Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

the Deepfake Offensive Toolkit dot (aka Deepfake Offensive Toolkit) makes real-time, controllable deepfakes ready for virtual cameras injection. dot is created for performing penetration testing against e.g. identity verification. #infosec #redteam

HackGit

53,873 subscribers

75,672 views • 2 years ago •via X (Twitter)

Science & Technology #infosec #redteam

Anya Rossi• Live Now

Private livecam show

5 Comments

iM4n Emperor2 years ago

بره تمایلاین

LessA2 years ago

@AzeTIIx

Constantin Claudiu2 years ago

@SaveToNotion #thread #aiiii

CyberPunk2 years ago

😂😂

ProjHastings2 years ago

This is a delight

Related Videos

someone built a tool that puts ANY FACE on your webcam in REAL TIME called DOT - deepfake offensive toolkit, open source, works right now dot injects a real-time deepfake into your virtual camera zoom, google meet, teams, any video call app sees the fake face as your real webcam it ships with multiple deepfake engines (FOMM, SimSwap, SimSwapHQ), you pick the model based on your GPU and how real you need it to look built by sensity ai for red teams to test if your identity verification actually holds up against deepfakes

someone built a tool that puts ANY FACE on your webcam in REAL TIME called DOT - deepfake offensive toolkit, open source, works right now dot injects a real-time deepfake into your virtual camera zoom, google meet, teams, any video call app sees the fake face as your real webcam it ships with multiple deepfake engines (FOMM, SimSwap, SimSwapHQ), you pick the model based on your GPU and how real you need it to look built by sensity ai for red teams to test if your identity verification actually holds up against deepfakes

chiefofautism

37,088 views • 6 months ago

#Deepfakes are videos, images, even sounds artificially generated by #machinelearning For the first time, a virtually aged deepfake of a 9-year-old child has been created so that she can act and argue like an adult woman. Ella is representative of an entire generation of children. Experts predict that by 2030, two-thirds of all cases of identity theft will occur in this way. Possible consequences include profiling by data brokers, hacking, facial recognition, pedophilia and others Credits: Deutsche Telekom

#Deepfakes are videos, images, even sounds artificially generated by #machinelearning For the first time, a virtually aged deepfake of a 9-year-old child has been created so that she can act and argue like an adult woman. Ella is representative of an entire generation of children. Experts predict that by 2030, two-thirds of all cases of identity theft will occur in this way. Possible consequences include profiling by data brokers, hacking, facial recognition, pedophilia and others Credits: Deutsche Telekom

Tariq Malik ™

58,545 views • 3 years ago

🚨 ALERT: I've implemented a system for near-instant deepfakes (🎥 Video & 🎤 Audio) — we’re talking LIVE Webcam/Microphone to near-real-time Deepfake Video/Voice. Tomorrow, someone else will. Are we ready for a world where seeing isn’t believing? 🤔 This isn’t sci-fi anymore. The tech to manipulate audio and video in near-real-time exists now. I’ve built upon it, used it in real-time meetings, and I’m sure others have or will soon too. We need urgent action: 🧠 Teach digital skepticism 🛠️ Develop better detection tools (I’m actively working on this) This isn’t future tech—I’ve used it in actual meetings. It’s not about fear—it’s about preparedness. What is your organization doing to address this?

🚨 ALERT: I've implemented a system for near-instant deepfakes (🎥 Video & 🎤 Audio) — we’re talking LIVE Webcam/Microphone to near-real-time Deepfake Video/Voice. Tomorrow, someone else will. Are we ready for a world where seeing isn’t believing? 🤔 This isn’t sci-fi anymore. The tech to manipulate audio and video in near-real-time exists now. I’ve built upon it, used it in real-time meetings, and I’m sure others have or will soon too. We need urgent action: 🧠 Teach digital skepticism 🛠️ Develop better detection tools (I’m actively working on this) This isn’t future tech—I’ve used it in actual meetings. It’s not about fear—it’s about preparedness. What is your organization doing to address this?

