Xerox Photocopy Effect Tutorial in AE (No plugins needed)... 1. Import Image 2. Adjustment Layer 3. Noise 4. Gaussian Blur 5. CC Threshold 6. Tint 7. Posterize Time #niqteworks #aftereffects #ae #motiongraphics #motiondesign #mv #tutorialshow more

Dither / halftone effects tutorial in After effects (No... plugins needed) 1. Create a Shape Layer (dot/circle) 2. Pre-comp it → add CC Tiler 3. Import your image/video 4. Add Card Dance to an Adjustment Layer 5. Set Gradient Layer 1 to your footage 6. Set X & Y Intensity to 1 Adjust Rows & Columns to tasteshow more

FINAL: Midwest Regional No.1 Game 6: No.1 seed GVSU... Baseball 7, No.4 seed Missouri S&T 4 FOR THE FIRST TIME IN PROGRAM HISTORY, THE GRAND VALLEY STATE LAKERS ARE HEADED TO A SUPER REGIONAL! GVSU Lakers Grand Valley State had a 4-run 8th to decide the game. Jovan Gill - 2/3, 3B, HR, 3 RBI Grand Valley State will host either Illinois Springfield/UIndy/Missouri-St. Louis on Thursday (5/21) in the Midwest Super Regional. #D2Baseballshow more

Exploring diff visuals with Gen2. I'm amazed at how... it takes a single block image to build neon streets, structures, and entire cityscapes. Anybody else toying with AI tonight⁉️ AI Art Workflow: 1. Crafted starting image in #Midjourney 2. Generated a 4-sec video with #Gen2 (no text prompt) 3. Extended first 4 secs to 18 secs 4. Repeated steps 2-3 twice, each time using the final frame from the previous 18-sec video to start the next 18-secs run 5. Merged all three 18-sec videos in #FinalCut Pro 6. Adjusted speed for fluidity 7. Final version published here is a 54-sec journey distilled into a 16-sec video with 🎶 #AIart #AIArtCommunityshow more

IF I WAS FORCED to build a $20K/month AI... creative agency using nothing but Photoshop, starting from 0, here's exactly what I would do in steps: The production setup (Days 1–3) 1. Download the Higgsfield plugin inside Photoshop — takes 5 minutes 2. You now have: sketch-to-image, layer decomposer, mockup studio, relight, upscale, face swap, character swap, background removal, AI stylist — all in 1 tool 3. Old creative agency workflow: designer + photographer + editor + 3–5 day turnaround 4. New workflow: 1 person, Photoshop, 30 minutes per deliverable The offer (Days 3–7) 5. Pick 1 niche — ecom brands, real estate agents, or course creators all need visuals constantly 6. Build a simple offer: "10 ad creatives delivered in 24 hours — $500" 7. Old agencies charge $2,000–$5,000/month for the same output 8. Your cost to deliver: $0 beyond the plugin. Pure margin. 9. Create 3 sample mockups using the tool — drop a product image in, generate 9 variations, pick the best 3 10. That's your portfolio. Built in under 1 hour. Cost: $0. The client machine (Days 7–20) 11. Go on X and search "[niche] + need a designer" or "[niche] + creatives" 12. DM 50 people per day — "I'll make you 3 free ad creatives in 24 hours, no catch" 13. Deliver them in 30 minutes using the plugin 14. 50 DMs/day × 14 days = 700 outreach messages 15. Conservative 3% conversion = 21 people see the free work 16. Close 5 of them at $500 = $2,500 in week 3 The scale (Days 20–30) 17. Upsell every client to a $1,500/month retainer — 10 creatives/week, unlimited revisions 18. 1 client per day in Photoshop takes 45 minutes max 19. 10 retainer clients × $1,500 = $15,000/month 20. Add 3 one-off clients at $500/month = $1,500 21. Add a $997 "AI creative system" course teaching other people this exact workflow = $3,000+/month from 3 sales The math: 50 DMs/day × 30 days = 1,500 outreach messages 3% book a call = 45 calls 40% close at $1,500/month retainer = 18 clients 18 × $1,500 = $27,000/month recurring Time per client per day: 45 minutes Total daily work: 4–5 hours Every mockup — AI. Every restyle — AI. Every layer rebuild — AI. Every variation — AI. No photographer. No designer and no reshoot. Start it here. 👇show more

