Perplexity's new AI-powered Comet web browser, initially exclusive to Max subscribers at $200 monthly, is now accessible to Pro subscribers for $20. I tested it and am impressed—both by Perplexity's hypocrisy and Comet's truly innovative browsing experience. This post shares my rant against Perplexity, while my next post will praise Comet, the first browser on the battlefield in the new browser war.

Source

In my previous Ollama series post, I explained how to access self-hosted AI models in Ollama via HTTP. While this method provides convenient connectivity from different apps, it's only secure for on-premises Ollama hosts. For remote internet access, you can set up an SSH tunnel. Alternatively, you can securely connect from GitHub Copilot to your self-hosted AI models by using Microsoft's VS Code extension Remote - SSH.

Source

In my previous post in the Ollama series with OpenAI OSS models, I explained how to install Ollama on an Ubuntu EC2 instance. If you followed my guide, you now have everything set up to remotely access gpt-oss-20b on your self-hosted Ollama installation via HTTP. I will also showcase how to connect to remote AI models in VS Code using popular coding assistants like Cline and Roo Code, as well as through AI apps such as the local Ollama client, BoltAI, and Anything LLM.

Source

In my previous post, I explained why you might want to run OpenAI gpt-oss-20b or gpt-oss-120b on your own virtual machine and how to select an appropriate EC2 instance type. Today, I will show you how to install Ollama via script or snap on Ubuntu within an EC2 instance. Most steps can be adapted for other cloud providers, on-premises setups, or Linux distributions with minor modifications. I will also briefly discuss the benefits of running open-weight models on an EC2 instance with Ollama compared to Bedrock, SageMaker, vLLM, or Hugging Face Transformers.

Source

OpenAI recently released its first open-weight models, gpt-oss-20b and gpt‑oss-120b, which you can download and run with Ollama. If security and privacy are essential when using AI models, you might not want to use the vendor's chat UI or API. The most powerful AI models require significant computing power and memory, and many desktop devices cannot run these models. Although Ollama was not designed with remote connections in mind, it is still possible. In this series, I will explain how to install OpenAI's open-weight models using Ollama and connect to them remotely from various applications, including VS Code GitHub Copilot. In this first post, I will outline how to choose the right size for your virtual machine, using EC2 instance types in AWS as an example.

Source

The key new feature in the just-released AI Shell Preview 6 is run_command_in_terminal, among several new built-in tools that allow the AI model to run commands and review outputs directly in your PowerShell terminal. Additionally, AI Shell now supports local and remote MCP servers, transforming it into a versatile AI assistant across Windows, Mac, and Linux.

Source

The ChatGPT Gmail connector was first released in June 2025 for Plus, Pro, Team, Enterprise, and Edu users. However, because of alleged data protection rules, users in the European Economic Area, Switzerland, and the UK were excluded. OpenAI is now deploying connectors for Gmail, Google Calendar, Google Drive, and GitHub across Europe for all plans—notably after Anthropic released their MCP connectors weeks ago in Europe without such concerns, suggesting OpenAI's initial data protection claims were merely pretextual or that MCP has no priority status for the AI company. The announcement mentions Google Contacts, but this connector isn't yet accessible in my ChatGPT with a Plus plan. I tested the Gmail Connector, and following GPT-5's disastrous launch, it appears OpenAI released another flop immediately afterward.

Source

The web is flooded with complaints from GPT-5 users, prompting OpenAI's CEO to reintroduce GPT-4o as an alternative. This marks the third setback following the departure of chief scientist Ila Sukestver, who led GPT development until version 4. One likely cause of user frustration is that, following Altman's recent boasts, everyone anticipated a major advance towards "AGI," but we are merely observing another crash into the AI Wall.