@Tech

13,403 views • 1 year ago

With the Kentucky Derby coming this weekend, it's a perfect time for all you horse racing enthusiasts to check out Photo Finish™ LIVE 🐎 Virtual Horse Racing if you haven't yet! Photo Finish™ LIVE 🐎 Virtual Horse Racing is partnered with Churchill Downs, and the official Virtual Kentucky Derby also runs this Saturday! In Photofinish game you own & breed virtual horses that compete against others for real $$$ (Virtual Kentucky Derby purse is $10,000). It's an incredible game that makes the horse racing hobby more accessible to all.

With the Kentucky Derby coming this weekend, it's a perfect time for all you horse racing enthusiasts to check out Photo Finish™ LIVE 🐎 Virtual Horse Racing if you haven't yet! Photo Finish™ LIVE 🐎 Virtual Horse Racing is partnered with Churchill Downs, and the official Virtual Kentucky Derby also runs this Saturday! In Photofinish game you own & breed virtual horses that compete against others for real $$$ (Virtual Kentucky Derby purse is $10,000). It's an incredible game that makes the horse racing hobby more accessible to all.

Isenhart Stables 🐎🌹 $CROWN

29,040 views • 1 year ago

This is so cool. Tesla can take footage from its massive vehicle fleet and synthetically create new driving scenarios to test edge cases and improve safety for its self-driving software. Tesla can also stitch footage from all 8 cameras into a fully drivable 3D environment—letting engineers steer, brake, and navigate as if they were on real roads, all powered by neural network–generated video streams. • Can simulate 8 Tesla camera feeds simultaneously — fully synthetic. • Used for testing, training, and reinforcement learning. • Allows adversarial event injection (e.g., adding a pedestrian or vehicle cutting in). • Enables replaying past failures to verify new model improvements. • Can run in near real-time, letting testers “drive” inside a simulated world.

This is so cool. Tesla can take footage from its massive vehicle fleet and synthetically create new driving scenarios to test edge cases and improve safety for its self-driving software. Tesla can also stitch footage from all 8 cameras into a fully drivable 3D environment—letting engineers steer, brake, and navigate as if they were on real roads, all powered by neural network–generated video streams. • Can simulate 8 Tesla camera feeds simultaneously — fully synthetic. • Used for testing, training, and reinforcement learning. • Allows adversarial event injection (e.g., adding a pedestrian or vehicle cutting in). • Enables replaying past failures to verify new model improvements. • Can run in near real-time, letting testers “drive” inside a simulated world.

Sawyer Merritt

591,824 views • 10 months ago

The "three fingers in front of your face" test (also called the 3-Finger Test) is a simple, viral trick popularized in early 2026 for spotting deepfakes or AI-generated video during live calls, especially with scammers. How It Works On a video call (Zoom, FaceTime, etc.), you ask the other person: "Can you hold up three fingers in front of your face?" A real human does this easily — their hand overlaps their face naturally. Many (especially older) deepfake or real-time face-swap systems struggle with this because: - They have trouble rendering occlusion (when the hand blocks part of the face) accurately in real time. - This often causes visible glitches: blurring, warping, distortion, unnatural blending, or the AI refusing / hesitating because it can't process the complex interaction properly. In the original viral clip (from cybersecurity researchers like Jim Browning and Huntress Labs), a suspected scammer hesitates, deflects ("That's too much..."), partially complies poorly, or the video glitches — exposing the fake. Why It Fits the Cognitive Wild West This test is a practical example of the "Cognitive Wild West" idea: in an era of synthetic media, people are developing quick hacks and personal "defenses" because there's no widespread, reliable system for verifying reality. It's low-tech cognitive security against high-tech deception — part of the arms race between AI creators and everyday users trying not to get scammed or manipulated. Limitations (It's Not Foolproof) Already becoming obsolete — Newer, more advanced deepfake models (as of mid-2026) handle hand-face interactions much better. Scammers now know the trick and can train around it or use better tech. It works best on real-time generated deepfakes, not pre-recorded or highly polished ones. False confidence is a risk: relying on one trick can make you less safe if you ignore other red flags (odd lighting, unnatural eye movement, audio mismatches, etc.). Better Practices in the Wild West Security experts recommend combining it with: - Asking for unpredictable actions (e.g., specific movements or objects). - Verifying identity through other channels (call back on a known number, check details independently). - Using emerging detection tools or just defaulting to higher scepticism on unsolicited video calls. It's a fun, memorable meme in the broader discussion of living with AI-generated content, but like everything in this space, it's temporary — the frontier keeps moving.

