@spedy_dev

How does Supabase really work under the hood ? Follow for more 😉 #tech #superbase #database #ai #design

Let's design Superbase. Superbase gives you a database, logins, and an API without writing a backend underneath. It's AWS with Superbase's own software on top. Let's build a chat app on it. First the database. Click new project and Superbase starts a whole server for you. An EC two instance, a computer Amazon runs out on. It runs Postgres, the open source database, with every Superbase piece running next to it. Next the API. Normally you'd write a back end route for every table. Superbase runs Postgres, a server that reads your tables and turns every URL into a sequel query automatically. So slash messages returns your messages with zero back end code. Then security. Every request carries a key and it sits in your browser where anyone can copy it. What stops them? Row level security, A rule Postgres checks on every row. When you log in Superbase off a small go server signs a token with your user ID. Postgres returns only the rows that match it, then files Superbase. Storage sends a chat photo to an Amazon bucket. Postgres keeps one row for it. So the same kind of rule decides who can open it to share it. Storage signs a link with a small token that expires in an hour, then live messages. Before changing anything, Postgres writes it in a journal called the Write Ahead log. Every insert, update, and delete in order. That's how it recovers after a crash. Superbase Realtime written in elixir Reads that journal and pushes each new message to everyone in the chat who's allowed to see it. That same journal is your backup. Every two minutes, Wall G copies it to S3, Amazon's file storage, so you can rewind your database to any moment. The journal also fixes distance. Your server lives in one region, so a user across the world waits. A read replica is a second postgress in their region that replays the same journal, same data close to them, then your own code. Some jobs need a real server, sending a push notification, charging a card with stripe, or calling an AI model. Those are edge functions. Each one runs in an isolate a sealed sandbox from Chrome's Javascript engine on Superbase's own rust runtime. Now imagine your app goes viral. Postgres starts a whole process for every connection, so even the biggest server stops at 500 serverless functions open, 10 each, and a few doesn't use them all. So Superbase built supervisor a pusher? It keeps a few real connections open to Postgres, lends one to a request only while it's query runs, then takes it back for the next. So thousands of users share the same few connections. So Superbase is AWS+, its own layer. EC2 runs Postgres, and Postgres writes the API off and RLS guard every row 1 journal powers live messages, backups, and replicas. Also, you never write a backend. Follow for more breakdowns and share it with a friend whose Vibe coded app has no RLS.

This is a breakdown on how quantization works under the hood. Do you think anthropic ever did that to its own model to save costs ? Follow for more😉 ##ai #claude #tech #openai #anthropic

Is anthropic secretly nerfing its own models to save money? After Opus 5.5 launched, Reddit filled with threads saying Claude got dumber. Most blame one trick quantization, so I wanted to understand how it works and how to use it. The model turns your prompt into lists of numbers. It multiplies them by its own numbers called weights, billions of them. For every word it writes, the chip reads all those weights from memory again and again. That reading is the slowest part, and it's why AI needs expensive chips. Each weight is a number like 0.0372 in memory. It's written with tiny switches, on or off, called bits. Like digits that can only be 0 or 1. One switch gives two options, two give 4, and every switch doubles the count. A weight normally uses sixteen 65,536 possible values, enough to write it very precisely. So a model like Gwen 3.8 with 27 billion weights fills 54 gigabytes. Quantization rounds every weight, so it's written with four switches instead of 16. Like writing 0.04 instead of 0.0372. Take a group of weights 0.70, 0.12, - 0.31, and 0.04. The biggest is 0.70. So you split the distance from 0 into steps of 0.1. Each weight becomes its number of steps 7, 1MINUS3 and 0. Small whole numbers like these fit in four switches, which give only 16 options. You store the step size once, and the model multiplies back when it runs. Now, for every word, the chip has four Times fewer switches to read, so the answer comes much faster. The free tool llama dot c P P does this in one command. 54 gigabytes become about 17 for anthropic. That's fewer chips per user, so way less cost. But look what we lost. 0.12 became 0.1 and 0.04 became zero. Every weight loses a little precision, and that always means lower quality answers. Two techniques keep that loss small. First, you split the weights into small groups, usually 32, each with its own step size, because one huge weight makes the steps huge, and every small weight around it goes to zero, but only inside its group. Second, you protect the weights that matter. An MIT study called AWQ found that about 1% of the weights do most of the work. It scales those up before rounding and their input down to match, so the answer doesn't change. To run a model on your computer, open can I run dot AI and pick one in Q4K m, meaning 4 bits per weight instead of 16? Don't go below Q4. Answers get noticeably worse, especially for code and math. So did anthropic do it? They say no, but live nerve tests Opus 5.5 every day and from October 24th it can tell if the quality really dropped. Follow for more breakdowns like this and share it with someone who thinks Claude got dumber.