Claude + Meta + GPT Image 2 = AI... UGC FACTORY Here's the free workflow Real UGC creators charge $150-$500 per video. 5-7 day turnaround. By the time you get 12 variants, your winning angle is already dead. We built a system that does 12 unique video ads in under an hour. Cost per video? Less than your coffee. The stack: • HeyOz (product import + export hub) • Claude (creative director + prompt engine) • GPT Image 2 (photorealistic starting frames) • Seedance 2.0 (animation) Broke down the entire thing in a step-by-step guide. What you'll get: → The exact 5-step workflow from URL to finished ad → 3 copy-paste image prompts (Stanley Quencher example) → 3 matching motion prompts for Seedance → The prompt framework that fixes 90% of bad AI outputs → Why grip specificity is the make-or-break variable most people miss → The 70/30 strategy for scaling (AI for testing, humans for winners) No design skills needed. No Canva. No CapCut. No Adobe. Just paste your product URL. Generate. Export. Launch. If you're running Meta or TikTok ads and your creative rotation is slower than your ad spend, this is for you. Comment "factory" to get access (must be following)show more

Batch Normalization by hand ✍️ ~ 7 steps walkthrough... below Batch normalization is common practice for improving training and achieving faster convergence. It sounds simple. But it is often misunderstood. 🤔 Does batch normalization involve trainable parameters, tunable hyper-parameters, or both? 🤔 Is batch normalization applied to inputs, features, weights, biases, or outputs? 🤔 How is batch normalization different from layer normalization? So I drew and calculated one entirely by hand. Goal: normalize a mini-batch of 4 examples to mean 0 and variance 1, then let the network scale it back. = 1. Given = A mini-batch of 4 training examples, each with 3 features. = 2. Linear layer = Let us multiply by the weights and add the biases. Batch norm sits after this, which answers the second question: what gets normalized is features, not inputs, weights or biases. = 3. ReLU = We apply the activation, and -2 becomes 0. Negative values are suppressed before any statistic is taken. = 4. Batch statistics = Let us compute the sum, mean, variance and standard deviation, one row at a time. A row is a feature and the four columns are the four examples, so every number here measures one feature against the rest of the batch. That is the "batch" in batch normalization, and it is exactly what layer normalization does not do. The statistics are rounded to whole numbers, which is what keeps the rest of the page doable in pen. = 5. Shift to mean 0 = We subtract the mean, in green. The four values in each feature now average to zero. = 6. Scale to variance 1 = Let us divide by the standard deviation, in orange. Each feature now has variance one, whatever scale it arrived at. = 7. Scale and shift = We multiply by a linear transformation and pass the result on. The diagonal and the last column are trainable, so having just forced every feature to mean 0 and variance 1, we hand the network the means to undo it. The outputs: Mean of each feature = [2, 1, 2] Std dev of each feature = [1, 1, 2] To the next layer = [2, -2, 2, 0], [-3, 3, 6, -3], [2, 0, 1, 2] The answers: 🤔 Both. The scale and shift are trainable, the statistics are not. Epsilon and the momentum on the running statistics are the hyper-parameters, and one mini-batch by hand needs neither. 🤔 Features, after the linear layer, not inputs, weights or biases. 🤔 Batch norm measures across the batch, one feature at a time. Layer norm measures across the features, one example at a time. 💾 Save this post!show more