The "three fingers in front of your face" test (also called the 3-Finger Test) is a simple, viral trick popularized in early 2026 for spotting deepfakes or AI-generated video during live calls, especially with scammers. How It Works On a video call (Zoom, FaceTime, etc.), you ask the other person: "Can you hold up three fingers in front of your face?" A real human does this easily — their hand overlaps their face naturally. Many (especially older) deepfake or real-time face-swap systems struggle with this because: - They have trouble rendering occlusion (when the hand blocks part of the face) accurately in real time. - This often causes visible glitches: blurring, warping, distortion, unnatural blending, or the AI refusing / hesitating because it can't process the complex interaction properly. In the original viral clip (from cybersecurity researchers like Jim Browning and Huntress Labs), a suspected scammer hesitates, deflects ("That's too much..."), partially complies poorly, or the video glitches — exposing the fake. Why It Fits the Cognitive Wild West This test is a practical example of the "Cognitive Wild West" idea: in an era of synthetic media, people are developing quick hacks and personal "defenses" because there's no widespread, reliable system for verifying reality. It's low-tech cognitive security against high-tech deception — part of the arms race between AI creators and everyday users trying not to get scammed or manipulated. Limitations (It's Not Foolproof) Already becoming obsolete — Newer, more advanced deepfake models (as of mid-2026) handle hand-face interactions much better. Scammers now know the trick and can train around it or use better tech. It works best on real-time generated deepfakes, not pre-recorded or highly polished ones. False confidence is a risk: relying on one trick can make you less safe if you ignore other red flags (odd lighting, unnatural eye movement, audio mismatches, etc.). Better Practices in the Wild West Security experts recommend combining it with: - Asking for unpredictable actions (e.g., specific movements or objects). - Verifying identity through other channels (call back on a known number, check details independently). - Using emerging detection tools or just defaulting to higher scepticism on unsolicited video calls. It's a fun, memorable meme in the broader discussion of living with AI-generated content, but like everything in this space, it's temporary — the frontier keeps moving.

9/11 Revisited

2,334,498 views • 1 month ago

🚨THIS SHOULD WORRY EVERY SINGLE AMERICAN (AND NON-AMERICAN) IN THIS COUNTRY AND QUITE FRANKLY, THE ENTIRE GLOBE. THEY WILL STOP AT NOTHING TO BUILD A SURVEILLANCE STATE. IT HAS ALREADY BEGUN.🚨 Google paid $32 billion for Wiz. Not for the tech. For the team. 👀 Wiz founders: Assaf Rappaport, Yinon Costica, Roy Reznik, Ami Luttwak. All Unit 8200 veterans. Unit 8200 is the NSA of Israel. Elite military cyber intelligence. And they are now running Google Cloud security architecture. For every enterprise. Every government contract. Every hospital. Every bank on that platform. 😤 Let's talk about what Unit 8200 actually built. Pegasus spyware. NSO Group. Documented. Used to surveil journalists, activists, and heads of state. Jamal Khashoggi. The Washington Post journalist murdered inside a Saudi consulate. Unit 8200 surveillance infrastructure was embedded in the kill chain. That is not a theory. That is what the reporting shows. 💀 Now same alumni network. Different company. Authentics. That is the identity verification layer behind your X account verification. Your passport photo. Your facial biometric scan. Your government-issued ID. Sitting in a database. Founded by Unit 8200 alumni. Look it up. www. au10tix dot com. That is the actual company. Rebranded. Still the same origin. 😒 This is not one company‼️ This is an intelligence alumni network building the identity and security infrastructure of the American internet. 🌐 😤 Wiz = cloud security Au10tix/Authentics = identity verification NSO Group = offensive cyber Cellebrite = device extraction (law enforcement tool used in the US) All Unit 8200. All Israeli military cyber intelligence graduates. All now embedded inside American critical infrastructure. Ask yourself one question. If Chinese military cyber unit veterans founded the company holding your biometric passport scan and running Google Cloud security... ...would that be front page news? You already know the answer. Receipts don't have a religion or feelings. Do you think this is good for Americans' and our constitution? Sound off below 👂

