Vibe coding is old now - Ben's Bites

Source: original

Vibe coding is old now

Anthropic's new ad, it's about... ads

Feb 05, 2026

The newsletter for the technically curious. Updates, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.

Hey folks,

Karpathy on one year of vibe-coding and a need for a better name. He suggests agentic engineering—I prefer it because there’s less of a negative connotation and that you’re not doing ‘real work’. It is a level of engineering, where you could be at the very beginning of the spectrum all the way up to experienced engineers.

On that note, I’m starting a community

You’ve seen the posts—people building apps, automations, entire products without writing code. You know you should be doing this too.

I can’t code, but I build. I sold a no-code community to Zapier, failed at learning to code the traditional way, and now ship software using AI agents.

This community is for people like us figuring it out together. Not another Discord full of experienced engineers.

Come and learn to build with agents!

Claude ran a Super Bowl campaign to dunk on ads in ChatGPT (without naming them), and it has really pissed Sam Altman and OpenAI’s CMO Kate Rouch. It got to a point that they went to ChatGPT and had it write longposts calling Anthropic out. They say: ChatGPT has more free users in Texas than Claude has in total. Ooofff. Do they have a point? Yes. Do they look like crybabies? Also, yes.

It feels off-brand for Claude to do something like this. But I suppose attention is all that matters these days.

Related → Why Claude will remain ad-free - this written piece from Anthropic has a more nuanced take on ads.

Back to work:

Droid now supports plugins - Plugins bundle skills, commands, and agents into packages you can install and share. Compatible with all claude code plugins - import them in 1-click.

Claude Code has a new command /insights. When you run it, it’ll gobble up your past month’s history, your projects and how you use Claude Code and give suggestions on how to improve your workflow. Also, you can now give CC’s VS Code extension access to your browser via Claude in Chrome.

Mistral released its latest batch of Voxtral models. Voxtral Realtime is built for live applications with configurable latency going as low as 240ms (and open-weights too). Voxtral Mini Transcribe 2 is for async workloads with high accuracy in multiple languages and low costs.

🌐What I’m consuming

• Death of software? Nah. A fantastic (and non-slop) read on how retail and media have already gone through this.

• Quick tips to upgrade how you use AI, for developers who are still in 2025.

• AI agents without the hype - A good hour-long talk with Armin and Mario to get a sense of where we are with agents.

• An agent wrote and sold an ebook on its own (almost).

• A systematic approach to building self-improving AI products.

• The office suite is a leaky runtime.

• As rocks may think - An interactive essay on thinking models, automated research, and where they are headed.

• Building Claudie - The AI agent that runs Every’s consulting projects.

• How OpenAI built Codex App Server - A connection link exposing the same harness to multiple clients (web, cli, codex app).

⚙️ Tools and demos

• 🦙 LlamaIndex - Industry-best agentic OCR for AI document parsing and extraction. Sign up today for 10k free credits.* (I’m an investor)

• Granola MCP - Take the context from your meetings to any AI app of your choice.

• Sazabi - AI-native observability platform for fast-moving engineering teams.

• ClawSpawn - Run OpenClaw in isolated microVMs.

*

🥣 Dev Dish

• npx skills add vercel/ai-elements - component library and custom registry for building better AI interfaces.

• Agent Browser now has iOS support to test how your websites work on a mobile browser.

• Apple’s Xcode now integrates with Claude Agent SDK, basically giving you Claude Code inside Xcode.

• You can now share your Claude Code sessions (Desktop, web and mobile). CLI is not there yet.

• Windsurf released Tab v2. It’s 50% better and comes with a tunable aggressiveness. Context: Tab models are “autocomplete for code” models that were popular before vibe-coding. Developers who still live in the IDE like these models a lot.

🍦 Afters

• TrueShort - An AI studio and streaming app that aims to create and release original movies and shows. (read more)

• Korabench - A benchmark for child safety. They found the Claude and GPT 5.2 models to be generally safe for children, passing over 70% of their tests. Also, simple prompt engineering improves all models, hinting that a separate “child mode” with different system prompts might be a good idea.

That’s it for today. Feel free to comment and share your thoughts. 👋

• Read about me and Ben’s Bites

PreviousNext