I write a newsletter about startups and investing—for ai builders of all levels.

I record mini-tutorials, review tools I’m testing, share my insights and give you a peek behind the digital curtain from an exited founder turned investor.

codex just launched - a web-based coding agent. you write tasks, it spins up an environment, writes code, and you get a pull request with the new code

sounds familiar but no other apps look like codex. its a new UI that others say suit senior engineers really well but less so for everyone else, including vibe coders.

i identify as me, so this is how i use it as a non-technical user (and how you can too)

(you can watch on yt³⁷³ too)

there’s lots of issues with it, but it’s an early preview and you can see where things are going (or guess)

I actually think there’s going to be a lot of re-imagining of workflows with llms in the middle. and coding like this makes it feel like writing. i write a task, hit code, go off to do something else, then accept changes.

yes background agents do this (and hit the same issues with merge conflicts), but getting out of the IDE (the coding tools) is going to happen more and more.

who’s reimagining the IDE experience anyway? ping me!

onwards, queue the Oprah meme for ‘everyone gets a coding agent’…

🔎 News worth knowing

• More codex resources:

  • Codex vibe check from Every²⁴⁴

  • the missing manual for Codex²¹⁷

  • how to use replit and codex¹⁴² together

• GitHub Copilot also introduced a new coding agent⁷⁹. You can assign GitHub issues to it directly or use it via VS Code. Unlike Codex, the developer assigning the issue cannot approve the agent's PR, maintaining a typical review pipeline. It also has limited internet access, which can be an advantage over Codex's current offline environment.

  • Also Github Copilot is going open-source¹⁰⁴! This is a big deal. You can read it as they struggle to compete with windsurf/cursor but regardless its a huge win for developers. being able to go open-source is a moat i think.

  • There’s a bunch more in Microsoft’s Build conference⁶⁶ - MCP servers for files and apps in Windows, turning websites into agentic apps, some protestors and more.

• Google also started giving more people access to their coding agent “Jules²⁵⁴” last night. Let’s see what google has in store today at I/O.

• Build AI Voice Agents that actually listen. Missed names. Garbled numbers. Dropped domain terms. Speechmatics delivers enterprise-grade Speech-to-Text and Voice Agent APIs. Built to capture critical information across 55+ languages - in noisy, fast-paced speech. On-prem, in-cloud, on-device. Try free.¹²⁸*

• Anthropic also added an SDK for Claude Code⁸⁴ and some $2.5B in credit line⁵⁹ between all this chaos.

*sponsored

want to partner with us? Click here¹⁴

🌐 What I’m consuming

• What does it take to transform a company into an 'AI-native' one^650†?

• A useful thread if you're looking to run AI locally: a list of recent ultra-small models²³⁷, mostly under 1 billion parameters.

• A chat between Aaron Levie (CEO of Box) and Kevin Weil (CPO of OpenAI) about AI agents in the enterprise²⁵⁹.

• Google's ex-CEO (who now has a secret AGI company) claims that AI is underhyped.³⁴⁶

• Replicate is making it easy¹⁰⁰ for AI code editors and LLMs to use their APIs. Copy a model page as markdown or even create an llms.txt file for each model. simplifying for llms⁶³ continues

⚙️ Tools I’m tinkering with

• Infinite Chat API¹³⁶ by Supermemory - Extend the context length of any model by embedding past tokens. (demo⁴⁶)

• The folks at Topology VC built LatentZip⁴⁷ - A text compressor that uses LLMs to make text files way smaller.

• Workflow Use¹⁷² by Browseruse - Create automated workflows on browsers by recording your actions and then passing in variables. Still an early tool.

• PromptWatch³³⁷ - Monitor how often your company gets mentioned by AI tools like ChatGPT, Claude, or Perplexity, and get suggestions for improvement.

• Stagewise¹²⁹ - A browser toolbar that connects your front-end UI to your coding agent (currently supports Windsurf and Cursor). Select an element, add a comment, and it's passed to your agent.

• MCP Boilerplate²⁹⁰ - Simple starter kit to help vibe coders quickly create and monetise MCP servers. chat with pdfs era is gone, it’s time for mcp

• Aurachat²⁴⁷ for creating beautiful designs that you can export to HTML or Figma.

• Two cool discovery apps:

  • en.app¹⁹¹ -Discover books based on vibes or vague descriptions.

  • discover.movie^210† - Vibe-based movie discovery

• Grok by xAI⁷⁹ can now generate charts. It currently works only in the browser.

⚔ Battle of IDEs

• Replit has introduced safe vibe coding⁴⁸, which flags accidental pasting of secret keys or environment variables in prompts, guiding users to use the secrets option instead. They've also partnered with Semgrep for security scans on apps²⁵ built on their platform.

• Windsurf Wave 9 is out⁴⁸, and the big news is they're building their own models: SWE1 (comparable to Claude 3.5 Sonnet, unlimited on paid plans), SWE1 Light (upgrade to Cascade Base, free & paid), and SWE1 Mini (upgrade to their tab/autocomplete model, free & paid).

• Cursor’s 0.50 update⁶⁰ brings a new model for in-file editing, a new tab model, and a background agent in preview, allowing multiple agents in parallel on a remote VM (sounds a bit like Codex, innit?).

🍦 Afters

• Agents Hackathon⁶⁸ on 31st May - June 1st by Cerebral Valley and AI Engineer conference.

• New challenge from OpenAI: OpenAI to Z¹⁵¹. Use o3/o4-mini to find the “lost city of Z” in the jungles of Amazon and win up to $250k in cash/API credits.

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

Enjoy this newsletter? Please forward to a friend.

• Find me on X²⁰, Linkedin²⁰, or Instagram¹⁸

• Read about me¹⁵ and ben’s bites