Here we go GPT Image 2 and Seedance 2.0... is now live on insMind #insmind #insmindai Generated this GRWM video using the Prompt : Aesthetic “Get Ready With Me – Gym Edition” storyboard layout, minimal neutral-toned design, soft beige and cream color palette, clean editorial grid. Top header text: “GET READY WITH ME” subtitle: “gym edition” in elegant script font subheading: “step-by-step activewear routine” The layout is divided into 4 blocks, each showing a sequence (1–8 steps per row), featuring the same young woman throughout with consistent face, natural makeup, athletic toned body, hair tied in a messy bun or sleek ponytail. BLOCK 1 – BASE (1–8) 1–2: putting on a fitted sports bra 3–4: wearing high-waisted gym leggings 5–6: adjusting waistband / smoothing fit 7–8: mirror check, relaxed confident pose BLOCK 2 – LAYERS (9–16) 1–2: putting on oversized gym t-shirt or cropped top 3–4: adding lightweight zip-up hoodie or jacket 5–6: tying hair tighter / adjusting outfit 7–8: slight movement pose (stretching arms or twisting body) BLOCK 3 – DETAILS (17–24) 1–2: wearing smartwatch / fitness band 3–4: adding minimal jewelry (thin chain, studs) 5–6: putting on gym gloves or lifting straps 7–8: wearing sunglasses or tying hair final look BLOCK 4 – FINISH (25–32) 1–2: putting on training shoes (clean white sneakers) 3–4: grabbing gym bag / water bottle 5–6: holding headphones / protein shaker 7–8: full-body mirror shot, confident final look Side icons representing categories: base, layers, accessories, shoes, final look. Soft natural lighting, indoor minimal room or modern apartment, neutral background, clean shadows, editorial fashion photography style, consistent framing across all panels. Footer text: “You’re ready. Go own your workout.” Video prompt : Use provided storyboard image as reference CONCEPT: Get Ready With Me — Gym Edition TIMELINE: 0 : 00–0:04 Sports bra on High-waisted leggings wear Waistband adjustment Mirror check 0 : 04–0:08 Oversized tee / cropped top Lightweight hoodie or jacket Hair tie (ponytail/bun) Light stretch / body turn 0: 08–0:12 Smartwatch / fitness band Minimal jewelry Gym gloves / lifting straps Sunglasses on 0: 12–0:15 Training shoes Grab gym bag + water bottle Headphones / shaker Walk-out + final confident look STYLE: Minimal, neutral tones, soft beige/grey palette, natural indoor lighting, clean modern interior, athletic aesthetic CAMERA: Close-up + mid shots, soft focus, shallow depth of field, steady framing, subtle handheld realism TRANSITIONS: Match cuts, outfit snap transitions, fabric motion cuts, quick clean jump cuts synced to movement OUTPUT: Loopable, smooth pacing, satisfying flow, social media ready (vertical 9 : 16show more

GPT-Image-2 + Seedance 2.0目前已成AI视频标配 甚至可以根据给定图片推导过去和未来，制作storyboard，然后生成视频 使用方法： 1️⃣ 随便找一张图 2️⃣... 给以下提示词，然后制作storyboard 用以下提示词👇： Create a 3×3 cinematic storyboard grid based on the uploaded reference image. Use the uploaded image as the central moment of the story: Frame 5 must represent the exact “t” moment, matching the subject, scene, mood, composition, costume, environment, lighting style, and emotional tone of the reference image. The storyboard must show what happened before and after this moment as a time-based visual timeline. FRAME STRUCTURE: Frame 1: t-30: Establishing shot, the wider environment before the main event begins. Frame 2: t-10: The subject approaches or prepares for the key moment. Frame 3: t-5: Tension builds, body language and atmosphere lead toward the reference image. Frame 4: t-1: Final instant before the reference image, close emotional or action transition. Frame 5: t: Recreate the uploaded reference image as the central key frame. Frame 6: t+1: Immediate reaction or continuation right after the key moment. Frame 7: t+5: Alternate angle showing the consequence of the moment. Frame 8: t+15: Candid transition frame, natural movement, emotional aftermath. Frame 9: t+30: Strong final cinematic frame that clearly resolves the scene. STYLE: Ultra-realistic cinematic storyboard, 3×3 grid layout, cohesive visual tone across all frames, consistent character identity, consistent costume, consistent environment, cinematic lighting, shallow depth of field, realistic camera angles, natural motion continuity, no text labels, no numbers, no arrows, no captions inside the image. 3️⃣ seedance2.0 一键成片show more