Source

In my previous post, I covered Microsoft’s new Edge Copilot Mode, introducing only one significant change. However, shifting Copilot from the sidebar to the center in Edge is more than just a design update. Microsoft’s move marks the first strategic step in the unfolding new browser war involving Google, OpenAI, and Perplexity—companies that exploit broad 'fair use' claims to harvest web content for their own greedy gain. But the upcoming battles aren’t just about who controls the web’s entry point—it could signal the beginning of the end of the web as we know it. The fall of Google’s dominance as the web’s central policymaker now seems inevitable. Once the dust settles from the new browser war, a decentralized web shaped by independently communicating AI agents might emerge. Think that’s far-fetched? Keep reading.

Source

OpenAI has just released its first open-weight models, gpt-oss:20 and gpt-oss:120, enabling you to run these surprisingly powerful models locally, which benefits security and privacy. I downloaded gpt-oss:20 and tested it on my laptop using Ollama and within VS Code with GitHub Copilot. To my surprise, the 20 billion parameter model gpt-oss:20 responded to my prompt.

Source

In Edge Copilot Mode, announced a few days ago, Copilot shifts from being just an AI assistant to becoming the main browser interface. Now front and center in Microsoft's browser, Copilot has evolved into a serious ChatGPT competitor. In this post, I explain how to enable Edge Copilot Mode and discuss whether it introduces any new features compared to the previous version of Copilot in Edge.

Source

Windows-MCP is an open-source MCP server that enables AI agents to control the Windows operating system, supporting tasks such as file navigation, application launching, and GUI automation. It features mouse/keyboard input tools, clipboard management, and window and app state detection. I tested Windows-MCP, one of the extensions listed in Anthropic Claude Desktop that can be installed using the free plan.

Source

💾

💾

💾

Reports indicate that ChatGPT Agent Mode might be capable of solving CAPTCHAs, which are often the final line of defense against persistent bots like AI crawlers on many websites. However, these claims originate from Reddit, known for the frequent hallucinations of its users. To verify, I decided to test it myself.

Source

The new ChatGPT Agent Mode combines the Operator's ability to browse and interact with websites with Deep Research's ability to analyze and synthesize information from multiple sources. I tested ChatGPT's Agent Mode and was quite disappointed. Honestly, I find it surprising that OpenAI made this feature available to the public, considering that they have led the way with their chatbot so far. In my view, ChatGPT's Agent Mode is a flawed concept because it forces an AI agent to use a web browser, a relatively primitive tool designed for inferior human intelligence.

Source

Are you skeptical about AI? Do you believe that AI is overhyped? Then, the latest episode of the Big Technology podcast is for you. AI champion Alex Kantrowitz and fierce AI critic Ed Zitron tackle the gap between AI hype and reality. However, both prominent journalists overlook an essential aspect when assessing AI's economic potential.

Source

I tested three MCP servers in Anthropic Claude Sonnet 4: Claude Gmail Search, Zapier Gmail MCP server, and GongRzhe’s Gmail MCP Server. I also submitted the same search prompt to Gemini in Gmail. The results varied greatly. Only two AI tools passed my test, one failed, and the most promising MCP server turned out to be useless.

Source

Anthropic has introduced two new concepts for connecting AI models to resources through MCP: desktop extensions and connectors. Extensions are available only on Claude Desktop (Windows and Mac), while connectors are supported on both the desktop and the Claude.ai web app. All desktop extensions are local MCP servers, but not all local MCP servers qualify as extensions. Conversely, all remote MCP servers are considered connectors; however, not all connectors are remote MCP servers. Although some theoretical documentation about these new concepts exists, I mainly categorized them based on how Claude Desktop organizes the different types of MCP servers.

Source

VS Code version 1.102 introduced several useful new features in GitHub Copilot. Command allow/deny lists let you specify which commands the Copilot agent can execute without asking for permission. The resubmit feature enables you to edit a previous request and resubmit the prompt. Additionally, the new VS Code version includes various updates for managing MCP servers, including a curated catalog of MCP servers.

Source

A new study by METR got a lot of attention online. It claims that developers often overestimate the productivity boost from AI tools. In a coding test, the researchers found that AI caused a 19% "slowdown" for developers. In my opinion, it's not AI coding that's overhyped, but rather the study's findings.

Source