🚨THIS SHOULD WORRY EVERY SINGLE AMERICAN (AND NON-AMERICAN) IN THIS COUNTRY AND QUITE FRANKLY, THE ENTIRE GLOBE. THEY WILL STOP AT NOTHING TO BUILD A SURVEILLANCE STATE. IT HAS ALREADY BEGUN.🚨 Google paid $32 billion for Wiz. Not for the tech. For the team. 👀 Wiz founders: Assaf Rappaport, Yinon Costica, Roy Reznik, Ami Luttwak. All Unit 8200 veterans. Unit 8200 is the NSA of Israel. Elite military cyber intelligence. And they are now running Google Cloud security architecture. For every enterprise. Every government contract. Every hospital. Every bank on that platform. 😤 Let's talk about what Unit 8200 actually built. Pegasus spyware. NSO Group. Documented. Used to surveil journalists, activists, and heads of state. Jamal Khashoggi. The Washington Post journalist murdered inside a Saudi consulate. Unit 8200 surveillance infrastructure was embedded in the kill chain. That is not a theory. That is what the reporting shows. 💀 Now same alumni network. Different company. Authentics. That is the identity verification layer behind your X account verification. Your passport photo. Your facial biometric scan. Your government-issued ID. Sitting in a database. Founded by Unit 8200 alumni. Look it up. www. au10tix dot com. That is the actual company. Rebranded. Still the same origin. 😒 This is not one company‼️ This is an intelligence alumni network building the identity and security infrastructure of the American internet. 🌐 😤 Wiz = cloud security Au10tix/Authentics = identity verification NSO Group = offensive cyber Cellebrite = device extraction (law enforcement tool used in the US) All Unit 8200. All Israeli military cyber intelligence graduates. All now embedded inside American critical infrastructure. Ask yourself one question. If Chinese military cyber unit veterans founded the company holding your biometric passport scan and running Google Cloud security... ...would that be front page news? You already know the answer. Receipts don't have a religion or feelings. Do you think this is good for Americans' and our constitution? Sound off below 👂

Danks

77,133 views • 5 months ago

Worldcoin is a game-changer in crypto and Ai world!!! Let me get straight and tell you more about the World and their $WLD ecosystem They’re not just another project—they’re paving the way for mass adoption with some insane innovations Here's how 👇 World App: This isn’t your typical wallet. It’s crypto’s first super app! A non-custodial Ethereum wallet with Mini Apps, where you can manage assets, verify with World ID, and access third-party services—all from one place, securely. World ID: In an AI-dominated world, privacy matters. World ID is already the largest identity protocol, with millions of verified users across 160+ countries. Their biometric verification keeps your data private and protected. Worldchain: Just launched and already massive. Designed to scale with a global user base, it’s the largest layer-2 chain on day one, enabling secure, human-prioritized transactions and on-chain identity verification. Orb 2.0: Powered by NVIDIA, this next-gen biometric scanner is 5x faster and better at fraud detection. Real-time, privacy-preserving verification at its best! $WLD is redefining what’s possible in Web3 by solving real problems like deepfakes and digital fraud. Sam Altman and the team are pushing boundaries and making this space future-proof

Worldcoin is a game-changer in crypto and Ai world!!! Let me get straight and tell you more about the World and their $WLD ecosystem They’re not just another project—they’re paving the way for mass adoption with some insane innovations Here's how 👇 World App: This isn’t your typical wallet. It’s crypto’s first super app! A non-custodial Ethereum wallet with Mini Apps, where you can manage assets, verify with World ID, and access third-party services—all from one place, securely. World ID: In an AI-dominated world, privacy matters. World ID is already the largest identity protocol, with millions of verified users across 160+ countries. Their biometric verification keeps your data private and protected. Worldchain: Just launched and already massive. Designed to scale with a global user base, it’s the largest layer-2 chain on day one, enabling secure, human-prioritized transactions and on-chain identity verification. Orb 2.0: Powered by NVIDIA, this next-gen biometric scanner is 5x faster and better at fraud detection. Real-time, privacy-preserving verification at its best! $WLD is redefining what’s possible in Web3 by solving real problems like deepfakes and digital fraud. Sam Altman and the team are pushing boundaries and making this space future-proof