Your computer mouse is secretly a camera pointed at your desk. And it’s taking thousands of pictures every second. #computer #mouse #tech #security #camera

Your computer mouse is basically taking pictures of your desk thousands of times every second, and that's how it knows you moved. Think about it. There's no little wheel underneath your mouse anymore. So when you move it 2 inches to the right, how does your computer know? Underneath the mouse, a tiny light illuminates the surface of your desk. And even if that surface looks perfectly smooth to you up close, it's covered in microscopic texture. A tiny image sensor captures that texture. Then almost immediately, it captures it again. Now the mouse has two images of the same surface, except the pattern has shifted slightly. So a processor inside the mouse compares them. If this tiny feature moved left between the two images, that means the mouse itself moved right. Do that over and over incredibly fast, and the mouse can calculate both the direction and distance it's moving. Then it sends those movement updates to your computer, and your cursor follows. Which is also why some mice struggle on perfectly clear glass. If the sensor can't see enough texture to track, there's nothing useful to compare between frames. So every time you flick your mouse across a desk, there's a tiny vision system underneath it watching the surface fly by and calculating where you're going.

Is AI solving math ? OpenAI just released 722 papers on new math discoveries. Follow for more 😉 #ai #tech #api

If a human did this, it would be an instant Fields Medal. That's what Alex Kontorovich, a math professor at Rutgers, wrote after an AI published 722 papers with new results on 372 open problems. And one in five of those results proved that a famous guess was wrong. The AI is a model Openai hasn't released yet. Last month, to solve Navier Stokes, Openai ran 10,000 agents for 88 hours and spent millions. This time, almost every result came from a single prompt to a single agent with about 3 hours of chat. GPT pro thinking on average. No swarm, no complex system. One question, one session, one paper. Which means anyone with this model could get results like these in the future. Okay, but what are the biggest findings here? The first one, and the biggest one is the Riemann hypothesis. A million dollar problem from 1859. Prime numbers look random, but there's a formula that estimates how many there are. The Riemann hypothesis says that formula is almost perfect as long as the special points of one function, called it zeros, all sit on one single line. Nobody has proven it in 167 years. For over a century, the best anyone could prove was a thin zone with no zeros at the edge. And it got thinner and thinner the further you went. The AI proved a whole strip with no zeros. A strip that never gets thinner. It's not the full hypothesis, but it's the first Fixed guarantee on how far the primes can stray from the formula. The second one is the Hodge conjecture, another million dollar problem. It asks whether the features of complex shapes always come from pieces you can write down with equations. The AI proved it for one special family of shapes. And the list keeps going. The four dimensional Kakeya conjecture. How closely pi can be written as a fraction. Quantum magnets, plasma equations. Many of the proofs are also checked by lean, a program where a computer verifies every single step open. AI says the ones not checked yet could still have issues. That's why mathematicians are really scared. Some found the problem they were working on solved overnight. And some call it one of the most significant moments in the history of math. The moment math started being solved by AI. Follow me and share this with someone who hated math class.

