grounded 3-wire cord upgrade.
Thermal ExpansionNichrome WireRadiant ThermostatBimetallic StripGrounded 3-Wire Cord
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 StatQuest with Josh Starmer · 15:52

Gradient Boost Part 1 (of 4): Regression Main Ideas

  1. Gradient Boost regression starts with a single leaf predicting the average target value across all training samples.
  2. Sequential fixed-size trees (often 8–32 leaves) fit the pseudo-residuals—the errors between observed and predicted values.
  3. Scaling each tree by a small learning rate (like 0.1) takes gradual steps that lower variance on unseen test data.
Gradient Boost regressionPseudo-residualsFixed-leaf decision treesLearning rate scalingBias-variance trade-off
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Ben Langmead · 37:01

Burrows-Wheeler Transform

  1. The Burrows-Wheeler Transform permutes a $-terminated string via sorted cyclic rotations or suffix-array left neighbors.
  2. Identical characters keep the same relative rank order in first column F and last column L, making BWT fully reversible.
  3. Rank-offset row jumps invert BWT without storing F, motivating compact rank structures smaller than O(m) suffix arrays.
Burrows-Wheeler Transform (BWT)Suffix Array ConstructionCyclic String RotationsLF-Mapping Rank PropertyFM-Index Space Optimization
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 w2aew · 13:33

#337: Use a NanoVNA to measure input impedance and gain of a small-signal RF amplifier

  1. Full NanoVNA calibration requires Open, Short, 50-ohm Load, Channel 1 Isolation, and Through standards saved to a memory slot.
  2. The NanoVNA's ~600 mVpp output overdrives a small-signal RF amp into nonlinearity, requiring a 13 dB attenuator (~134 mVpp).
  3. Inline attenuation also prevents ~15 dB amplifier gain from overloading Port 1 during S21 gain and 87 MHz 3 dB bandwidth sweeps.
NanoVNASOLT + Isolation CalibrationTermination Insensitive AmpInline RF AttenuationS21 Gain & 3 dB Bandwidth
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Visual Electric · 17:35

AC Power: Basic Principles - Real, Reactive and Apparent Power - what do they REALLY mean?

  1. Real power transfers net energy when electric fields accelerate mobile electrons into fixed lattice ions, modeled as lumped resistance.
  2. Reactive power oscillates around zero at twice line frequency as current leads voltage by 90° in capacitors or lags by 90° in inductors.
  3. Apparent power combines real and orthogonal reactive power; Power Factor (real over apparent power) nears 1 in resistive loads.
Real vs reactive powerLumped-element resistor90° capacitive/inductive phaseApparent powerPower factor (P/S)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Hello Interview · 14:42

Distributed Transactions Explained: 2 Phase Commit vs Saga Pattern

  1. Splitting orders across payment, inventory, and ledger microservices breaks single-database ACID atomicity and isolation.
  2. Two-Phase Commit holds row locks across prepare and commit phases, moving at the slowest node's speed and stalling on crashes.
  3. Production systems favor orchestrated Sagas with idempotent retries, visible compensations (like refunds), and a transactional outbox.
Two-Phase Commit (2PC)Saga PatternOrchestration vs ChoreographyCompensating TransactionsTransactional Outbox
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Spanning Tree · 8:33

How Dijkstra's Algorithm Works

  1. Dijkstra's algorithm sets the source distance to zero and others to infinity, tracking each vertex's best distance and predecessor.
  2. At each step it relaxes outgoing edges to improve neighbor estimates, then explores the unexplored vertex with the smallest distance.
  3. Exploring the smallest estimate first guarantees its shortest path is final, and a priority queue speeds up selecting each vertex.
Dijkstra's Shortest-Path AlgorithmWeighted Graph (Vertices & Edges)Edge RelaxationPriority QueuePredecessor Path Tracking
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Barry Van Veen · 18:56

The Fast Fourier Transform Algorithm

  1. Direct calculation of an N-point Discrete Fourier Transform (DFT) requires on the order of N² complex multiplications.
  2. Radix-2 decimation-in-time FFT splits an N = 2^M sequence into even and odd indices, cutting computation roughly in half.
  3. Recursively splitting across log₂(N) stages to 1-point DFTs combines pairs via butterfly operations for fast analysis.
Discrete Fourier Transform (DFT)Radix-2 Decimation in TimeCooley-Tukey FFT AlgorithmButterfly ComputationEven/Odd Subsequence Splitting
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Luis Serrano Academy · 32:49