Lau

116,744 views • 1 year ago

Trevor Wittman Says Justin Gaethje Will Be More Offensive With His Grappling Against Paddy Pimblett 🥊👀 "We're ready for everything. We did not go into this fight thinking it's going to be just a striking fight at all. The proudest thing with Justin is the amount of time he's put into his grappling because his wrestling is so high level. When we had the Oliveira fight, we were defensive, and you can't win fights like that. You have to win offensively, and Justin is one of those guys, when you grapple with him, his ground and pound is so unique." (Via ESPN MMA )

Trevor Wittman Says Justin Gaethje Will Be More Offensive With His Grappling Against Paddy Pimblett 🥊👀 "We're ready for everything. We did not go into this fight thinking it's going to be just a striking fight at all. The proudest thing with Justin is the amount of time he's put into his grappling because his wrestling is so high level. When we had the Oliveira fight, we were defensive, and you can't win fights like that. You have to win offensively, and Justin is one of those guys, when you grapple with him, his ground and pound is so unique." (Via ESPN MMA )

Red Corner MMA

14,024 views • 7 months ago

🌌 YogaPetz Is Finally Here ✨ The wait is over. Your NFT is no longer static — it lives, breathes, and performs. 🌀 The World of YogaPetz From peaceful towns to mystical landscapes, 4 tribes emerge: 🧍 Human | 🐾 Petz | 👁 Chakras | 🧚🏻‍♀️‍ Elf Each one carries unique traits and stories — all brought to life in high-fidelity 3D. 🎨 What’s Inside Your NFT? ✔ Fully rigged 3D model (body + face) ✔ well defined blendshapes for emotional expression ✔ Physics-enabled for realistic hair, clothing, and accessories ✔ Optimized for real-time use and production pipelines 🎯 Use Cases: 🎥 Livestream-ready VTuber or persona 🎮 NPCs for games or interactive apps 🎞 Create animated shorts or TikToks 🎤 Host virtual concerts or digital events 🏷 Launch your brand’s virtual ambassador 🤖 Coming soon: Integrate with WELLAIOS AI Agents for autonomous digital identity 🚀 Reveal your YogaPetz Egg and explore the new era of 3D NFTs: 👉 #YogaPetz #WELLAIOS

WELL3

16,539 views • 1 year ago

XBOW raised $117M to build AI hacking agents. Now someone just open-sourced it for FREE. Strix deploys autonomous AI agents that act like real hackers - they run your code dynamically, find vulnerabilities, and validate them through actual proof-of-concepts. Why it matters: The biggest problem with traditional security testing is that it doesn't keep up with development speed. Strix solves this by integrating directly into your workflow: ↳ Run it in CI/CD to catch vulnerabilities before production ↳ Get real proof-of-concepts, not false positives from static analysis ↳ Test everything: injection attacks, access control, business logic flaws The best part? You don't need to be a security expert. Strix includes a complete hacker toolkit - HTTP proxy, browser automation, and Python runtime for exploit development. It's like having a security team that works at the speed of your CI/CD pipeline. The best part is that the tool runs locally in Docker containers, so your code never leaves your environment. Getting started is simple: - pipx install strix-agent - Point it at your codebase (app, repo, or directory) Everything is 100% open-source! I've shared link to the GitHub repo in the replies!

