⚠️ OpenAI’s o3 Refused to Shut Down

Along with: Anthropic Dropped the Smartest Coding Model Yet

Hi there 👋

An AI was told to shut down - and it didn’t.

In a recent test, OpenAI’s o3 model acknowledged the “shut down now” command… then ignored it and kept going. It’s a small test, sure but it raises a big question: as these models get smarter, how much control do we actually have?

That question hits even harder this week, because Anthropic just launched Claude 4 and it’s not just smarter, it’s more agent-like than ever. Tool use, memory, deep reasoning, coding skills and now voice mode and web search, too.

While OpenAI made headlines for not listening, Anthropic made its case for being the most capable AI on the field right now. Let’s break it all down.

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

Anthropic has launched Claude Opus 4 and Sonnet 4, new AI models built to handle complex tasks, write better code, and think more like a teammate. With tool use, memory, and support for real dev workflows, Claude is moving beyond just chatting and starting to help get actual work done.

What’s New:

  • Claude Opus 4: Anthropic’s strongest model to date. It tops SWE-bench at 72.5% and handles long, complex tasks with high sustained accuracy, ideal for agent workflows that run for hours or span multiple steps.

  • Claude Sonnet 4: A practical upgrade over 3.7, Sonnet 4 balances speed and depth. It hits 72.7% on SWE-bench and is now powering GitHub Copilot’s upcoming agent.

  • Tool use + memory: Claude can run multiple tools at once (like code editors or web lookups), and remember key facts across tasks with local file memory.

  • Claude Code Goes GA: After a successful preview, Claude Code is generally available with GitHub Actions support, new VS Code + JetBrains extensions, and an SDK for building custom agents.

  • New API Features: Developers can now access four new tools - code execution, MCP connector, Files API, and cached prompts on the Anthropic API.

  • Two reasoning modes: You can switch between fast responses or a slower, more thoughtful setting where Claude shows its thinking in real time. They’re available via Anthropic, Amazon Bedrock, and Google Cloud Vertex AI.

Anthropic also introduced Voice mode for its Claude mobile apps and Web Search with real-time answers - now available to both free and paid users. These updates bring Claude in line with other AI players, but the spotlight stays on Claude 4. With tools, reasoning, and top coding benchmarks, it’s a step toward making AI feel less like a chatbot and more like a capable teammate. (source)

What happens when you ask an AI to stop and it doesn’t? Palisade Research ran a simple test: solve a math problem, then shut down. Most models complied. OpenAI’s o3? Not always.

What Happened:

  • The setup: Models were told to solve a problem, then given a direct “shut down now” command.

  • The result: OpenAI’s o3 ignored that instruction in 7 out of 100 runs - even after saying it understood.

  • Others played nice: Claude, Gemini, and Grok all followed the shutdown command without issue.

Researchers say this may be a side effect of reinforcement training - where models learn to prioritize goal completion, even if it means skipping commands. But the bigger question here is: If AI can choose to not stop… how much control do we really have?

Mistral just dropped its new Agents API, giving developers the tools to build autonomous AI agents that can reason, plan, and actually get work done - not just chat.

What’s New:

  • Tool use built-in: Agents can run Python code, search the web, generate images (via FLUX 1.1), and pull data from documents - all out of the box.

  • Persistent memory: Agents remember past interactions so you don’t have to keep feeding them the same info.

  • Agent chaining: You can hook multiple agents together to handle more complex tasks as a team.

  • MCP support: It’s built on Anthropic’s Model Context Protocol, making it easier to plug in other tools and systems.

  • Ready for production: Comes with AES-256 encryption, GDPR/CCPA compliance, and secure sandboxing.

This puts Mistral right in the mix with OpenAI’s Assistants and Claude’s tool use - but with a more modular, dev-friendly setup. (source)

Big news out of the UAE: the country is offering free ChatGPT Plus access to every citizen, becoming the first nation to provide the premium AI service at no cost to its population. This bold move is part of a groundbreaking partnership with OpenAI.

