Welcome back. Decision models are quickly becoming Silicon Valley's newest obsession. An AI startup just dropped two new models that handle text, images, and audio seamlessly. But as hype builds, an Optimizely breach has exposed 178 tech giants to malicious code. Today, we break down what happened and how to stay secure.
Also: Microsoft’s VP on what took Outlook down (it started earlier than you think), a prompt library for developers, and a laid-off engineer’s job hunt that’s got devs talking.
Today’s Brief
A staff engineer’s playbook to find work worth doing
An app to keep running your AI models without network
Building an always-on agent with Pi (tutorial)
How to build an agentic engineering workflow (cookbook)

TODAY IN PROGRAMMING
Liquid AI’s new models top the benchmark charts: The AI startup just dropped d1.2 and d1.2-ultra, two omnimodal decision models, with ultra taking first place on the Artificial Analysis Decision Index. Both handle text, images, and audio and return structured answers with confidence scores. That means you can automatically approve confident calls and send uncertain ones for review. They’ll even say “not found” instead of making up missing values. Pricing starts at $0.02 per million tokens, and you can try them here.
Hackers use Optimizely to spread malware across 178 websites: The AI marketing platform says an attacker exploited a flaw to sneak malicious code into customer projects. The code then spread to 178 websites, including Zoom, GitLab, and the BBC, through Optimizely’s own trusted CDN. It swapped crypto wallet addresses, skimmed card details, and even pushed iOS exploit kits to iPhones. If your team uses Optimizely Web, now’s a good time to audit your Project JavaScript. You can check the affected sites and learn what to do next.
Tencent’s cheapest model rivals Claude Haiku for 18x less: The Chinese tech giant just dropped Hy4 flash, its most cost-efficient model yet. In Tencent’s own tests, it goes toe-to-toe with Claude Haiku 5.5 on everyday tasks at 18x lower cost per task. Under the hood, it packs 224B parameters yet activates only 12B per token. Like its bigger sibling Hy4, it’s natively multimodal with a 1M-token context window. Input costs $0.07 per million tokens, output costs $0.281, and cached input costs only $0.003. Try it here.

Most voice tools rent someone else's speech model. Wispr Flow trained its own, called Canto. On 10 hours of real dictations, it had the lowest word error rate of every model we tested.
So you can talk through the bug, the file names and the edge cases, and get clean text in Cursor, VS Code or Slack. camelCase, snake_case and file paths come through intact. 4x faster than typing, on Mac, Windows, iPhone and Android.

INSIGHT
A staff engineer’s playbook to find work worth doing

Source: The Code, Superhuman
Platform teams don’t have a PM handing them a roadmap or revenue numbers telling them what to prioritize. So where does the work come from? In a widely shared essay, software engineer Sujith Jay Nair argues that finding the right problems to solve is a big part of a staff engineer’s job. He points to four places to look and eleven signals to watch:
Your systems
Crashes: Dig through postmortems and see how users worked around outages.
Costs: Check your cloud bills and vendor contracts. Old buy-versus-build calls are worth revisiting.
Toil: Your team’s repetitive work hurts efficiency. Your users’ toil hurts their experience.
Your users
Keep talking: Ask users about their last task and three biggest headaches. No workaround usually means the pain isn’t that bad.
Watch the workarounds: Users bending your platform to do something it wasn’t built for? They’ve built you a free prototype.
Build together: Pick a pain point and prototype a fix with users. Just don’t promise it’ll ship.
Your organization
Team goals (OKRs): If an existing goal already covers it, the work is decided. Move on.
Manager concerns: If you’ve heard the same concern twice in a week, write it down, but treat it as a weak signal.
Migration holdouts: Ask the teams that never migrated why. They’ll show you who your solution left behind.
Your industry
Write it down: Document how your system works, then compare it with what others are building. You’ll spot decisions that haven’t aged well.
Learn from others: Other companies already paid for these lessons. Read their postmortems and migration stories first.
The loudest problem isn’t always the biggest. It’s easy to fill your backlog with the latest outage or whatever your manager mentioned last. The hard part is knowing what to focus on. Nair’s full essay walks through which signals deserve attention and which don’t.

PRESENTED BY AMD
What would you build with up to 192GB of unified memory? AMD Ryzen™ AI Max+ PRO 495 processors give you room for demanding AI workloads locally.
With them, you can:
Run larger local AI workloads and multiple agents at once
Fit the full workflow: model, context, runtime, and dev tools
Experiment freely with more memory
The next build is yours.

IN THE KNOW
What’s trending on socials and headlines

Meme of the day.
Microsoft Meltdown: Their VP just revealed the freak sequence of events that took down M365 worldwide. It started way back in July, and nobody noticed (3.2K likes).
Brutal Market: A laid-off engineer’s job hunt is being called concerning for anyone working in tech. She claims 500 rejections in a year, despite having strong credentials (2.1M views).
Mental Models: A robotics engineer claims 10 thinking frameworks run almost 93% of engineering and turned them all into visual cheat sheets (7.7K bookmarks).
Prompt Goldmine: A dev stumbled across a site packing thousands of ready-to-use AI prompts for coding and automations, the collection is wild (13K bookmarks).
Factory Fire: The engineer who popularized software factories says he’s burned five of his own and warns most teams shouldn’t build one. He’d rather you master something smaller first (3.5K bookmarks).
Chip Laundering: A thread alleging how Nvidia’s export-blocked Blackwell chips reached China through Southeast Asian data centers is blowing up tech Twitter. It’s using NYT investigations to back it up (7.7K likes).
Offline AI: An engineer shared an app that runs a local AI model on your phone. It keeps answering on planes, trains, and dead zones, no account or signal needed (5K bookmarks).

TOP & TRENDING RESOURCES
Top Tutorial
How to build and deploy an always-on agent with Pi: This tutorial shows how to combine Pi, Neon, and Claude Code to build a durable agent that can keep state across sessions. You’ll learn how to set up the backend, add persistence and auth, build the app, and deploy it to Vercel so you can actually use it from your phone.
Top Tool
Kitesurf (2.1k ⭐): Cloudflare’s experimental browser engine runs entirely on Workers, so agents can browse and render pages without spinning up Chrome or a browser container. It parses HTML and CSS, runs JavaScript, renders screenshots, and exposes a Chrome DevTools Protocol endpoint, which means you can drive it with Puppeteer and other CDP clients.
Trending Cookbook
How to build an end-to-end agentic engineering workflow (by GitHub): This guide shows how to use coding agents beyond just writing code. You’ll learn how to bring them into planning, parallel implementation, review, security checks, and post-deploy debugging so they can help across the full software lifecycle.

AI CODING HACK
How to make Claude Code run its own backlog
Plan mode asks 3-4 questions, then your agent builds from a half-understood spec and loses the plan when context runs out. A tech founder shared a fix to turn the plan into linear issues.
Step 1: Connect Linear’s MCP server (free tier works):
claude mcp add --transport http linear https://mcp.linear.app/mcpStep 2: Run “/mcp” inside Claude Code to authenticate. Then paste this with your feature filled in:
I want to build out [describe your feature in detail]. Ask as many
questions as you need to fully understand every detail of what I
want to build. Then take everything you learn, and create super
focused and detailed Linear issues. Then begin work.Step 3: Answer everything. When Claude stops, say “work on the next thing” until the feature ships.
Expect 40+ scoped issues. The backlog outlives any context window, so a fresh session picks up exactly where the last one stopped.
Whenever you're ready to dive deeper
If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:
We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.
Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team





