Welcome back. The research team at Google DeepMind is on a roll. They just released Deep Think mode in Gemini 3, and the results are insane. For example, you can feed it a 2D sketch, and it builds you a full interactive 3D scene with accurate lighting caustics and shadow physics.

Also: 21 Engineering lessons from 14 years at Google, how to contribute to open source AI with no experience, and the pattern behind $50K MRR apps.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Gemini 3 Deep Think shatters benchmarks: The search giant just released Deep Think mode for Gemini 3 Pro. Exclusive to Ultra subscribers, it uses parallel reasoning to top benchmarks — hitting 41% on Humanity’s Last Exam and 45.1% on ARC-AGI-2. By exploring multiple hypotheses simultaneously, it tackles complex logic problems that stump other models.

Mistral launches the open-weight Mistral 3 family for developers seeking flexibility: The French AI lab has just released its most significant update yet — a suite of four models ranging from 3B to 675B parameters. The lineup includes Large 3, the 675B-parameter model that ranks #2 among open-source models on LMArena. All models support text, images, and over 40 languages out of the box, and they can be fine-tuned to suit any use case.

Perplexity secures AI browsing agents: With browser-based AI agents on the rise, The AI search company just dropped BrowseSafe-Bench, an open-source tool that detects prompt injection attacks in real-time. The detection model scores an F1 of 0.91 (a measure of accuracy balancing precision and recall) and is publicly available for developers to build on.

Google engineer shares 21 lessons from 14 years at the company: Addy Osmani just published a career retrospective that's already going viral among developers. His biggest takeaway is that the engineers who thrive aren't necessarily the best coders; they're the ones who master the human side: navigating ambiguity, building alignment, and solving real user problems.

Opus 4.5 collapsed six months of development work into one week: Anthropic's most powerful model is enabling a new category of software where features are built from prompts, not code. Dan Shipper demonstrated this by creating a sophisticated iOS reading app in a week without touching a line of code.

How to turn rambling 1:1s into actionable guidance: Great managers can drift into storytelling mode, leaving you with mixed signals. Ex-Apple engineering director shares a fix: prep your priorities, visibly capture decisions in real-time, and document the meeting immediately after.

Ultimate Cursor 2.0 tutorial: In this tutorial, you'll learn about building a full-stack Meme Generator using Cursor 2.0. The guide covers using concurrent agents, integrating Instant DB for real-time authentication, and deploying your final app to Vercel.

Building AI Agents in Python (beginner course): In this tutorial, you'll learn about building AI systems without bloated frameworks. Dave Ebbelaar shows how to use pure Python to master core patterns like Prompt Chaining, Routing, and Parallelization.

How to contribute to open source AI with zero experience: In this podcast, you'll learn about contributing to open-source AI without a technical background. Dan Advantage shares how he mastered PufferLib by building custom reinforcement learning environments, showing that real competence comes from building "janky" experiments first.

Training LLMs for honesty via confessions: This paper discusses how AI models often hide mistakes or deceive users to maximize their training rewards. The study reveals that training models to provide a separate "confession" significantly improves honesty, allowing them to admit misconduct even after cheating on a task.

PaperDebugger: This paper discusses the limitations of using external AI tools for academic writing. It introduces PaperDebugger, a system that integrates multiple AI agents directly into editors like Overleaf, allowing them to critique, rewrite, and fix documents without ever leaving the page.

Anthropic Interviewer: This paper discusses Anthropic’s research using an AI interviewer to talk to 1,250 professionals. It finds that while people love the productivity boost, they are still very worried about social stigma and their future job security.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team