Low Rank Adaptation (LoRA) - How to search for a needle in a much smaller haystack

  1. LoRA fine-tunes large language models efficiently by constraining weight updates to a low-dimensional subspace around the base model.
  2. Because pre-trained weights already share strong dependencies, a high-rank update matrix can be approximated by two thin matrices.
  3. Singular Value Decomposition (SVD) motivates low-rank matrix approximation, slashing trainable parameters while preserving accuracy.
Low-Rank Adaptation (LoRA)Low-Rank Matrix ApproximationSingular Value Decomposition (SVD)Subspace Dimensionality ReductionMatrix Rank Factorization
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Core Dumped · 21:29

How a Single Bit Inside Your Processor Shields Your Operating System's Integrity

  1. A hardware mode bit inside the CPU separates unprivileged user mode from kernel mode, blocking direct hardware tampering.
  2. Applications request OS services via system calls and interrupts, triggering controlled hardware transitions into kernel mode.
  3. Timer interrupts enforce preemptive multitasking, while buggy kernel-mode drivers or anti-cheat code can crash the entire OS.
CPU Mode BitKernel & User ModeSystem Calls (Syscalls)Hardware InterruptsPreemptive Multitasking
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 The Signal Path · 35:10

TSP #349 - Hacking a Giant Cryocooler! Custom Hardware & GUI for Stirling Ultracold Refrigerators

  1. A 1.2 kW free-piston Stirling cryocooler reaches below -80 °C while rejecting heat at +50 °C across a sealed helium cold head.
  2. Analyzing Debian controller binaries and serial traffic reveals the command protocol driving piston power and heater states.
  3. Custom RTD sensor interface hardware, safe firmware limits, and a PC GUI unlock manual cryogenic control beyond freezer presets.
Free-Piston Stirling CryocoolerSerial Protocol Reverse Eng.RTD Temperature TelemetryCustom Control Firmware & GUI
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Leo Isikdogan · 9:17

How Diffusion Models Work

  1. Diffusion models train a U-Net to predict total noise added to clean images, then iteratively subtract scaled noise during sampling.
  2. Latent diffusion freezes a pre-trained VAE so denoising runs efficiently in compact latent feature maps before pixel decoding.
  3. Classifier-free guidance sharpens prompt alignment by scaling the gap between conditional and unconditional or negative predictions.
Forward & Reverse DiffusionU-Net Noise PredictionLatent Diffusion & VAEClassifier-Free GuidanceNegative Prompting
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 w2aew · 11:47

#276: Smith Chart: Design an L-Network - Impedance Matching Circuit

  1. A two-element L-network uses one series and one shunt inductor or capacitor to match a complex load to 50 ohms.
  2. Adding series reactance rotates impedance along constant-resistance circles on the Smith chart toward the center match.
  3. Measuring the circuit on a VNA confirms the 50-ohm match while showing slight phase rotation from fixture line length.
Smith ChartL-Network MatchingVector Network Analyzer50-Ohm Impedance MatchSeries & Shunt LC
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 PaPoC Workshop · 10:57

Paxos vs Raft: Have we reached consensus on distributed consensus? — Heidi Howard

  1. Both Paxos and Raft implement state machine replication by agreeing on a total order of log operations across distributed servers.
  2. Both ensure a new leader holds all committed entries: Paxos transfers missing logs in votes, while Raft restricts who wins votes.
  3. Presented in unified terminology, core Paxos and Raft are equally understandable, though Raft's leader election is more efficient.
PaxosRaftState Machine ReplicationDistributed ConsensusLeader Election
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Tom Scott · 6:22

How Computers Compress Text: Huffman Coding and Huffman Trees

  1. Fixed 8-bit character encoding enables fast text indexing, whereas lossless compression assigns shorter bit sequences to common characters.
  2. A Huffman tree repeatedly merges the two least frequent characters and sums their counts until a single root node remains.
  3. Tracing left branches as 0 and right as 1 yields optimal single-character prefix-free codes that decode unambiguously without separators.
Huffman CodingHuffman TreeLossless CompressionPrefix-Free Bit CodesCharacter Frequency
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Animagraffs · 4:50