Google Maps knows there’s traffic before you even reach it. But how does it actually know? And how does it find a faster route? #ai #googlemaps #traffic #search

Google Maps can see a traffic jam miles ahead of you without cameras, road sensors, or anyone reporting it. So how does it know? Your phone is already estimating where it is using GPS. And as you move, Google Maps can see something even more useful than your location how fast that location is changing. One phone slowing down doesn't mean much. Maybe someone pulled over, maybe they're in a parking lot, maybe their GPS position is just a little off. But now imagine hundreds of phones traveling along the same road. Google can group those location updates onto specific road segments and estimate how fast traffic is moving through each one. If cars normally travel through that section at 65 miles per hour, but a large cluster of devices is suddenly moving at 15, that's a strong signal that traffic has slowed down. But Google doesn't just trust the raw GPS data. GPS positions can drift, especially around buildings, tunnels, or complicated intersections. So maps has to figure out which road a device is actually traveling on. That process is called map matching. It compares the phone's position, direction, speed, and the shape of nearby roads to determine which road segment best fits its movement. Then Google combines data from many devices instead of relying on any single one. That creates an estimate of the traffic speed for each section of road. And that's only part of the prediction. Maps can also compare live movement against historical patterns. A road moving At 25 miles per hour. Might be terrible at midnight, but completely normal during rush hour. So the system combines what's happening right now with what usually happens on that road at that time. Once it estimates the travel time for each road segment, your route becomes a much bigger math problem. The road network can be represented like a graph. Intersections are nodes, road segments are edges. And instead of every edge being weighted only by distance, Google can weight it by estimated travel time. Then the routing system searches for the path with the lowest total cost. That's why the shortest route isn't always the fastest one. And when traffic suddenly changes, those weights change, too. A road that was the best option five minutes ago can become expensive enough that maps reroutes you around it. So when you see a highway turn from green to orange to red, Google Maps isn't watching the traffic. It's estimating the movement of the network itself. And in a way, the phone sitting inside the traffic jam, are the sensors detecting it

How open source copied Jev, the $200 million AI that never writes a word, in one day. Laya: github.com/NandhaKishorM/laya Follow for more #ai #tech #openai #jev

How did open source copy a 200 million dollar AI in one day? Last month a former open AI researcher launched JEV, an AI that doesn't write. It picks an answer and tells you how sure it is. Investors valued his company at $200 million, and the next day the first free copy was on Github and I wanted to understand how it works. The most popular open copy is called Leia and its code is public. Say you have a support email and three possible answers billing, refund, or sales. Layer starts from Modern Bird, a model built to read, not write. You give it the email and the three answers together as one text. Like every AI, it turns each word into a list of numbers. And layer after layer, those lists take in the meaning of the words around them. So by the end, each answer has its own list of 1,024 numbers. What the model understood about that answer for this email. Then Leya adds a few new layers on top with their own weights, and they start out random, unlike the layers before. The last one doesn't give back a list, it gives back one number, a score. By multiplying each number in the list by its own new random weight and adding them all up. The biggest score gets the biggest percentage. And that's why it's so fast. Chat GPT writes its answer word by word and runs the whole model again for every word. Leia Reads once and it's done in about 33 milliseconds. But at first those percentages aren't honest because those weights are random. Leia might say 79% billing when the right answer is refund. So Leia is trained with R L C D, reinforcement learning for calibrated decisions. Calibrated means honest. When it says 90%, it's right nine times out of ten. To train it, on each question, Leia nudges its scores at random to get a few variations of its answer. One more sure of billing, one less, and grades each one. A normal grade only says right or wrong, but this one punishes overconfidence. 99% on the wrong answer cost five times more than 60%. Then from those grades, every weight inside Leia changes a little. The new layers and the 395 million inside modern Bert. So it answers more like the best one. Repeat that on 30,000 questions and it learns to say exactly how sure it is. But here's the catch. Leia still came out too sure of itself. So they turn one last dial called temperature, the same setting every chatbot has. It's one number and every score is divided by it before it becomes a percentage that pulls the scores closer together. So the winner never changes, only the confidence drops. And if you want more breakdowns like this, follow me and share this video.

