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.

Hey folks, i hope you had a wonderful Easter & 4/20 for those who celebrate. We hosted (pics below), overate and fully relaxed…

we have a new home, Substack! There’s several reasons why but it feels like a great home for the future of this newsletter. Today, we’re trying something a little different, let me know what you think.

p.s. our old site¹⁴⁰ will still be available for existing members.

p.p.s. Add this email as a contact to protect us from spam flags 🙏

Let’s get into it

🔎 News worth knowing

• o3 in ChatGPT has a hidden new talent. It can play geoguesser³⁰⁸ to scary accuracy (so it can help fight crime or something, right?). The Claude + MCP workflow I recorded last week (icymi^108†) can now just be done straight in o3¹⁸⁸ - with no extra tools to install. I dunno if that deflates me or excites me…however we found o3 + tasks wasn’t that great (more on that below), so rattles my confidence on the founder analysis workflow.

• Gemini 2.5 Flash⁴¹ is now available as an experimental model (across Gemini app, APIs and AI Studio). It’s a hybrid reasoning model where you can turn the “thinking” off or allocate a thinking budget to the model (upto 24k tokens). 2.5 Flash with no thinking is better than 2.0 Flash, but 50% more expensive. Quick impressions: This model feels undercooked.

• Anthropic released a 4k+ word guide on best practices for using Claude Code⁵⁸⁶, their tools for agentic coding, right from the terminal. It’s a relatively easy read, and you should set aside 20-30 mins this week and read it. Another post from them tried to find Claude’s values¹⁴² by analysing 300k+ anonymised chats.

• Tired of AI that can't hear you properly? 👂

  Flow by Speechmatics gives your AI the best ears in the business - 25% better accuracy across real-time in 55+ languages, even in noisy environments.

  Your AI should keep up, not catch up. Try Flow free today¹²⁴.*

*sponsored

want to partner with us? Click here¹⁰

💬 Our thoughts on… o3 + tasks

I’m seeing examples/claims of using o3 with ChatGPT’s Tasks for real-time updates. Sad reality is: it doesn’t work yet. I set up a task asking it to give me real-time stock price of 5 NASDAQ companies. Here’s one of the outputs:

Two of the companies mentioned here ($FAST and $BKNG) haven’t had a stock price that low in the past 6 months.

o3 in chat is amazing at using the web search tool. But I don’t see any traces of o3 running my initial query or task prompt each time it sends me an update. This is just a bad implementation, which might get fixed soon, but until then, I wouldn’t rely on “o3 + Tasks” for real-time info.

⚙️ Tools I’m tinkering with

• Your Voice AI should understand users the first time 🎯 Flow by Speechmatics delivers 25% better real-time accuracy. Upgrade your app today.¹²⁴

• Ion¹⁷⁰ - AI visual editor for your codebase, so PMs and designers can ship. Curious on people’s thoughts, I tried but think hit a few early bugs and haven’t re-visited yet.

• Quadratic²⁸³ - AI that understands spreadsheets. Chat with data, create graphs, models, dashboards and more.

• Tablextract^206† - Extract tables from PDF and images and save hours of work.

• I found Stagehand²⁰⁵ when tinkering with MCPs. Made by browserbase, you can build repeatable browser automations with natural language and code.

• Conteflow³⁰⁶ looks interesting. Content creation for people who hate creating content. I do IF I feel like I have to, but its easy when there’s no pressure.

• Netlify Drop¹⁰⁸ - drag and drop your html/css/js files and you get a link to share your site.

• Pydantic launched a simple AI agent framework²⁷⁰ (28 lines of code)

• Figma¹⁷⁵ is working on a text-to-app tool¹⁷⁵. Feels like everyone will have their own, like they all have native automations now.

*sponsored

🌐 What I’m consuming

• Cline is like an AI coding agent in Cursor (like Devin), so i’ve been testing it. They just released their full system prompt¹⁷⁵ which is interesting to analyse for your own prompting. Funnily, they did it as lots of proprietary LLM instructions have been leaked¹²². Garry Tan one-shotted Manus for an online guide¹⁷⁸.

• Google released quantised variants of Gemma 3⁴⁶ (its open-source models).

• Vercel’s AI SDK - this video¹⁰¹ has the complete breakdown, how it works and cloning Deep Research in 30 mins.

• Cursor’s latest release⁸⁵ includes a bunch of (very welcomed) features. I’m excited by; automated rules, images in MCP, improved agent, and project structure in context.

• Building Windsurf²⁰⁸ and the magic of AI coding. (remember it’s being bought by openai)

• AI-assisted search²¹³ actually works now (almost).

• We always read what Ethan has to say^221† on new models.

• This was a really good read on putting AI agents to the test¹³⁵, Dex feels like most ‘agents’ are not actually agentic. So what makes AI agents actually good enough?

• Can AI run a vending machine business⁴⁰¹? This was such an interesting read.

🍦 Afters

• Tell Claude to ‘ultrathink²¹⁵’ instead of ‘megathink’, or ‘think’. but apparently saying please and thank you wastes 10s of $millions every year in GPU costs 😅

• Who’s building terminal wrappers? I’m testing Warp which I like so far. I want a terminal that can code like Claude Code/Codex, but I haven’t a clue on using the terminal. which seems problematic…

• I switched my phone to dark mode to stop myself going on it so much. It feels like its working but we shall see…

• I can write more about why we switched to Substack but it all fits into a theme of mine atm…use less shit. I want to use less tools to run my business, less things to use altogether and stop trying to ‘optimise’ everything. im in a digital refresh, i’ll write about what gets cut and what stays.

• this tweet⁵³¹ was a banger (and true) - about people who think everything is too competitive

• im just about to kick off fundraising for the latest Ben’s Bites Fund, starting with a trip to SF on Sunday (and a very busy week of meetings). i’ve been a lot before but always welcome food recs 😋. Also I’m very inspired by founders inc model¹⁰³.

• We hosted Easter, lots of family, food and very hyper kids

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 X³⁶, Linkedin³⁴, or Instagram²⁷

• Read about me⁴⁴ and ben’s bites