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 from an exited founder turned investor.
Hey folks,
I’ve been testing Perplexity Comet and Dia - two early-access ai-first browsers. I’m not the biggest fan of Perplexity(‘s ceo - poor product experiences, calling competitors out on social, bragging about their investors, etc) BUT I also went off Dia’s CEO after they seemingly ditched Arc (their previous browser) which I loved - I like Josh again.
Dia has a v simple interface, its clean, it feels easy to use and no fluff.
Comet is clunkier and less snappy (new tabs, loading a page etc).
Both browsers can summarise the content of the page you’re on and let you ‘tag’ in multiple tabs to the assistants too. Comet has a voice mode which I haven’t tested.
Dia has ‘skills’ so you can say everyday look at the wimbledon order of play and tell me what it is, by typing /tennis in a new tab. Skills are cool but not game-changing yet.
Comet has Perplexity as the default search. I have an extension changing all search to chatgpt by default.
Comet doesn’t have skills but it has something Dia is missing…taking actions on your tabs (wait for it…)
So I was looking at my Google sheet of Fund of Funds I’m going to reach out to, some have urls but most don’t. I asked Comet to find the urls for them. It started off promising…listed them all and said no url, searching… So I assumed it would update them, but it never did. Uhm thanks.
I tested it on a Figma image (a grid of 9 logos with a ‘9x’ on each to show how much they’re marked up). I said change the 6x on [redacted 😉] to 6.4x. There was a pulsing on my screen and it was calling actions to find the label, edit it etc. It looked promising, but it didn’t do it (says automated browser actions are unavailable). It was also super slow but I’m not marking them down for that because models improving will solve this eventually.
Comet’s assistant panel also doesn’t let you scroll up to see the question you asked, and threads all your chats in one place. So having tried the Figma test, I went back to sheets and hit ‘recent’ to see my figma conversations - my sheets one is like 20 q’s back - think this needs to be fixed.
Dia doesn’t have any actions yet (I have requested them…)
But this is what both browsers need to do.
I need to be able to be on a tab, say update this list with urls, then research the people behind the funds and craft a simple email to them etc. Or on a tab testing something I’ve built “Open the console logs, copy/paste them to Cursor”.
Everything is becoming an AI Operating System. It’s just going to matter which flavour you like most. Currently I’m more Dia than Comet, but I like the direction of both (and I’ve used Dia more than Comet).
Related here is Claude Code. People love it for all sorts of reasons but I think they’ve just nailed the agentic experience; context, tool use, parallel agents, planning, etc. I read ‘Claude Code is my computer’ which made me think more about the AI OS. You can conceivably do all sorts of work with Claude Code that isn’t coding; research people, draft emails, heck even write some applescript to open mail and create the draft (I haven’t tried it yet but have done some LLM-applescript building).
I think ChatGPT is going to be the OS for most people. But I think more people will start creating their own OS from the flavour of tools they like to use, especially with so many tools out there to use, MCPs providing integration, and, of course, the fact that AI can create its own software.
I’ll say it again, AI is a new computer.
Replit is introducing replit.md - a text-based document to give Replit Agent the context about your app and set some custom instructions for it. Take a look at these examples to get started. Similar to Claude.md - essentially just some rules and overview of your project to help steer the LLM.
Cursor’s agent now has to-do lists and supports queuing up messages which I’ve been using A LOT. There was a big outrage on Twitter over their new “unlimited” plan, and although they’ve refunded people, looks like they took a big hit on user trust.
CodeRabbit is your AI co-pilot for code reviews. Get instant comments, one-click fix suggestions, and custom AST Grep rules. Reviewed 10M+ PRs across 1M repos. Trusted by 70K+ OSS projects. Free for open-source. Try it now*
xAI released the system prompt behind ask @grok on Twitter. It’s fairly simple.
*sponsored
want to partner with us? Click here
🌐 What I’m consuming
Google for a new internet—of tools and MCPs. Smithery (I invested) just hired a co-founder, Anirudh, who wrote this piece which is a really good read on how the internet works and what it could look like in the age of AI.
Against brain damage - looking into the claims on how AI hurts our thinking.
Anthropic’s proposal and framework for transparency in frontier AI.
Shreya’s thoughts on background agents.
Why Meta and Google learned to love art. This is an amusing read.
François Chollet at YC Startup school - How we get to AGI.
The 10-minute AGI-proof stress test
Langchain made a video on context engineering for agents. I really like this little graphic describing four parts of dealing with context.
⚙️ Tools I’m looking into
Ability AI Agent Hub – crafted and hand-picked AI agents for business automation. 500+ templates, setup guides, and a community of builders.*
OpenSearch - Better Perplexity but personal to you with your own supermemory.
Ambient Canvas - Simple text editor that detects when you're thinking and prompts you with thought-provoking questions that help you continue. (in action)
Huddle - Plan hangouts with your friends. Built on Replit.
Orchids - Generate apps and websites that don’t look and feel AI-generated. (free credits)
Chat Capsule – Convert ChatGPT chats to markdown (for Notion, etc.)
Context - AI for all your office needs—documents, presentations and spreadsheets.
Yoink - Use AI to write wherever you are on Mac.
Clueso - Turn rough screen recordings into stunning videos & documentation.
Viseal - Upload a picture and this tool creates dialogue based on that picture in different languages to help you learn a language.
*sponsored
🥣 dev dish
llm-bridge - interoperability between input formats of different AI providers.
backlog.md - manage project collaboration between humans and AI Agents in a git ecosystem.
ccundo - intelligent undo for Claude Code sessions.
NotebookLlama - an open-source version of NotebookLM.
chapters.py - tiny utility script to make chapters for long youtube videos.
Gemini API can be 50% cheaper if you can wait for a day to get your answers.
🍦 Afters
AI Tinkerers meetup in London on Wed 16th July - I’ll be there
Zuck is also shopping from Apple, not just OpenAI. The manager of Apple’s foundation model team just joined Meta’s superintelligence team. Catch the glimpses of Zuck’s haul.
Lenny hacked together a tool to see how your thumbnails look in a generic YouTube home feed.
Ex-X CEO Jack Dorsey’s weekend project was building a chat app that works over bluetooth (testflight)
That’s it for today. Feel free to hit reply and share your thoughts. 👋
Enjoy this newsletter? Please forward to a friend.
Apologies if I've missed that but what about user's data in terms of privacy? Where do those browsers stand on a "How much I harvest your personal data for my own profit" scale between let's say Tor and Chrome?