fal ships a video model faster than real time

Welcome back. Most video models take forever to render anything useful. One startup just flipped the script: you can now generate AI videos faster than you can watch them. Developers are already live streaming with it. Plus, Cursor users are about to lose access to OpenAI models permanently.

Also: How cutting internships could backfire, how a Claude download link can get you hacked, and how to stop Codex from over-engineering your product.

This new AI model generates videos in real time: Fal just shipped a video model that generates AI videos faster than you can watch them. They post-trained MiniMax H3 into a leaner version called H3 Max, which spits out a five-second video in under three seconds. That’s about 35 times faster than the original model, and it still tops the quality leaderboards. This model allows you to live stream the video while it's still being generated. 

OpenAI ends partnership with Cursor (and how it affects you): Just weeks after SpaceXAI's $60B Cursor acquisition, OpenAI is cutting ties and blaming SpaceXAI for contract breaches. While co-founder Michael Truell remains unfazed, noting OpenAI drives only 5% of their traffic, Replit is already funding migrations for Cursor users. If you use Cursor, switch by mid-November or prepare to use your own API keys. 

Developers say Rust's ugliness is now a feature: Ruby on Rails creator David Heinemeier Hansson (also known as DHH) just trashed Rust, calling it the ugliest language he's seen in decades. He later clarified that he actually likes it when agents write the code, since humans don't have to look at it. A viral follow-up pushed this further, arguing that Rust’s unfriendliness to humans is exactly why it's perfect for agents. For a lot of teams, safety and speed are what count now, even if the language isn't the most fun to write. 

Tiger Data gives Postgres the performance it needs for demanding time-series workloads with columnar storage, automatic partitioning, and purpose-built time-series capabilities.

Keep the SQL and tools you already use. Skip the extra database, complex pipelines, and added infrastructure.

One Postgres. Built to handle more.

Try it free at tigerdata.com. No credit card needed.

The instructions backfire. Since coding agents do most of the heavy lifting now, your spec doc is more important that ever. That’s why GitHub’s last review of 2,500 agent files is such a wake-up call; the verdict was that most specs fail because they’re just too vague. Now, your first instinct may be to fix that by piling on rules and details, but that’s actually where things go south. If you overload a model with directives, it runs out of "attention budget." It’ll focus on the first few instructions and just quietly ignore the rest. 

Structure carries the weight. Google Engineer Addy Osmani's guide reframes the spec as a product requirements doc with labeled sections a literal-minded model can actually parse. Osmani says the best specs swap a flat list of don'ts for three tiers, each signaling how much freedom the agent has:

The spec is the new code review. The big takeaway here is that you're shifting where the work actually happens. If you take the time to make your spec sharp, well-structured, and clearly bounded, you won't have to spend hours untangling a mess of confused output later. And that’s definitely a win.

P.S: GitHub’s Spec Kit, now past its 1.0 release and supporting 30-plus agents, bakes that discipline into a gated Specify, Plan, Tasks, Implement pipeline.

Claude Code, Cursor, Codex. The output ceiling is insane. But "fix the bug" gets you slop. Detailed context gets you shipping code.

The problem: typing out file references, edge cases, and architecture context 40 times a day is exhausting. So you cut corners. Your agent gets lazy input.

Wispr Flow lets you speak full prompts with syntax-aware dictation. camelCase, snake_case, file names all preserved. Auto-tags files in Cursor and Windsurf.

Used by engineers at OpenAI, Vercel, and Clay. Works on Mac, Windows, iPhone, and Android.

Better prompts in. Better code out. Try Free

How to use Herdr to run agents from terminal: Herdr is blowing up among developers right now because it makes running agents from your terminal seamless. After a rocky start two months ago, HuggingFace developer Alejandro AO revisited Herdr and got hooked. In this 17-minute guide, he covers installation, terminal multiplexing, and remote VPS sessions. 

Monid (free search & fetch): The tool gives agents free web search and page fetching with no subscriptions, no API keys, and no credit card. It is one command away in Claude.

Rewriting the rules of context engineering for Claude 5 models (By Anthropic): If your CLAUDE.md and skills are stuffed with rules and examples, you're probably fighting the model. Anthropic stripped over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with no drop in coding evals.

You start a task in the terminal, then want the desktop app's file view. Switching used to mean starting from scratch. Anthropic's Claude Developers account shared a fix: the app now picks up any CLI session where you left off.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Check out how a a senior dev packaged 20 years of dev experience into 3 Claude skills for building production-ready Next.js apps.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ 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