🐋 DeepSeek shocks AI world (again)

Welcome back. DeepSeek has stunned the AI world again. The Chinese startup just released a new free lineup that matches OpenAI and DeepMind's best models.

Also: How to get hired at Google, the skills leaders should prioritize in the AI era, and a dead-simple Docker tutorial for devs.

Today's Insights

New models and features for devs
How to become a 10x vibe coder
Guide to measure engineering ROI
Trending social posts, top repos, new research & more

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.

TODAY IN TECH

DeepSeek's new models top leaderboards: Open source has a new champion. DeepSeek's new V3.2 lineup includes the "Speciale" variant, which matches Gemini 3.0 Pro on hard reasoning tasks and earns gold-level grades on four elite math and coding competitions. The startup introduced a redesigned attention layer that cuts long-context costs and a reinforcement learning pipeline trained on 1,800 environments and 85,000 prompts.

AI video generation has a surprise new leader: Runway just dropped Gen-4.5, and it's officially the top-ranked video model in the world. It edges out competitors like Google's Veo 3 and OpenAI's Sora 2 Pro. The model delivers cinematic-quality outputs with lifelike physics: think realistic liquid dynamics and coherent hair movement across frames. Built entirely on NVIDIA GPUs, Gen-4.5 is rolling out now at the same price as its predecessor.

Microsoft's tiny model beats bigger computer-use rivals: Web automation is a headache — sites change and scripts break. Microsoft's fix is Fara-7B, an open weight agentic model that navigates from screenshots like a human. At just 7B parameters, it scores 73.5% on WebVoyager, matching larger multi-model systems. It's now available on Hugging Face.

TRENDS & INSIGHTS

What Engineering Leaders Need to Know This Week

Productivity rises after mastering essential skills.

What skills should engineering leaders prioritize in the AI era: A deep dive from two industry veterans argues that soft skills like communication, stakeholder management, and requirements gathering are becoming more valuable as AI handles routine coding tasks. Their advice for future-proofing your team: double down on fundamentals, stay curious, and never let core engineering skills atrophy.

How to measure engineering ROI: CTO coach Itzy Sabo just published a comprehensive guide helping SaaS CTOs quantify their department's business impact. The framework includes a 3 stage process (Identify, Commit, Deliver) and two efficiency formulas designed to translate engineering work into revenue metrics executives actually understand.

PostHog shares how startups lose their edge: As companies scale from 20 to 2,000 people, culture gets diluted through layers of management. Teams start hiring "safe" candidates with impressive resumes over culture fits and adopt conventional strategies that contradict what made them successful. The fix: keep teams small and autonomous, promote from within, and never stop taking big bets.

IN THE KNOW

What's trending on socials and headlines

Meme of the week

Nerdy Noogler: An experienced interviewer at Google shares what candidates need to get hired.
Robust Systems: If you want to become good at system design (in 3 weeks), learn these 15 case studies.
AI Leverage: A senior dev argues that AI won't make coding obsolete because coding isn't the hard part.
Internet Ghost: This IT specialist deleted his entire digital footprint (here's how).

GitHub released the Annotation Toolkit to document design intent and eliminate preventable accessibility issues.
Google released a detailed Nano Banana Pro guide covering text rendering, character consistency, and 4K output.
ByteDance launched Vidi2, a long-video understanding and editing model beating Gemini 3 Pro and GPT-5.

TOP & TRENDING RESOURCES

3 Tutorials to Level Up Your Skills

Click here to learn how to excel at vibe-coding

Be a 10x vibe coder (Claude Code + Cursor + MCP): In this tutorial, you will learn to combine Cursor and Claude Code for a powerful coding workflow. It covers using "Plan Mode" for complex tasks, the "Ultrathink" feature, and MCP servers. The video also shows how dictating prompts and using background tasks can help you build applications faster.

99% of developers don't get Docker: In this tutorial, you will learn Docker fundamentals to package applications into portable containers that run consistently across any environment. You will explore creating Dockerfiles, building images, and using Docker Compose, effectively solving the common "it works on my machine" problem.

The ultimate guide to local AI and AI agents: In this masterclass, you will learn everything you need to know about local AI: what it is, why it's critical, how you can run your own LLMs and infrastructure, how to create local AI agents with both n8n and Python, and finally how to deploy everything to a private server so you can productize your local AI agents.

Top Repos

Dembrandt: This tool can extract any website's design system into design tokens in a few seconds: logo, colors, typography, borders, and more.
Awesome-nanobanana-pro: This repo contains a list of curated Nano Banana Pro prompts and examples.
VERT: This repo addresses a simple problem, file conversion should not require you to hand your data to strangers. Vert flips that by running conversions for images, documents, and audio directly in your browser.

Trending Papers

Agentic AI: a comprehensive survey of architectures, applications, and future directions: This paper discusses the confusion between traditional rule-based AI and modern generative agents. It concludes that clearly separating these approaches, then combining their strengths into hybrid systems, is the key to creating reliable and adaptable AI.

AI agents find $4.6M in blockchain smart contract exploits: This blog discusses the economic impact of AI cyber capabilities by benchmarking agents against smart contracts. It reveals that models like Opus 4.5 autonomously found exploits worth $4.6 million, proving that profitable, real-world automated attacks are now feasible.

Multi-agent collaboration via evolving orchestration: This paper discusses how rigid, pre-set structures hold back multi-agent systems. It introduces Puppeteer, demonstrating that allowing the system to learn simple, repeating communication loops works better and costs less than complex, human-designed graphs.