Welcome back. Alibaba just cracked the code on image editing with their new Qwen model. It can decompose images into separate RGBA layers allowing easy editing with AI.

Also: Why engineering leaders should rethink junior hiring, Google engineer’s playbook for AI-assisted coding, and the Supabase tutorial.

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.

Alibaba's Qwen releases open-source model for editable image layers: The Chinese e-commerce giant just dropped Qwen-Image-Layered, an open-source model that decomposes images into separate RGBA layers, each holding a distinct object with its own transparency channel. Recolor, resize, or delete elements without side effects. It supports variable layer counts and recursive decomposition for finer control, making it ideal for designers and developers building image editing tools.

OpenAI launches a model for extended coding sessions: The ChatGPT maker dropped GPT-5.2-Codex, a focused update to its agentic coding stack optimized for real software engineering rather than quick code snippets. The key upgrade is native context compaction, which compresses prior steps while preserving intent and state, letting developers run extended workflows without losing track or hitting context limits. The model scores 56.4% on SWE-Bench Pro and 64.0% on Terminal-Bench 2.0, edging out GPT-5.2 on both.

Google releases a new model that makes edge AI agents reliable: Developers building on-device AI agents now have a fix for flaky function calling. Google's FunctionGemma, a 270M-parameter model, encodes function calling directly into its weights—no more broken JSON or failed offline requests. After fine-tuning, accuracy jumps from 58% to 85%. It's ideal for apps needing deterministic behavior on phones and edge devices. Available now on Hugging Face and Kaggle.

Why engineering leaders should rethink junior hiring: Engineering leader Kent Beck makes the case that AI tools have transformed the economics of junior developers. By using AI as a patient tutor rather than a code generator, juniors reach profitability faster — and more of them survive the "valley of regret" before leaving.

Why your team's decisions keep failing: Most decision records miss the point, says engineering leader Peter Gillard-Moss. His Decision Triangle framework forces teams to articulate three things: what triggered the decision, the future they're trying to create, and the action they'll take. He suggests you should document your learnings after decisions play out — that's where patterns emerge.

New research reveals why daily stand-ups actually work: A study of 318 agile employees found stand-ups don't directly boost performance — they build psychological safety, which then drives results. The takeaway for engineering leaders: focus less on status updates and more on creating space where team members feel safe sharing blockers.

How to build & deploy an AI SaaS end To end: In this tutorial, developers learn to build and launch a full-stack AI SaaS using Next.js and Prisma. You will see how to use AI coding agents to accelerate development and how to self-host affordably on a VPS, covering the entire process from code generation to production deployment.

Google engineer shares the playbook for AI-assisted coding: Addy Osmani says the secret to AI-assisted coding isn't letting the model run wild — it's treating it like a junior dev that needs clear direction. His workflow: brainstorm a spec first, tackle tasks one by one, and always review and test. The human engineer stays in the driver's seat.

Supabase tutorial: In this tutorial, you will learn to build a real-time sales dashboard using React and Supabase. You will master database setup, secure authentication, and Row Level Security. The guide also covers implementing real-time subscriptions and database triggers, providing a practical foundation for building robust full-stack applications.

Bloom (by Anthropic): This paper discusses the challenge of slow, outdated AI testing methods. It introduces Bloom, an automated tool from Anthropic that generates fresh testing scenarios, allowing developers to reliably evaluate model behaviors in days rather than weeks.

Evaluating chain-of-thought monitorability (by OpenAI): This paper discusses the challenge of understanding why AI models behave the way they do. It finds that analyzing the AI's internal "thought process" is a much better way to predict its actions than simply looking at the final answer.

Stanford AI experts predict what will happen in 2026: This report discusses moving beyond AI hype to focus on real results. It argues that instead of making broad claims, companies should use precise, task-level measurements to track the actual cost and accuracy of their AI systems in daily workflows.

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

Until next time — The Code team