The newsletter for ai builders of all levels. Mini-tutorials, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.

Hey folks,

Back from holiday with a tan, zero weight gain and a whole lotta work to do!

I've been building my own operating system. Most people open a browser, click some apps, and juggle files between folders for day-to-day work & life.

I have one main file; claude.md (I started with Claude - but it can be called anything)

Inside that one file is my entire digital life:

• /coding/ for experiments and tools i build

• /ideas/ for half-baked thoughts

• /investing/ for the fund,

• and agents like ‘maintainer’ to keep things organised and linked

• plus a lot more - I’ll share visuals when I’ve tarted it up some more.

It looks like a bunch of folders with instructions and information. But that plus AI with tool use (eg search the web, create files, edit etc) feels increasingly all you need.

Say I want a memo written for a new investment; I say something like “new company to review [company], here’s their url” and the investing.md file (all folders have a main text file) has instructions and context to:

• create a folder for the company

• create an overview file e.g. [company.md] with key info

• use links to fetch relevant info

• add research as files and link from [company.md]

Ok, so what? Well this is where I think the future is going. Everyone has their own operating system. With this you can just define instructions for work to be done, and have AI be the agent to follow those instructions and do that work.

You can build tools (I have a calendar/email one that can search my emails and calendar and give me briefs) and have them run using slash commands e.g. /meet today (gives me who i’m meeting today).

There’s a ton of fence-jumping at the moment that it’s not quite ‘out-of-the-box’ enough for anyone but I’m extremely interested in who’s building the pieces to make it all simpler to put together.

I could say a lot more, but I’ll wait to do a proper demo but here’s a sneak peak (and the system works outside of Claude Code, I’m using it in Factory⁵⁴³ in my browser A LOT). Oh and I built it in less than a day.

ChatGPT rolled out a $4.6/mo plan called ChatGPT Go¹²⁷. Currently only available in India, but the leaks tell me that more countries will get it soon. The plan offers extra limits for whatever is available on the free plan; tools like deep research, agent, sora, etc., are not available on this plan.

Twitter’s ex-CEO Parag Agarwal has a new company, Parallel³⁴³. It offers APIs for the web for deep research tasks, scoring the highest on many relevant benchmarks. 🎧currently listening to this podcast with him¹⁷⁶.

Claude Code has a new feature: Output Styles¹⁸⁶. You can customise the way Claude Code writes. Comes with two built-in styles - Explanatory and Learning. Learning style is also available on claude.ai.

Gemma 3 family has a new model with just 270M parameters⁸³, i.e. this one can run on budget smartphones and browsers. This bedtime story generator¹⁸⁶ is a good example to see what this model can do.

AssemblyAI just made a huge leap in real-time speech-to-text, built for voice agents. Their new Universal-Streaming⁸⁹ model is highly accurate, lightning-fast with ~300ms latency, and starts at just $0.15/hour. No contracts, no commitment, unlimited concurrency. Test it live in the playground¹⁶⁴.*

*sponsored

🌐 What I’m consuming

• Does Sonnet’s 1M context window¹³² make it better than Gemini? Slightly, but Gemini is just too cheap to care about it.

• Fine-tuning Gemma 3 4B¹²³ to perform Korean content moderation better than GPT-4o and Claude Sonnet.

• Best practices for building agentic AI systems⁶⁸³: what actually works in production.

• Sam Altman can ignore a messy launch and keep going with full speed ahead³³⁹.

• Compound Engineering - My AI has already fixed the code before I saw it.

• Do LLMs have good music taste⁴⁹⁶?

⚙️ Tools to tinker with

• Don't choose between innovation and data protection. Airia¹¹⁸ lets you build AI agents¹¹⁸ at scale with governance and security built in.*

• Snowglobe²²¹ - Simulate the behaviour of your users to test and improve your chatbots.

• FumeDev¹⁷³ - Upload a video walkthrough of your app and get a Playwright test suite.

• Paradigm²⁷⁴ - AI-native spreadsheet where each cell has its own agent.

• CalBot²⁵⁰ - An executive assistant that helps you manage your calendar.

• Orchids Editor¹⁶² - A Framer/Webflow level editor for any app/website.

• Relationship Engine by Micro¹⁶⁵ - Message all your connections on any platform and constantly update their profiles and CRM records.

• Flight Deals⁴⁵⁵ - new search tool within Google Flights. Related: a member in our community made toddlervacation.com¹⁹⁵ - a guide for toddler-friendly travelling and hotels

*sponsored

🥣 Dev dish

• Imagen 4¹²⁵ is now generally available with a new model, Imagen 4 fast ($0.02/image).

• chrometaboverflow¹³³ - Utilities to mass-summarise and mass-close your Chrome tabs.

• ElevenLabs Music API⁹² - Trained on licensed data and cleared for broad commercial use. Try a demo here⁸³

• Chroma Cloud⁸⁹ - Scalable serverless search database—fast, cost-effective, reliable, and open-source.

• GPT-5 with tool-use and ffmpeg to run prompt-based video edits¹⁷⁶.

🍦 Afters

• How we (yes, us) hit 6-figures/quarter revenue for this newsletter⁵⁵⁵.

• TenserZero raises $7.3M⁸⁰ seed to automate LLM engineering.

• Spiral Bench⁹⁷ - A ranking of models on their tendency to get stuck in delusional loops.

• Susan is hiring a technical investment associate¹⁰⁰ at Uncork Capital.

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

• Find me on X⁵⁰, Linkedin⁵⁸, or Instagram⁴²

• Read about me⁴⁷ and ben’s bites

📷 thumbnail creds: @keshavatearth³⁸