Weekly News #12
This Week’s Updates Worth Your Attention 👀
Hi,
Here is a summary of tech news worth your attention from the week of October 16-22, 2025.
ChatGPT Browser
While many expected OpenAI to challenge Google by changing how people search for information through chat, the company took things a step further by releasing its own browser. With Atlas, you’ll be able to browse websites and ask ChatGPT questions about the page you’re on, instantly. The new Agent feature can even
perform tasks that require navigating websites. Thanks to Atlas, these agents will be able to access your logged-in accounts and carry out more complex actions on your behalf.
Another major addition is memory: ChatGPT will remember your online activity so you can ask things like “What was that hotel I looked at last week?” Of course, this also means giving ChatGPT broader insight into your browsing history, potentially letting it use that data to personalize future answers or even ads. For now, Atlas is only available on macOS.
Claude Code Open Access
Claude Code, which was previously available only through the command line, is now open to everyone through a web app and a VS Code extension. Another piece of news about Anthropic is that the company spent more money on AWS infrastructure than it earned in revenue in 2024. This imbalance shows how the current AI boom may be forming a bubble, as companies spend heavily not only on model training but also on the expensive cloud systems needed for user memory, speed, and data storage.
AWS Global Outage
This week, a major AWS outage showed how much of the internet still relies on Amazon’s cloud. A DNS issue caused major services such as IAM and DynamoDB to stop working in AWS’s oldest and busiest
region, us-east-1. The failure affected banks, hospitals, and security systems. Even after the issue was fixed, recovery was slow because of the backlog of requests that had built up.
DeepSeek OCR
DeepSeek launched a new open-source OCR model that converts text from images with almost perfect accuracy. In response, Andrej Karpathy suggested that large language models might work better if they processed pixels directly instead of text. Some disagreed, saying that language models are designed for words, not images.
Still, this sparked an important discussion. Large language models do not truly understand words; they interpret them as coordinates in a multidimensional space. Feeding them text as images might allow
them to capture more information with less computation, which could help reduce memory and token limitations that models face today.
⚡️ Quick Bits
- Anthropic released a new model designed for the healthcare industry. Anthropic →
- X (formerly Twitter) is selling inactive usernames. TheVerge →
- The AI companion startup Friend faced a protest in New York under the slogan “Make real friends”. Yahoo →
- Samsung launched its Galaxy XR headset to compete with Apple Vision, priced at $1,799. Samsung →
tio
12/10/2025