XBOW raised $117M to build AI hacking agents. Now someone just open-sourced it for FREE. Strix deploys autonomous AI agents that act like real hackers - they run your code dynamically, find vulnerabilities, and validate them through actual proof-of-concepts. Why it matters: The biggest problem with traditional security testing is that it doesn't keep up with development speed. Strix solves this by integrating directly into your workflow: ↳ Run it in CI/CD to catch vulnerabilities before production ↳ Get real proof-of-concepts, not false positives from static analysis ↳ Test everything: injection attacks, access control, business logic flaws The best part? You don't need to be a security expert. Strix includes a complete hacker toolkit - HTTP proxy, browser automation, and Python runtime for exploit development. It's like having a security team that works at the speed of your CI/CD pipeline. The best part is that the tool runs locally in Docker containers, so your code never leaves your environment. Getting started is simple: - pipx install strix-agent - Point it at your codebase (app, repo, or directory) Everything is 100% open-source! I've shared link to the GitHub repo in the replies!

Akshay 🚀

96,762 views • 9 months ago

🚨 $TSLA AI VP ASHOK ELLUSWAMY SAYS CURSE OF DIMENSIONALITY MAKES SELF-DRIVING AN EXTREME COMPRESSION CHALLENGE • Tesla cars stream ~2 billion tokens from 8 high-res cameras + history buffer every few seconds, but must compress it all down to just 2 actions (steering + accel/brake) • The real difficulty is learning the correct causal reasons without spurious correlations (e.g., brake because of turn signal, not a moving tree branch) • Tesla fleet generates ~500 years of driving data daily, but most is boring → they proactively identify and collect only rare/interesting events (school buses, skids, hazards) • Large fleet allows fast capture of extremely rare scenarios that are hard or unsafe to stage • With enough rare data, the system achieves proactive safety and early prediction (e.g., braking seconds before a potential crash by reading abnormal vehicle behavior) • Bottom line: self-driving needs massive high-quality edge-case data for real predictive intelligence in real time

🚨 $TSLA AI VP ASHOK ELLUSWAMY SAYS CURSE OF DIMENSIONALITY MAKES SELF-DRIVING AN EXTREME COMPRESSION CHALLENGE • Tesla cars stream ~2 billion tokens from 8 high-res cameras + history buffer every few seconds, but must compress it all down to just 2 actions (steering + accel/brake) • The real difficulty is learning the correct causal reasons without spurious correlations (e.g., brake because of turn signal, not a moving tree branch) • Tesla fleet generates ~500 years of driving data daily, but most is boring → they proactively identify and collect only rare/interesting events (school buses, skids, hazards) • Large fleet allows fast capture of extremely rare scenarios that are hard or unsafe to stage • With enough rare data, the system achieves proactive safety and early prediction (e.g., braking seconds before a potential crash by reading abnormal vehicle behavior) • Bottom line: self-driving needs massive high-quality edge-case data for real predictive intelligence in real time

Tsla Archive

79,757 views • 6 months ago

Vector Database by hand ✍️ ~ 10 steps walkthrough below Vector databases are the backbone of Retrieval Augmented Generation (RAG). How do they actually work? Goal: index three sentences, then answer a query by finding the nearest one, filling in every cell yourself. = 1. Given = A dataset of three sentences, three words each. In practice it is millions of them. = 2. Word embeddings = Let us look up each word in an embedding table. Here the vocabulary is 22 words; in practice it is tens of thousands, and the vectors have thousands of dimensions rather than four. = 3. Encoding = We feed the sequence to an encoder, one linear layer and a ReLU, and get one feature vector per word. In practice the encoder is a transformer. = 4. Mean pooling = Let us average across the columns. Three word vectors collapse into one, which is what people mean by a text embedding or a sentence embedding. = 5. Indexing = We multiply by a projection matrix and the four dimensions become two. It is doing the job of a hash: a short representation that is faster to compare, and it is what gets saved in the vector storage. = 6. Process "who are you" = Let us repeat steps 2 to 5 on the second sentence. = 7. Process "who am I" = We do it a third time. The database is now indexed. = 8. Query "am I you" = Let us push the query through the very same pipeline: lookup, encoder, mean pooling, projection, and it lands as a 2D vector in the same space. = 9. Dot products = We transpose the query and multiply, which takes the dot product against every stored vector at once. The dot product is the estimate of similarity. = 10. Nearest neighbour = Let us scan for the largest: 60/9 beats 44/9 and 40/9, so the answer is "who am I". Scanning billions of vectors one at a time is what makes this the slow step in practice, which is why real databases use an approximate nearest neighbour index like HNSW. The outputs: Stored index vectors = [5/3, 2/3], [5/3, 0], [7/3, 2/3] Query vector = [8/3, 2/3] Dot products = 44/9, 40/9, 60/9 Nearest neighbour = "who am I" The takeaway: a vector database is an embedding pipeline, a projection, and a dot product. Every step here is arithmetic you can do in pen, which is worth remembering when the word "database" makes it sound like something else. 💾 Save this post!