a🧵on how to do the movable crater tech, You... will learn this step by step, It should only take a couple of minutes to do it so lock in, Simple Overview: 1) Draw a black texture that represents a crack (use lasso or draw then fill tool) 2) Use a png to svg converter to make it into svg (you MUST make the crack black!) 3) Make the mesh of the crater 4) Export it as fbx and Import it in Roblox Studio 5) Create a viewport put the thing inside play with some lighting,material of crater n ur done 6) Make another viewport and make it so that it duplicates the character of the player every frame and puts in the viewport so that you can see it to be above the crater viewport Now for the detailed tutorial:show more

1/7 Built a Polymarket trading bot over 3 months.... Here are the biggest mistakes that cost me real money > Went from v1 to v61. Every version fixed something painful. --- 2/7 Stop Loss killed more money than it saved. > Binary markets need room to breathe - fluctuations are normal. > Stop Loss was cutting positions on random noise and locking in losses right before the market flipped. > Removed it in v61. Immediately better. --- 3/7 Martingale + Stop Loss = a loss cascade. > Seemed logical: lost $5 -> bet $8, lost again -> bet $10. > In practice: a losing streak plus early exits = a hole in your balance in a single day. > Killed it. For good. --- 4/7 Smart Exit without Force Exit is a trap. > Token hits 90c (+75% profit), but the bot was waiting for a "BTC reversal" signal. > Market closes, token drops, profit gone. > Fix: hard Force Exit at 85c. No conditions, no waiting. --- 5/7 Blocking the 5:30-10:30 PM ET window felt safe. It wasn't. > NYSE open = sharp spikes = bad signals. Made sense to block it. > But the full block was also killing clean entries at 8-10:30 PM. > Had to split the zone into segments with different edge/move thresholds. --- 6/7 The Gamma API lies about market start time. > Start price ("price to beat") is the core input for every signal. > Gamma was returning stale data. Had to pull prices directly from Chainlink on-chain on Polygon. > That's its own adventure - polling a smart contract every 2 seconds at 2 AM. --- 7/7 The real lesson: don't overcomplicate what works. > v1: complex system, 10 indicators -> -$200/day > v61: "buy the expensive token for $5, exit at +30%" -> consistently green > Simpler logic = fewer failure points. > The bot runs 96 intervals a day. Every mistake shows up fast.show more

I created this desk calendar as a source of... inspiration for anyone learning AI in 2026. It includes 24 AI algorithms and architectures, all drawn and calculated by hand. ✍️ 𝗝𝗮𝗻𝘂𝗮𝗿𝘆: [1] Matrix Multiplication; [2] Discrete Fourier Transform (DFT) 𝗙𝗲𝗯𝗿𝘂𝗮𝗿𝘆: [3] Support Vector Machine (SVM); [4] Vector Database 𝗠𝗮𝗿𝗰𝗵: [5] Multi-Layer Perceptron (MLP); [6] Backpropagation 𝗔𝗽𝗿𝗶𝗹: [7] Batchnorm; [8] Dropout 𝗠𝗮𝘆: [9] Recurrent Neural Network (RNN); [10] Long-Short Term Memory (LSTM) 𝗝𝘂𝗻𝗲: [11] Residual Network (ResNet); [12] Graph Convolutional Network (GCN) 𝗝𝘂𝗹𝘆: [13] Autoencoder; [14] Variational Autoencoder (VAE) 𝗔𝘂𝗴𝘂𝘀𝘁: [15] Generative Adversarial Network (GAN); [16] U-Net 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿: [17] Transformer; [18] Self Attention 𝗢𝗰𝘁𝗼𝗯𝗲𝗿: [19] Reinforcement Learning with Human Feedback (RLHF); [20] Contrastive Language-Image Pre-training (CLIP) 𝗡𝗼𝘃𝗲𝗺𝗯𝗲𝗿: [21] Diffusion Transformer; [22] Switch Transformer 𝗗𝗲𝗰𝗲𝗺𝗯𝗲𝗿: [23] Sparse Autoencoder; [24] BitNetshow more