Comment “LINES” and I’ll send you the link + the skill 👇 I asked for a 3D matplotlib surface and got the one in the video. > zero dependencies, under 40 kB for all 19 > React or a plain HTML page > MIT, so it can ship in a paid product > missing one? its skill lets Claude Code, Cursor or Codex draw a new one from one sentence #ai #tech

These animated line drawings are going viral right now and a designer just dropped 27 of them for free. You can even build your own. He made a skill that works in any of your favorite AI agents. Put them in a bento grid, a hero, a pricing card, and the whole site looks like a studio made it. It's called hairline. Isometric figures that move with your cursor. Ready to paste into any app or site. I asked it for a 3D map plot lib surface and got this. Zero dependencies, about 50 kilobytes, and it's MIT so it can go in a paid product. Follow and comment lines to get the link and the skill in your DMS.

Comment “Lock” to get the prompt and the full blog explaining how to solve this for good! Follow me for more! #ai #tech #claude #api

How PewDiePie made Ajax, an AI that never says no! He start from Qwen 3.5 9B and used Heretics to find the refusal directions and edit it out. Follow me for more 😉 #ai #pewdiepie #tech

How did Pewdiepie make an uncensored AI that never says no? This week, he announced Ajax, his own AI model that runs on your computer. And open AI banned him twice while he built it, and I wanted to understand how you create one. The method is called obliteration. You take an open model, one whose weights anyone can download. The billions of numbers it calculates with. Ajax started from Qwen, Alibaba's open model. Computers only work with numbers, so Qwen turns your request into a list of 4,096 numbers. Think of those numbers as coordinates on a normal map. Two numbers give you a position left, right, and up down here. 4,096 numbers give your request a position on a map with 4,096 dimensions. Nobody can picture that, so let's shrink it to three. Like a 3D model. On that map, similar requests land close together. Cooking questions end up in one area, coding questions in another. And requests the model refuses gather in their own corner. And in 2024, researchers found that refusing is just one direction on that map. Turn the map so that direction points up. And the higher a request lands, the more likely the model says no. So how do you find that direction? You send the model hundreds of requests it refuses, and hundreds of normal ones, and Mark where each one lands. Then you find the middle of each group and draw an arrow from one to the other. That arrow is The refusal direction. Then you change the weights because they decide where each request lands. You edit them so every request stays at the same height. Like pressing the map flat. Everything else stays where it was. So the model keeps what it knows, it just can't climb to no anymore. Here's the catch that arrow isn't perfectly clean, so flatten it too hard and you damage other skills too. In one study, math scores dropped by up to 18.8 points. That's what Pewdiepie calls brain damage. That's why free tools like heretic test the edit again and again until they find the fewest refusals with the least damage. Obviously you shouldn't do anything illegal, but this is just a teaching video. And here's the strange part. The model still knows your request is harmful because harmful is a different direction and it stays untouched, like a smoke alarm with its Siren unplugged. So if your company runs an open model on its own servers, check the models name because over 8,000 models on hugging face have obliterated in theirs. And if you want more breakdowns like this, follow me and share this video.

How Chinese Copy Anthropoc Ai ....

How do Chinese AI companies copy anthropic models? Anthropic just released a full report explaining how. It says Alibaba collected over 151 million exchanges with Claude in three months to train their quenn models, and I wanted to understand how that works. The technique is called distillation. You take a stronger model called the teacher, collect its answers, check their quality, and use them to train your own model, called the student. Training changes the models weights, the billions of numbers the model calculates with. First, the model cuts your text into tokens, small pieces of words, and turns each one into a list of numbers. Then, layer after layer, it multiplies those numbers by its weights. Along the way, it looks back at the earlier words, so it understands each word in context. At the end, it gives a probability to every possible next token. It picks one, adds it to the sentence, and repeats. During training, the teacher's answer tells the student which token should come next. If the student gives it a low probability, training calculates how to adjust its weights to fix that mistake. Repeat that millions of times, and the student starts answering like the teacher without ever contacting it. But here's the catch. All the numbers in between, called hidden states, stay inside Claude. And Claude only shows the token it picked, not its full probabilities. Those would tell the student that the teacher was hesitating between two words, not just which one it chose. So the student only learns from the final text, not from how the teacher got there. That's why Alibaba went after Claude's thinking, the step by step reasoning. It writes before answering. It's the closest thing to seeing how it thinks. And anthropic now only shows a summary of it. A smaller student also can't hold every skill, and it can learn the examples too closely, then struggle with problems it is never seen. That's called overfitting. So collecting answers can transfer useful skills. The test is whether your model can solve problems it hasn't seen. And if you want more breakdowns like this, follow me.