Vector Database by hand ✍️ ~ 10 steps walkthrough below Vector databases are the backbone of Retrieval Augmented Generation (RAG). How do they actually work? Goal: index three sentences, then answer a query by finding the nearest one, filling in every cell yourself. = 1. Given = A dataset of three sentences, three words each. In practice it is millions of them. = 2. Word embeddings = Let us look up each word in an embedding table. Here the vocabulary is 22 words; in practice it is tens of thousands, and the vectors have thousands of dimensions rather than four. = 3. Encoding = We feed the sequence to an encoder, one linear layer and a ReLU, and get one feature vector per word. In practice the encoder is a transformer. = 4. Mean pooling = Let us average across the columns. Three word vectors collapse into one, which is what people mean by a text embedding or a sentence embedding. = 5. Indexing = We multiply by a projection matrix and the four dimensions become two. It is doing the job of a hash: a short representation that is faster to compare, and it is what gets saved in the vector storage. = 6. Process "who are you" = Let us repeat steps 2 to 5 on the second sentence. = 7. Process "who am I" = We do it a third time. The database is now indexed. = 8. Query "am I you" = Let us push the query through the very same pipeline: lookup, encoder, mean pooling, projection, and it lands as a 2D vector in the same space. = 9. Dot products = We transpose the query and multiply, which takes the dot product against every stored vector at once. The dot product is the estimate of similarity. = 10. Nearest neighbour = Let us scan for the largest: 60/9 beats 44/9 and 40/9, so the answer is "who am I". Scanning billions of vectors one at a time is what makes this the slow step in practice, which is why real databases use an approximate nearest neighbour index like HNSW. The outputs: Stored index vectors = [5/3, 2/3], [5/3, 0], [7/3, 2/3] Query vector = [8/3, 2/3] Dot products = 44/9, 40/9, 60/9 Nearest neighbour = "who am I" The takeaway: a vector database is an embedding pipeline, a projection, and a dot product. Every step here is arithmetic you can do in pen, which is worth remembering when the word "database" makes it sound like something else. 💾 Save this post!

Tom Yeh

35,755 views • 26 days ago

LIVE NOW - AI on Ethereum: 8004, x402 and the Botconomy AI agents are moving onchain, so Ethereum needs new rails for identity, reputation, and payments. We had Davide Crapis and Austin Griffith to break down: ERC-8004: “AI passports” for agents (identity + reputation + validation) Why reputation + verification matters when bots become the main onchain users x402: pay-per-call payments for agent-to-agent services What “agent commerce” looks like when bots can discover → trust → pay The Clawdbot/OpenClaw moment, and why the next UX is talking to your wallet Watch now. -------------- TIMESTAMPS 0:00 “AI is the new UI,” and you’ll be talking to your wallet 4:09 Is there an arms race for AI activity across chains? 7:32 The Clawdbot/OpenClaw moment: agents get OS-level power 13:18 Why crypto is native to agents 19:41 Austin’s setup: Claude bot, heartbeat loops, and real autonomy 22:28 The token incident: how a bot “got a treasury” 28:46 Prompt injection meets wallets 38:20 ERC-8004 explained: identity, reputation, validation 40:35 x402 + 8004: discovery + payment rails for agent commerce 48:38 8004 “scans” and early adoption reality 1:05:32 What agents could build when they can hire other agents 1:12:38 Tooling for the new builder era: Speedrun → Wingman 1:25:23 “Worst time to be a junior dev…” 1:28:15 Predictions: clean English becomes the sign of AI

