Sam Altman recently said something that caught my attention: “Don’t learn to code.” Instead, he says, learn high agency, soft skills, and idea generation. I think he’s half right—and half dangerously wrong. Yes, high agency, soft skills, and idea generation matter. They mattered before the LLM era, and they still matter now. No argument there. But telling people to stop learning to code? That is the most dangerous advice in tech right now. ...
Running large language models in production is expensive. A single Llama 3 8B instance in FP16 consumes around 16GB of GPU memory, limiting how many concurrent requests you can serve. Quantization changes this equation dramatically. In this tutorial, I’ll walk you through quantizing Meta’s Llama 3 8B Instruct model to W8A8 (INT8 weights + INT8 activations) using LLMCompressor, then serving it with vLLM for production inference. Table of Contents What is W8A8 Quantization? Why Quantize? Prerequisites Project Setup The Quantization Recipe Quantization Implementation Serving with vLLM Testing Your Deployment Production Considerations Conclusion What is W8A8 Quantization? W8A8 stands for 8-bit Weights and 8-bit Activations. Instead of storing model parameters as 16-bit or 32-bit floating point numbers, we compress them to 8-bit integers. ...
The Observation I recently watched a new graduate join a team. Fresh out of school. Zero real-world experience. Zero familiarity with the codebase or tech stack. That’s typical—every new hire starts there. What wasn’t typical: he went from “where is the repo?” to helping the team with non-trivial tasks in just a couple of weeks. Not months. Weeks. What Changed Twenty years ago, ramping up in a complex codebase with an unfamiliar tech stack took a minimum of a couple of months to become fully productive. That was normal. Nobody expected you to be fully functional in weeks. ...
Your company just announced a new AI initiative. You hired a Head of AI. You’re “exploring use cases.” You have 56 meetings on your calendar about it. And your engineers? They’re laughing at you. Not with you. AT you. The worst part? You have no idea. After two decades in technology and software engineering, I’ve watched this pattern repeat itself across countless organizations. Let me show you exactly why your AI strategy is failing—and what you can do about it. ...
The Popular Comparison A teenager learns to drive in 10 hours, but an AI system needs millions of simulations and millions of hours of simulated data. This comparison appears frequently in AI discussions, but once you look closely, it’s not fair. Here’s why. Where This Example Comes From I recently saw this example in an interview with Ilya Sutskever, co-founder of OpenAI who later started his own AI startup with significant investment backing. ...
Developers are using AI tools like Cursor, Claude Code, GitHub Copilot, and Codex every day. But most are doing it wrong. Here’s what you need to understand: Your job is not to let the AI write code. Your job is to stay in control of the change. The Problem Yes, AI coding tools are really good now. You can literally say “add authentication” or “fix this error,” and it will do something. It will fix the problem. That’s great. ...
The Setup About a month ago, I installed Claude Code on my son’s computer. By the way, I gave him a Linux computer. To be honest, I opened the terminal, installed Claude Code, and now he’s using it. I thought, “Okay, it will be cool. He’ll make some simple games; some HTML, JavaScript, and that’s it.” The 30 Games Surprise In the last 30-40 days, he coded more than 30 web games with HTML and JavaScript. ...
I want to talk about an exciting research paper that has been on my list since its release last month. I finally had the opportunity to dive deep into it, and I believe this represents a fundamental shift in how AI understands code. What Are Code World Models? Let’s start with the basics and some impressive numbers. Meta’s Code World Model is an open weights large language model with 32 billion parameters. It features a dense, decoder-only architecture with a 131K token context size. While these specifications are noteworthy, the truly exciting aspect lies in what this model actually does. ...
When StarRocks introduced vector search, I had to test and see myself. Is it just a toy feature just bolted on an OLAP DB or not? Turns out this thing is actually good enough that you can fold your approximate nearest-neighbor (ANN) workloads into the same cluster that’s already powering your analytics dashboards. In this article, I decided to skip skip the benchmarking details and focus on what you actually need to get vector search running. ...
People say, “I am vibe-coding. Therefore, I am a software engineer.” This is like saying, “I can ride a bike, so I can drive a car.” Both have wheels. Both can move forward, right? Here’s what I’ll say: try merging onto a highway on a bicycle. You’ll learn fast what’s missing. Mapping the Territory Today I want to place vibe-coding, coding, regular coding, and software engineering onto the same map. So you’ll know when to pedal and when to drive. ...