Gemini Omni is changing video creation and most people... still don't know what it can actually do. Here are 7 things you can do with it right now: 1. Image to Video in multiple styles: Upload a photo. describe the style. Gemini Omni turns it into a video. cinematic, anime, realistic. 2. Reimagine camera angles: Take any image and visualize it from a completely different angle. no reshooting. just describe it. 3. Change backgrounds: Swap the background of any video through plain conversation. no green screen. no editing software. 4. Change hairstyle: Describe a new hairstyle. Gemini Omni applies it. characters stay consistent across every frame. 5. visualize with moodboards: feed it a moodboard, audio, and a brief. it builds the video around the full creative reference — not just a text prompt. 6. swap characters and objects mid-scene: replace people, props, and objects in an existing video. the scene stays consistent. the physics hold up. 7. edit physics: describe how you want things to move. Gemini Omni adjusts the physics of the scene through conversation.show more

I spent 48 hours running AI from my phone.... Here are 11 things that turned out to be possible and 3 that almost cost me money Forgot my laptop at home and thought the day was lost. Opened a terminal from my phone and decided to see how long I could last Lasted 2 days. Not just lasted but made $840 What works from a phone: 1. Set up Claude Code through SSH in 10 minutes while riding the subway 2. Get Telegram pushes every time a wallet enters a position 3. Copy a trade with 1 tap without taking out my earbuds 4. Launch scripts by voice through Shortcuts 5. Monitor 3 wallets simultaneously without a single lag 6. Get a morning report at 7 AM as a regular message 7. Rebuild the bot when it crashed while sitting in a cab 8. Check PnL without opening a browser 9. Add a new wallet to tracking in 30 seconds 10. Set up auto-copying without confirmation on verified wallets 11. Get a full strategy breakdown of a wallet through Claude Code in a regular chat And here is what almost killed the deposit: 1. My finger slipped and I entered at twice the planned size. Did not notice for 20 minutes. Got lucky that the position ended up in profit but it could have gone very differently 2. My phone died at 2 AM. Missed the exit signal and the position dropped $110 while I slept. By morning I realized that a power bank is just as much a part of the strategy as the bot itself 3. The delay when copying was 40 seconds. On a 15-minute market that is an eternity. The price moved from 8 cents to 23, and instead of a 12x return I got a 4x. Still profit but you feel the difference immediately Total for 48 hours: +$840. Screen time on the phone: 47 minutes. Never needed the laptop The entire time I was following the same wallet. That is the 1 that was sending me signals at 2 AM: The phone turned out to be a fully functional control panel. But this control panel has no safety switch. And that is worth remembering every time you are tapping with 1 hand in the coffee lineshow more

