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,
‘Nano-banana’ aka ‘Gemini 2.5 flash’ is Google’s new native image generation and editing model, is now out to everyone on the Gemini app and API. It’s the best image gen model out there, it’s very fast and cheaper than others. It feels very creative for infographics too.
I tested it against a few of the thumbnails for the newsletter where I either was using ChatGPT’s image generations or I had to open up pixlr (photoshop-lite on web). The quality of generation is much better than ChatGPT, it doesn’t have the glossy look or the sepia finish on all the images like ChatGPT’s images do. Plus it can handle basic text additions etc. really well. I had issues with modifying aspect ratios and making collages though.
Here’s an example: Original image → nano-banana → ChatGPT in order.
— Keshav
Other examples; annotating the real world, 13 items into one image, people/character swaps in video, isometric buildings and a million more - if you’ve spent any time on Twitter in the past 24 hours you’ll have seen a TON.
Anthropic has opened the research preview for “Claude for Chrome” and from the looks of it, it feels like a pretty nice implementation of the “do tasks for me” agent. It triggers like a chrome extension, open up in a sidebar, listens to your prompt and then takes over the open tab to do your job for you. I’m just waiting for my approval before testing it out.
OpenAI’s Codex-CLI now has image inputs. I saw a tweet covering other improvements too; IDE extension, code reviews in GitHub and more.
A live coding challenge of humans vs AI is happening soon. The group will be randomly split in half and assigned manual or AI code. $15k prize for the winners. It’s approval only but if you tell them I sent you, you’ll get approved :) - it’s hosted by Factory (one of my portfolio companies).
Shanice, my wonderful business partner and COO, wrote a post “What happened when we stopped posting daily newsletters”.
🌐 What I’m consuming
How to index an entire repo and update it as code changes.
Vibe coding apocalypse & how to survive it (by the ex-CTO @ GitHub). Essentially you should learn backwards (which is what I do) - build the thing and learn how it works on the go (plus rebuild it several times)
Top 100 AI Apps by a16z.
How educators use AI by Anthropic (thread)
⚙️ Tools to tinker with
Don't choose between innovation and data protection. Airia lets you build AI agents at scale with governance and security built-in.*
Lindy Build - Vibe code websites that fix themselves.
Agent Fill by Ramp - An agentic form filler for finance and ops teams. (how they built it)
Bitrig - Build mobile apps on your phone (no code required) - from the co-creators of SwiftUI.
Rube - Connects to hundreds of tools within chat
CivLab - a live, systems view of your government (launch thread)
Papyrus - A new text editor and AI copilot in one (waitlist)
Environments Hub by Prime Intellect - A community platform for crowdsourcing open environments for RL.
*sponsored
🥣 Dev dish
Building a search agent with Parallel AI’s API.
Claude Code Templates - Ready-to-use configs for Claude Code (25k npm downloads, 5k GitHub stars).
Customise your Replit apps with app themes
ARC Prize is hiring a backend engineer
web2mcp: a toolchain that auto-generates an MCP interface for any website
🍦 Afters
Context7 is the most used MCP on Cursor (its defo my most used!)
New AI-powered live translation and language learning tools in Google Translate
ChatGPT helped locate a stolen Lamborghini after two years.
Someone messaged the owner on Instagram asking about the Lamborghini and sent a photo - ChatGPT accurately identified a blurred vehicle in the background, leading the owner to Denver, where police recovered the car — tibor
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumbnail creds: @keshavatearth