Author: Sneha Bajaj

I started a Codex task from my laptop, walked into a meeting, and approved a breaking-change decision from my phone before I even sat down. That wasn’t possible a few weeks back. On May 14, 2026, OpenAI launched Codex inside the ChatGPT mobile app on both iOS and Android. All you need to do is just update your ChatGPT, scan a QR code from the OpenAI Codex desktop app, and your phone becomes the control system for any coding agent running. The most important workflow shift this creates isn’t about coding from your phone. It’s the ability to keep long-running…

Most people assumed OpenAI would win the AI race by default. It had the head start, the consumer brand, the Microsoft money, and ChatGPT. Anthropic had a safety pitch and a quieter model named Claude. But somehow that assumption got expensive. In May 2026, Anthropic closed a $65 billion Series H round, putting its post-money valuation at $965 billion. OpenAI, by comparison, was most recently valued at $852 billion. That means Anthropic, a company that didn’t exist six years ago, is now the most valuable pure-play AI company in the world. The question is not just how. It is why…

France just pulled off the biggest AI investment any European country has ever seen, and yet the power grid that is supposed to run it all is years away from being ready. That tension is what this piece is actually about. Over €110 billion in AI and data center commitments have landed in France across two summits. Whether those pledges turn into operational infrastructure is a different question, and I think it’s the more interesting one. Key Takeaways The Numbers Behind France’s AI Investment Surge The February 2025 Paris AI Action Summit was where the AI investment story started. French…

I stumbled upon research that made me realize we’ve been thinking about computer security the wrong way. Not disastrously wrong. Just wrong enough to change everything. Researchers at the University of Toronto built an AI worm. Not a simulation or a theoretical model. An actual, functioning malware that breached 73% of a simulated corporate network within a week, without any human help. But what got me was that it wasn’t locked into exploiting a single vulnerability. It taught itself to hack differently on each machine. It read live vulnerability reports from the internet in real time and figured out how…

AI chips are getting faster, but raw compute is only useful when data can reach those chips quickly enough. That’s where high bandwidth memory comes in. Large language models constantly move model weights, activations and cached data between memory and compute units. When memory cannot keep pace, expensive GPU cores can spend time waiting for data. HBM memory tackles this bottleneck by placing stacked DRAM close to the processor and connecting it through an unusually wide interface. In this post, I’ll break down what is high bandwidth memory, how HBM bandwidth differs from conventional memory, why AI GPUs depend on…

I’ve lost count of how many scrapers I’ve had to rebuild because a client’s site swapped a <div> for a <section>. That’s basically the whole story: traditional web scraping tools pull data by matching fixed HTML patterns, and AI crawling systems read a page the way a person would, figuring out what a “price” or a “product name” is instead of just where it happens to sit in the markup. One survives a redesign. The other doesn’t. I’ve maintained both kinds of pipelines. CSS-selector scrapers that snapped the moment a dev team shifted to a new frontend, and LLM-based extractors…

Smart home security fails less often to some genius hack and more often to something really ordinary. An app nobody bothered to encrypt properly. 20 companion apps were tested for major smart home brands and found 16 had critical cryptographic flaws that can let attackers intercept and alter their traffic. That’s 80% of the apps tested, failing at the most basic layer of protection. This guide covers exactly how these failures get exploited, the warning signs you should be aware of, what to do in the first ten minutes of a breach, and the specific settings that stop most attacks…

There’s a man from California who has said close to 2 million words in the last couple of years without moving his lips. His name is Casey Harrell. He has ALS; he can’t speak in a way people can understand. And he communicates through a set of electrodes sitting in his brain. Over nearly two years at home, he has produced more than 183,000 sentences at an average of 56 words per minute. This is something I want to walk you through. Brain-computer interfaces, or BCIs, are systems that read electrical activity in the brain and turn it into an…

I used to start every single ChatGPT conversation from scratch, explaining how I like my answers formatted. That’s not really true anymore. With ChatGPT memory turned on, the assistant can hang onto useful details about my preferences, my work, my goals, and things I’ve mentioned in past chats – then use all of that to make its answers actually relevant to me. OpenAI pushed this further with a background process it calls “dreaming.” The system quietly reviews information across several of the conversations to keep the picture more current, rather than waiting for us to spell everything out. It’s a…

A researcher bought a domain for five dollars and used it to walk sensitive CRM records out of Salesforce. No malware. No stolen password. Just some text hidden in a web form. That was ForcedLeak, and it’s where I want to start, because it kills the idea that prompt injection is a research toy. Prompt injection is unsolved at the model layer, and you can’t patch your way out of it. You have to design around it. And the model doesn’t know the difference between your instructions and instructions that arrive inside the data it reads, and that’s exactly the…

Somewhere on your company network right now, a valid password is probably being typed by someone who shouldn’t have it. That’s not a plot from a horror story. It’s the default condition of most corporate systems, because the old security model assumed that anyone who got past the front door belonged there. That assumption is what’s failing. The global average cost of a data breach has hit $4.44 million in 2025. And in the U.S it reached a record $10.22 million. And most of those didn’t start with a Hollywood-style hack. They started with a stolen credential walking through the…

In April 2026, an electric aircraft flew from JFK to a Manhattan heliport in just 7 minutes. I have done that trip by car. It took me 71 minutes and a measurable amount of my will to live. So yes, flying taxis are real. No, they’re not affordable, and they won’t be for a while. A seat on one of the first routes runs about $200, which is helicopter money. Getting an eVTOL ride down to Uber prices needs three things, all at once. Pilots out of the cockpit, better batteries, and enough vertiports to keep aircraft flying all day.…