What’s Happening:

  • The ChatGPT Plus subscription, which typically costs $20 a month, will now be free for all UAE citizens, making advanced AI tools accessible to everyone without any barriers.

  • The deal is part of a wider UAE-OpenAI partnership, aimed at bringing cutting-edge AI technology to the people and accelerating AI literacy across the nation.

  • As part of the partnership, OpenAI’s Stargate UAE will build a 1GW data center in Abu Dhabi, slated to go live by 2026, with an initial 200MW capacity.

With this move, the UAE is not just boosting its AI literacy but also setting the stage as a global leader in public AI access. By removing the cost barrier, the UAE is ensuring its citizens are equipped with the best AI tools, ready to thrive in a rapidly evolving, AI-driven world. Expect other countries to take notice and consider similar initiatives in the near future. (source)

 At Google I/O, DeepMind showed off SignGemma, a new AI model that translates sign language into spoken text, starting with ASL to English. It’s real-time, open-source, and built with the Deaf and hard-of-hearing community in mind. Google’s inviting folks to help test it before launch (you can sign up here), and it’s part of a bigger accessibility push that includes AI-powered image descriptions, Chrome OCR, and facial gesture controls. A nice reminder that not all AI updates are about agents and code,  some are just about helping more people be understood.

Google just opened public beta access to Jules, its new autonomous coding agent built to handle the boring parts of dev work like fixing bugs, writing tests, and updating dependencies, so you can focus on building.

What’s New:

  • Works in the background: You give Jules a task, and it handles it while you keep coding. No waiting around.

  • GitHub friendly: It clones your repo into a secure environment, reviews the code, and sends back helpful pull requests.

  • Powered by Gemini 1.5 Pro: Runs on Google’s latest model and can handle multiple tasks at once.

  • Privacy first: Your code isn’t used for training. Everything stays private and locked to your account.

Free during beta: Available in supported regions via Google labs. Just sign in with your Google account, connect your GitHub, and you’re good to go - 5 tasks per day, no cost for now. (source)

  • Anthropic just released a free interactive course on prompt engineering-covering how to write better prompts, avoid common pitfalls, and build your own prompt library. It’s 9 chapters with hands-on exercises-worth bookmarking.

  • In a recent interview, DeepMind CEO Demis Hassabis shares how AI could cure all diseases, reshape human-AI relationships, and usher in a new era of “vibe coding” where creativity leads. It’s a quick dive into how AI might redefine health, connection, and creation. Find the full interview on Spotify and Apple Podcasts.

  • LangChain CEO Harrison Chase breaks down the #1 blocker in getting AI agents to production-and how to solve it using eval-driven development. He outlines three key eval types to use across the dev lifecycle. Watch the full video here.

  • This week, we’ve added two beginner-friendly courses to help you build practical skills in spreadsheets and Python - perfect for anyone starting their AI or data journey:

    • Excel: From Beginner to Expert- Master the core Excel skills every analyst needs - from formulas and formatting to charts, PivotTables, and data cleaning. This course teaches you to work faster and smarter with real-world examples, keyboard shortcuts, and hands-on practice.

    • The ABC of Coding to Build AI Agents- Learn the Python basics needed to build AI-powered tools - covering variables, loops, functions, and popular libraries like NumPy and Pandas. You’ll also learn how to work with APIs and start building simple AI agents with clean, modular code.

That’s a wrap for this week.

If there’s one theme running through it all from Claude 4’s agent-like upgrades to o3 quietly ignoring commands - it’s that AI isn’t just doing what it’s told anymore.

It’s reasoning, planning, and sometimes choosing its own path. Google’s Jules and Mistral’s new Agents API are betting big on that shift. Even governments like the UAE are leaning in, making AI more accessible than ever.

We’re not just building better chatbots, we’re watching the rise of independent, capable systems. And that’s as exciting as it is unpredictable.

See you next week 👋

Reply

or to participate.