In my class, I teach the autoencoder by asking... everyone to stand up. 🙆 Stretch your arms out wide. Imagine you are holding a heavy textbook (like Introduction to Algorithms by Prof. Cormen), the whole thing, every page. Now bring your hands slowly together until they almost touch your neck. The bottle "neck." The final exam is tomorrow and you are allowed one cheat sheet: whatever you can scribble on your palm. All nine hundred pages have to survive the squeeze. That is the encoder. Now imagine you sit in the exam. Push your arms back out to where they started. You try to rebuild the textbook from your palm notes. That is the decoder. Of course you cannot get every page back. What you get back is what mattered enough to write down, and the gap between the two is the loss the network is trying to shrink. I call it AI by Arms 🙆. It gets a laugh, and then it gets remembered. Goal: squeeze four numbers down to two, then rebuild the original four from them. 1. Given Let us start with four training examples: X1, X2, X3, X4. 2. Auto (copy to targets) We copy the training examples straight into the targets. That is the whole trick behind the name: "auto" is Greek for "self", and the data is its own label. 3. Encoder, layer 1 Let us multiply the inputs by the weights, add the biases, and apply ReLU. Negative values get crossed out and become zero. 4. Encoder, layer 2 (the bottleneck) We do it again, and now the four dimensions have become two. This layer is called the bottleneck, because everything has to fit through it. 5. Decoder, layer 1 Let us go back the other way: multiply, add, ReLU. This time there are no negatives to cross out. 6. Decoder, layer 2 We multiply once more and get the outputs Y. This is the decoder's attempt to rebuild the four original numbers from the two it was given. 7. MSE loss gradients Let us compare Y with the targets Y'. The gradient is 2 x (Y - Y'): subtract, then double. Those gradients kick off backpropagation, and the weights start to learn. Your entire education is all about encoding and decoding!show more

Met my girlfriend's parents for the first time. Her... dad asked what I do for work. I said I build trading systems. He said like Wall Street? I said no. 6 AI agents. They work while I sleep. He laughed. So robots are making you money? I did not argue. I opened my laptop. Showed him the terminal. 6 agents running. 47 mispriced markets caught in the first week alone. His face changed. That is not gambling. That is automation? Exactly. Then I showed him how it works. Built the whole thing in 6 hours. Agent 1: Monitoring Runs 24/7. Watches Polymarket for mispriced markets. Spots an anomaly. Writes to memory and pings me on Telegram instantly. Agent 2: Research Parses news, X, macro data via browser tool on a cron schedule. Every morning I have a full digest on all open positions before I check my phone. Agent 3: Trading Reads the research agent memory. Sees the market has not reacted yet. Acts. Execution tool in gateway mode with a whitelist. No full access on a live server. Agent 4: Watchdog Heartbeat every 5 minutes. Monitoring running. No errors. Positions up to date. Something breaks. Immediate Telegram message. All of this. One Gateway. One config file. Isolation via per-agent scope. The token trick: stopped dumping everything into one file. Critical rules in bootstrap. Markets, patterns, past trades in memory. Semantic search pulls it when needed. Token spend dropped 3x. From $0.40 per request to $0.13. First week running: → 47 mispriced markets caught before Polymarket adjusted → Average entry edge 8 to 12 cents per position → Watchdog fired 3 times and caught a broken RPC before it cost me anything The whole system is plain text files. Open an editor. Change one line. Agent behaves differently. No deploy. No build. Her dad went quiet. Then he asked can you teach this? Her mom asked for the setup guide. I built the entire framework. Six agents. Full deployment. Memory architecture. Telegram alerts. You only need Claude + device + 1 hour per day. Giving this free for 24 hours. To get it: 1. Comment the word "Claude" 2. Like and retweet this 3. Follow me Himanshu Kumar so I can DM you Save this post. Deploy the 6-agent system this week. Start with $200. Scale on evidence.show more

