want another ai app builder?
how about a builder? for your ai app
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,
You get an app generation tool, you get one, you get one and now Github has one. It feels that Microsoft is trying to expand its market and reach less technical people by using GitHub Spark as an entry point. The idea (I assume) is to then get them into the full GitHub Copilot experience, allowing them to expand their applications. It feels like a reasonably good approach. Cursor could also benefit from some version of this - they do have their web agent which is like Codex (are people using this still? I dropped off). It just makes good business sense for Microsoft to do it (if it works).
Related…Lovable just reached $100 million ARR - the fastest co to do so? Part of me thinks this can't be sustainable – pretty wild numbers for what I assume are lots of PMs, designers, and indie makers building stuff – because I still don't know the longevity of these kinds of businesses, plus, lots of competition. It feels kind of up in the air in terms of what the world moves to and what we need in terms of applications. To me, agents are applications. A prompt, effectively, that you can just rerun and get the same thing out of a system like Claude Code is an app. But so impressive to see these companies take-off and get more people building software. I’m so here for that.
I think what Github is doing by having the full-blown power of the IDE as your MVP-plus-one system is not a bad approach here. It'll be interesting to see, but I think it's obviously an attempt by Microsoft to boost GitHub Copilot numbers and compete with the likes of Cursor, Windsurf, and Bolt, Lovable, et al.
Gemini 2.5 Flash Lite is now generally available. The model is insanely fast, very cheap and smart enough to enable applications like this and this. One startup idea you can build with Gemini 2.5 Flash Lite would be “a less annoying Grammarly”. You can process over a few million words (most users would barely hit that in a month) for under a dollar.
Seriously, these are mega impressive speeds - it spat out a timer in like 2-3 seconds (Lovable takes nearly 60s). I spent 3 weeks trying to build a text-to-app builder that wouldn’t overbuild (e.g. full stack for a timer) like many do now. And played with all sorts of funky ways to make it super fast and work, like creating an AppSpec that a tiny model interprets actions at runtime. But if I’d waited a few more weeks, I would’ve just used 2.5 flash lite!
The US government has released its AI action plan (led by Sriram Krishnan and David Sacks). This is one of the better thought-out AI plans. There's a lot of focus on building assets—including scientific datasets, data centres and energy supply. It covers good ground on measuring the economic impact, risks to the legal system and deploying AI systems with care. Plus, it's pro-open-source.
Airia is the enterprise AI orchestration platform that keeps your data secure. Build AI agents with built-in governance, role-based controls, and audit trails. Integrate with enterprise systems securely through no-code tools. Secure AI at scale.*
Launches from the portfolio:
Droids by Factory AI can now live in your Slack. They can read and write to your channels, helping you make sense of your engineering efforts.
Zams - The sales automation platform for B2B companies. I backed this team in 2021 with my first tiny fund!
Pika, the video creation startup, launched a new iOS app in beta - a social AI video feed.
NEW: We’re trialling a ‘Ben’s Bites AI expert matching service’ - connecting AI devs & no-code/automation experts with companies needing projects built. Be first to know when it launches and register your interest on either side here.
*sponsored
We’ve got a few open ad slots over the summer. Wanna partner with us?
🌐 What I’m consuming
The complete recipe to build a fully functioning, code-editing agent.
Real-time experiments with an AI co-scientist.
You don’t own your memory.
How and why teens use AI companions. (33% as a companion, 46% as a tool/program)
Goodbye, Featured Snippets: How SERP features have evolved in the AI era.
How to use Claude Code as your video editor.
⚙️ Tools I’m looking into
Chronicle helps you make designer-grade presentations with AI for free. It is like Cursor for presentations!*
Toad - A universal UI for agentic coding in the terminal.
Kafka - An AI employee with its own email, phone, Slack, and connections to 1000+ third-party applications.
Notate - Highlight, note & chat with any text on the internet.
Promptefy - Create detailed video prompts, optionally with image references.
*sponsored
🥣 Dev dish
Replit is partnering with Contra to match Replit experts with relevant projects.
Firecrawl Observer - Monitor any page or entire site with powerful change detection.
Jacquez - A guiding agent for open-source contributors to follow project guidelines.
Freewrite - open-source template for distraction-free writing. free to tinker with, add features, make it your own and sell.
A small 0.5B parameters model by Quotient AI to detect problems with tool usage in agents.
Voxtral’s tech report - How to train a transcription model.
🍦 Afters
Free live session today for Ben’s Bites readers from Hostinger Horizons on how to turn an idea into a real app in 5 minutes.
OpenAI is expanding its partnership with Oracle for Stargate.
Bee, the AI wearable company which actually got meaningful traction over the past year, has been scooped up by Amazon.
DeepMind, in partnership with the University of Nottingham, trained a transformer model to help historians interpret ancient inscriptions. The project is called Aeneas.
How much water does AI use? Well, Mistral has an answer for us—training and running Mistral Large 2 over a span of 18 months consumed less than 700 US households’ annual usage.
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumnail creds: @keshavatearth
On Lovable:
"Part of me thinks this can't be sustainable – pretty wild numbers for what I assume are lots of PMs, designers, and indie makers building stuff – because I still don't know the longevity of these kinds of businesses."
Some time back, Lovable announced that 75% of their traffic comes from apps built with the tool, not the builder itself.
To rephrase: for each pageview generated by an aspiring vibe coder, there are (at most) three from actual users. At most, because an aspiring vibe coder will be the first to play with their creation. And for some scenarios, testing in the builder is entirely unfeasible.
It's equivalent to saying "no one uses any of these."
So why should we expect these apps to be viable product ideas again?
Don't get me wrong, I love Lovable as a plaything. But assuming a significant portion of these new websites (they claim 10% of all new websites are Lovable's) have the capabilities of becoming commercial products is absurd.