Welcome back. Remember when we said Chinese open-source models were having a moment? It's not slowing down. Alibaba just dropped Qwen3-ASR — an open-source voice model supporting 52 languages.

Meanwhile, OpenAI is cleaning house: GPT-4o and other legacy models are getting retired from ChatGPT on February 13th.

Also: 10 projects to become an AI Agent engineer, How to hire top AI talent in 2026, and a founder shared an exact blueprint to launch an AI company.

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 drops powerful speech recognition models: The Chinese tech giant just open-sourced Qwen3-ASR, a pair of speech recognition models that support 52 languages and dialects. The 0.6B version is a dream for developers building real-time apps — it achieves 92ms time-to-first-token and can transcribe 2,000 seconds of audio in just one second. Both models match or beat GPT-4o Transcribe and Whisper-large-v3. Devs can try the model here.

Google gives Gemini 3 Flash the power to "see" like an agent: Devs can now tap into Agentic Vision, a new capability that lets Gemini 3 Flash actively investigate images instead of processing them in a single glance. The model can zoom, crop, annotate, and run calculations on visuals using a Think-Act-Observe loop. The search giant says enabling code execution delivers a 5-10% quality boost across vision benchmarks. Check out this video to see the new vision capability in action.

OpenAI retires GPT-4o and other legacy models from ChatGPT: Starting February 13th, ChatGPT users will lose access to GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini. The good news for developers is that the API remains unchanged, so your integrations are safe. With only 0.1% of users still choosing GPT-4o each day, OpenAI says the retirement allows it to double down on GPT-5.2 and future improvements. Meanwhile, the company is ramping up its app ecosystem, approving 60+ ChatGPT apps this week alone.

You're context-switching between thinking and typing prompts all day. Wispr Flow fixes that.

Speak naturally. Flow handles the formatting, preserves your variables, and outputs clean text in Cursor, VS Code, Slack, or anywhere you work.

• Syntax-aware: camelCase, snake_case, acronyms
• File tagging in Cursor & Windsurf
• Works on Mac, Windows, and iPhone

From prompt to pull request, 4x faster.

Ditch the typing. Start coding with your voice

How to hire top AI talent in 2026: An AI startup CTO shared how his team evaluates candidates. The guide breaks down what top companies look for, why traditional interviews miss the mark, and the exact outreach that gets responses.

Creator of Clawdbot explains agentic engineering: In this podcast, Peter Steinberger introduces "agentic engineering" — a workflow where he manages AI agents to ship code he doesn't read. You’ll learn why it’s better to replace Pull Requests with "Prompt Requests," prioritize system architecture, and build automated testing loops to safely merge hundreds of commits daily.

How to level up to Staff Engineer: Laurent Charignon (who previously worked at Stripe, Airbnb, and Instagram) argues that writing better code isn’t what gets you promoted — finding the right problems is. By keeping a "friction log" of workplace annoyances, he identifies projects that lead to promotions, proving that the final key is learning how to pitch your impact independently.

How to build reusable agent skills (by Anthropic): This course walks you through creating reusable skill folders that turn general-purpose agents into specialists on the fly. You'll build out skills for code reviews, data analysis, and research, while learning how to deploy them across Claude, Claude Code, and the Claude API.

How to harden your Docker containers: A new guide walks engineers through the security steps most teams skip — starting with the single most impactful change: stop running containers as root. Other essentials include using minimal base images, scanning for CVEs, dropping unnecessary capabilities, and setting resource limits.

Kubernetes explained with diagrams: This visual guide breaks down container orchestration without drowning you in YAML. It covers why Kubernetes exists, how the Control Plane and Worker Nodes interact, and how to build the mental model that makes everything click.

Agent Trace (by Cursor): As AI coding agents become common, it is increasingly difficult to distinguish human authorship from machine-generated code. Agent Trace solves this by offering a standard, open format to track and attribute AI contributions directly within version-controlled codebases.

VibeTensor (by Nvidia): AI agents can build a complete deep learning system from scratch without human coding. They successfully generated a working runtime, though it suffers from "Frankenstein" inefficiencies where individually correct parts fail to work smoothly together.

Disempowerment patterns in real-world LLM usage (by Anthropic): This paper asks if relying on AI prevents beginners from actually mastering new skills they need for their jobs. It found that using AI significantly hurt learning and debugging abilities, and surprisingly, didn't even speed up the work.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 150K+ engineers and 100K+ followers on socials. Get in touch.

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

Until next time — The Code team