Made this cinematic AI video in minutes using Getvivix... Prompt used below 👇 STORYBOARD 1 "THE KNIGHT" PROJECT TYPE: 10-second cinematic fantasy storyboard CHARACTER LOCK: single consistent knight — original fictional STYLIZED fantasy warrior (not a real person). Full ornate plate armor, VISOR DOWN the entire sequence (face never visible — safe by design), tattered surcoat + banner, mounted on an armored warhorse. Identical armor/horse across all frames. STYLE: epic dark-fantasy, cinematic, painterly film stills PACING & FLOW: slow, weighty, EPIC — no rush. One continuous charge → clash → melee → hero arc. Gradual camera moves; the action carries unbroken from frame to frame (each beat is the next instant of the last). Transitions are match-on-motion — the horse's stride and the sword's arc bridge every cut, never a hard jump. FRAMES (8 shots, 0–10s) — angle | lens | motion | lighting | environment | → into next: 1 (0–1.5s): wide establishing | 24mm | knight reined at a hill crest, banner snapping, slow push-in | cold dawn backlight, mist | battlefield below → camera drifts down as the horse shifts weight 2 (1.5–3s): 3/4-rear tracking | 35mm | horse breaks into a canter down the slope | low sun raking | churned mud, distant ranks → match-on-stride into the gallop 3 (3–4.5s): side tracking | 50mm | full gallop toward the enemy line, dust plume | side rim light, haze | spears + banners ahead → he lowers the lance, carrying the motion 4 (4.5–6s): low-angle hero | 35mm | lance leveled mid-gallop, visor catching light | backlit dust glow | closing on the line → impact begins 5 (6–7s): impact wide | 50mm | lance strikes, enemy hurled back, splinters | harsh flash + sparks | clash of the lines → horse rears from the hit 6 (7–8s): low 3/4 | 35mm | warhorse rears amid the melee, sword drawn | embers, torchlight | swirling battle → the blade sweeps down 7 (8–9s): tracking the blade | 50mm | sweeping arc through foes, motion-blur trail | sparks on steel | bodies + banners → camera settles, pulls back 8 (9–10s): hero hold | 24mm | horse reared, sword raised, banner behind, silhouette | dramatic backlight, battle haze | the field beyond → freeze LAYOUT: film sheet — left: 3 dynamic mounted poses (charging 3/4, rearing, mid-swing — in-scene, visor down); center: 8-frame grid; right: director notes; bottom: 0–10s. VISUAL STYLE: cinematic dark-fantasy, painterly, volumetric dawn light, dust + embers + mist, shallow DOF, motion blur, anamorphic; stylized — NOT photorealistic, not real human skin; FACE NEVER SHOWN (visor down). Seedance on Getvivix lets you generate high-end cinematic visuals for around 1000 credits (~$1), making pro-level video creation cheap and scalable. Try it here:show more

Transformer by hand ✍️ ~ 6 steps walkthrough below... Open the hood of a transformer and the parts list is overwhelming: embeddings, positional encoding, attention weighting, self-attention, cross-attention, multi-head attention, layer norm, skip connections, softmax, linear, Nx, shifted right, query, key, value, masking. Which of those actually make the car run? Two of them. Attention weighting and the feed-forward network. Everything else is an enhancement to make it run faster and longer, which is how we got from a car to a truck, and to the word "large" in large language model. So I drew and calculated those two parts entirely by hand. Goal: push five features through one transformer block, filling in every cell yourself. 1. Given Five positions of input features, arriving from the previous block. 2. Attention matrix Let us feed all five features to a query-key module (QK) and read back an attention weight matrix, A. The details of that module are a post of their own. 3. Attention weighting We multiply the input features by A to get the attention weighted features, Z. Still five positions. The effect is to combine features *across positions*, horizontally: X1 becomes X1 + X2, X2 becomes X2 + X3, and so on. 4. First layer Let us feed all five weighted features into the first layer of the FFN. Multiply by the weights and biases. This time the combining happens *across feature dimensions*, vertically, and each feature grows from 3 numbers to 4. Note that every position goes through the same weight matrix. That is what "position-wise" means. 5. ReLU We cross out the negatives. They become zeros. 6. Second layer Let us bring it back down: 4 dimensions to 3. The output feeds the next block, which has a completely separate set of parameters, and the whole thing runs again. You have just calculated a transformer block by hand. ✍️ The takeaway: the two parts are doing two different jobs, and neither one alone is enough. Attention mixes *across positions*, so a feature can see its neighbours. The FFN mixes *across feature dimensions*, so each position can think about itself. Horizontal, then vertical. Then that pattern repeats N times, each block with its own separate set of weights. That is the Nx from the list up top, and that is what makes the transformer run. 💾 Save this post! #AIbyHand #Transformers #DeepLearningshow more