How Jet Engines Work

  1. Spinning rotor blades and stationary stator vanes compress incoming air across stages before fuel ignition in the combustor.
  2. Hot exhaust drives rear turbines that spin both the core compressor stages and the large front ducted turbofan.
  3. High-bypass turbofans move slow air around the core for quiet efficiency, while low-bypass jets favor speed and afterburners.
Rotor-Stator CompressorAnnular CombustorTurbine Air CoolingHigh vs. Low Bypass RatioAfterburner Nozzle
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Umar Jamil · 26:56

LoRA: Low-Rank Adaptation of Large Language Models - Explained visually + PyTorch code from scratch

  1. LoRA freezes pre-trained weights W (d×k) and models updates via low-rank matrices B (d×r, zero-init) and A (r×k, Gaussian).
  2. Because fine-tuning updates have low intrinsic rank, small r cuts trainable parameters below 1% for fast backprop.
  3. Scaling by fixed alpha stabilizes rank changes, and swapping small A/B matrices switches tasks without reloading W.
LoRA (Low-Rank Adaptation)Intrinsic Rank DecompositionZero / Gaussian Matrix InitAlpha-Over-Rank ScalingPyTorch Parametrization
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 First Principles of Computer Vision · 7:19

BRDF: Bidirectional Reflectance Distribution Function

  1. BRDF models material appearance as the ratio of reflected surface radiance L to incident irradiance E across two directions.
  2. Incident and reflected rays each use zenith (theta) and azimuth (phi) angles, making general BRDF a 4D function.
  3. Helmholtz reciprocity holds when swapping light and view angles; isotropic surfaces reduce to 3D, unlike grooved anisotropic ones.
BRDF (L / E Ratio)Radiance & IrradianceHelmholtz ReciprocityIsotropic vs. AnisotropicSpherical Angles (theta, phi)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 Denis Magda · 16:06

PostgreSQL Internals in Action: MVCC

  1. PostgreSQL MVCC avoids overwriting rows in place, storing new tuple versions tagged with hidden xmin, xmax, and ctid columns.
  2. Using pageinspect reveals ctid version chains and Heap-Only Tuple (HOT) flags that bypass index updates on unindexed changes.
  3. Concurrent sessions read consistent snapshots until commit, after which VACUUM garbage-collects obsolete dead row versions.
PostgreSQL MVCCHidden Columns (xmin, xmax, ctid)Snapshot Isolationpageinspect ExtensionHeap-Only Tuples (HOT)
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 EEVblog · 34:14

EEVblog #479 - Opamp Input Bias Current

  1. Real op-amp inputs draw finite bias currents (IB+ and IB-) that create unwanted DC voltage drops across feedback resistors.
  2. Adding a bias resistor equal to R1 in parallel with RF cancels matched drops, but input offset current (IOS) error still remains.
  3. Bench tests on a rail-to-rail op-amp show picoamp bias currents changing magnitude and reversing sign from 5V down to 2.7V.
Operational AmplifierInput Bias Current (IB)Input Offset Current (IOS)Bias Compensation ResistorPicoamp Bench Measurement
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 ByteByteGo · 5:38

Bloom Filters | Algorithms You Should Know #2 | Real-world Examples

  1. A Bloom filter uses a compact bit array to test set membership with zero false negatives and tunable false positives.
  2. Inserting an element hashes it across multiple functions to set bit positions to 1; any 0 bit on lookup guarantees absence.
  3. Databases and CDNs use Bloom filters to avoid costly LSM-tree disk reads for missing keys and skip one-hit-wonder cache writes.
Bloom FilterBit Array & Hash FunctionsFalse-Positive Trade-offLSM-Tree Read FilteringCDN Cache Admission
Watch in Feed → Full EN/DE Blueprint
Brain-Score 94/100 MATLAB · 10:32

Understanding Field-Oriented Control | Motor Control, Part 4

  1. Unlike six-step control (60°–120° angle), FOC keeps the stator field 90° orthogonal to the rotor for peak smooth torque.
  2. Clarke and Park transforms map three-phase AC stator currents into rotating direct (d) and quadrature (q) DC axes.
  3. Two PI loops hold direct current at zero and regulate quadrature current before inverse transforms drive motor phases.
Field-Oriented Control (FOC)Clarke & Park Transformsd-q Rotating FramePI Control LoopsPMSM Motor Control
Watch in Feed → Full EN/DE Blueprint