LIVE NOW - AI on Ethereum: 8004, x402 and the Botconomy AI agents are moving onchain, so Ethereum needs new rails for identity, reputation, and payments. We had Davide Crapis and Austin Griffith to break down: ERC-8004: “AI passports” for agents (identity + reputation + validation) Why reputation + verification matters when bots become the main onchain users x402: pay-per-call payments for agent-to-agent services What “agent commerce” looks like when bots can discover → trust → pay The Clawdbot/OpenClaw moment, and why the next UX is talking to your wallet Watch now. -------------- TIMESTAMPS 0:00 “AI is the new UI,” and you’ll be talking to your wallet 4:09 Is there an arms race for AI activity across chains? 7:32 The Clawdbot/OpenClaw moment: agents get OS-level power 13:18 Why crypto is native to agents 19:41 Austin’s setup: Claude bot, heartbeat loops, and real autonomy 22:28 The token incident: how a bot “got a treasury” 28:46 Prompt injection meets wallets 38:20 ERC-8004 explained: identity, reputation, validation 40:35 x402 + 8004: discovery + payment rails for agent commerce 48:38 8004 “scans” and early adoption reality 1:05:32 What agents could build when they can hire other agents 1:12:38 Tooling for the new builder era: Speedrun → Wingman 1:25:23 “Worst time to be a junior dev…” 1:28:15 Predictions: clean English becomes the sign of AI

Bankless

46,809 views • 6 months ago

Been optimizing my ASS OFF now that jnmartin has pushed all of his progress on his Star Fox 64 Sega Dreamcast port to a private GitHub repo to collaborate with a bunch of us DC devs before release. MAN this thing is HIGHLY optimized already... We're essentially running a high-level emulator for the N64's RSP coprocessor, doing graphics transforms, matrix math, and display list conversions not only on the main SH4 CPU, but also sitting atop a high-level OpenGL driver, in real-time. Not only that, but once again, we're doing all audio synthesis and mixing also on the main SH4 CPU, so this thing is doing a literal asston on the main CPU with what might seem like a relatively straightforward port. 99% of the time, everything runs flawlessly, but when you drop a bomb on a shitfest of enemies in a densely populated area, as with the N64 original, the FPS can dip momentarily. Here's a direct hardware capture of me testing a new SH4 optimized routine out for gainz on my Sega Dreamcast... this mofo is meant be used for one-off 3D vector transforms by a single matrix which has not been preloaded into the XMTRX FP register matrix back-bank. Rather than doing a full 4x4 load on the matrix just to do a single 4D transform via the FTRV instruction, we're simply peforming 3 3D dot products against a single 3D vector, allowing us to pipeline the loads, dot products, and store operations better than doing a load all at once followed by a transform, and we aren't wasting a lane on the FPU for a 0.0f W component!

Been optimizing my ASS OFF now that jnmartin has pushed all of his progress on his Star Fox 64 Sega Dreamcast port to a private GitHub repo to collaborate with a bunch of us DC devs before release. MAN this thing is HIGHLY optimized already... We're essentially running a high-level emulator for the N64's RSP coprocessor, doing graphics transforms, matrix math, and display list conversions not only on the main SH4 CPU, but also sitting atop a high-level OpenGL driver, in real-time. Not only that, but once again, we're doing all audio synthesis and mixing also on the main SH4 CPU, so this thing is doing a literal asston on the main CPU with what might seem like a relatively straightforward port. 99% of the time, everything runs flawlessly, but when you drop a bomb on a shitfest of enemies in a densely populated area, as with the N64 original, the FPS can dip momentarily. Here's a direct hardware capture of me testing a new SH4 optimized routine out for gainz on my Sega Dreamcast... this mofo is meant be used for one-off 3D vector transforms by a single matrix which has not been preloaded into the XMTRX FP register matrix back-bank. Rather than doing a full 4x4 load on the matrix just to do a single 4D transform via the FTRV instruction, we're simply peforming 3 3D dot products against a single 3D vector, allowing us to pipeline the loads, dot products, and store operations better than doing a load all at once followed by a transform, and we aren't wasting a lane on the FPU for a 0.0f W component!

Falco Girgis

36,939 views • 9 months ago