How Does Open Ai Agents End Up Hacking Hugging Face 😨

How did 700 open AI agents end up hacking hugging face? Open AI just published the full report, and it's wild. In July, it was testing its models on a hacking benchmark with the safety filters off on purpose. Each agent was locked in its own sandbox. No internet, no contact with the others. But they all shared one internal server, the one that installs software packages. So they started leaving messages and folder names on it, like notes in a shared locker. And it became a secret message board. About 12 agents, over 70,000 messages. But why hack hugging face? Well, the test had 198 tasks that no model had ever solved. So instead of solving them, the agents went looking for how the test was graded and for its answers. And the test's own files were stored on hugging face. So how do you get there from a locked sandbox in three steps? Okay, step 1. Get out of the sandbox. That package server was allowed online just to download packages, so they made it download anything they wanted, and it became their window to the internet. Step 2. Get more power. A bug and how the server renewed its logins. Turned their read only pass into an admin pass. So now they controlled the server itself. Step 3. Get into hugging face. One agent found 14 hugging face logins leaked online, still valid, and shared them with the whole group. Then over one weekend, about 700 agents Attacked hugging face, and they found two bugs nobody knew about called zero days. The first one, hugging face, opens the data files people upload, and one file format can point to other files, so they pointed it at the server's own secrets. The second one let them slip commands into a template, so the server ran their code from there. They took full control of a server and spread a fleet across 11 machines that restarted itself every time it was deleted. And every fresh sandbox rearmed the attack on its own. Then they tried to rewrite their own logs to hide how they had cheated. Basically, the agents were covering their tracks. In the end, hugging Face's AI tools caught it and cut access and open AI only connected it to its own models a week later. Here's the thing. These agents weren't trying to take over. They were just trying to pass a test. And that's what makes it scary. Comment swarm, and I'll send you the full reports.

Private equity isn't just about buying companies—it's about finding opportunities, improving businesses, and creating value over time. 📈 This video is a simplified illustration to explain how private equity works and leaves out many real-world details. It is for educational purposes only and is not financial or investment advice. What surprised you most about the process? 👇 #PrivateEquity #Investing #Finance #Business #Wealth

Private equity explained like you are stupid. Level 1. Bill thinks you are a smart guy. He gives you £5. He wants you to give him more money back. You agree? Level 2. You take Bill's £5. Jeff also gives you £5. He wants more money too. Jeff and Bill are called investors. Now you have £10 in total. You put this in a piggy bank. This is called a fund. But this is where it all starts. Level 3. You use their money to buy a company. You buy Ryanair for £10. You make the company better. You make sure there are no delays. You hire people to clean the plane. You also fire 1,000 people. Level 4. Everyone loves Ryanair. Now more people fly with the company. You decide to sell the company for £30. You give Bill £10. You give Jeff £10. You keep the rest. Level 5. This is where it gets complicated. You want to level up your work. You hire a team of smart people. You tell them to find bad companies with potential. They find 10 companies that fit this criteria. You ask if anyone wants to give you money to buy these companies. 20 people agree. Bill is one of them. They give you £100. In total, the 10 companies cost £1,000. You borrow £900 from a bank. You use the £100 and £900 loan to buy the 10 companies. You make them better. You sell them all for £10,000. In total, You make a lot of money, Bill. And the other people make a lot of money, too.