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.

We’ve got a few open ad slots over the summer. Wanna partner with us¹²⁶?

Hey folks,

Windsurf’s deal with OpenAI is dead. Instead, founders and key researchers have been poached by Google¹⁹⁹ (for reportedly $2.4B¹⁵⁴) and the company (whatever’s left of it) is being acquired by Cognition¹⁸¹, the makers of Devin. The announcement from Cognition’s founder also mentions that Windsurf is a $82M business¹⁸³ with growing enterprise growth. Also, Meta acquired PlayAI a couple of days back. A wild few days in acquisition land.

Another Chinese lab, Moonshot AI is taking over the open-model scene with Kimi K2. It’s a massive 1T MoE model with no reasoning yet (tech blog²⁵⁷), and it easily beats Claude Sonnet 4 and pokes Opus 4 too. It’s great at coding and tool calling, but that’s not it. On this creative writing eval⁴²⁵, it just blows everything out of the water, avoiding most of the telltale signs of AI writing. This is what Llama 4 Behemoth was supposed to be.

You can use it on the web⁶³¹, it also has an API¹⁸⁰ (makes it possible to run in Claude Code¹⁸⁹), but that’s slow, and the folks at Groq are already serving it at superfast speeds³³⁷.

Grok has a new feature: Companions³⁷⁶. It’s built in partnership with Anichat. Grok powers the dialogue (with voices), and Anichat covers the generative animation. One of the avatars Rudi is a red panda, who can be good and bad, and the other one is Ani, a borderline nsfw avatar based on an anime character. That company Meta just acquired used to clone AI voices, so expect something like this from Meta too

Claude finally has a directory of pre-built MCP connectors⁴⁷³ to connect to Notion, Figma, Stripe and more. Plus this is also live for the Pro plan ($20/mo) and web app. The desktop app has some special ones like allowing Claude to use Chrome, read your files, fill PDF forms etc. You can also add a custom connector if you have a remote MCP url—like from Smithery²⁶⁸ which is a portfolio company I use several times a week.

Introducing one of my latest investments - Mirai³⁷⁴ just dropped an on-device AI SDK for iOS & macOS that lets any solo dev ship Llama-3, Qwen-3, Gemma-3 or DeepSeek-R1 (or bring your own model) locally—zero latency, zero cloud cost, zero privacy fuss. Grab the SDK³⁷⁴, the Swift engine¹⁴², the model-conversion toolkit¹³⁴ or poke around the platform¹⁴⁴ and docs¹³⁶.

Unwrap²⁰⁹ partners with DoorDash, Clay, Stripe, and others to power their customer intelligence. Its AI-powered platform aggregates and analyzes customer feedback across channels and proactively surfaces emerging issues to you before they snowball out of control. Check them out here²⁰⁹.*

Hostinger is doing a special free session for Ben’s Bites, covering how to build your idea into an app²⁹¹ even if you lack the technical chops. July 24th, RSVP here.²⁹¹

*sponsored

🌐 What I’m consuming

• The rise of agentic commerce⁵³⁵ and Stripe’s role in it.

• Stop saying RAG is dead.⁵⁵⁷

• Context Rot³³² - How increasing input tokens impacts LLM performance.

• Claude Code is all you need⁷⁹⁶ - Using CC for non-technical tasks. I’m doing this more and more now - I used it to help with a P&L and do research for an investment memo I am writing. It really is the agentic workflow they’ve built that is great at using tools.

⚙️ Tools I’m looking into

• ART²²⁵ - Open RL framework for improving agent reliability by allowing LLMs to learn from experience. Integrate GRPO into any application.*

• Cased²⁵⁴ - Interactive agents that do the busywork of DevOps for you.

• Basic Memory³⁸⁵ - AI conversations that build lasting knowledge over time.

• Consensus Deep Search³⁴⁰ - Conduct literature reviews across 200M academic papers in two minutes.

• Terragon³⁰¹ - Use Claude Code like OpenAI Codex/Cursor web agents

• The AI() function in Google Sheets⁴⁴⁸ is now available to everyone.

• BrowserOS⁴⁴¹ - An open-source agentic web browser.

• Factory³⁰⁸ - A top coding agent is offering 20m free tokens or 1 month free for new users

• Smithery Playground^218† - Explore, test and debug MCPs.

• Amazon has a new agentic IDE called Kiro³³⁰. The focus is on spec-driven development, which could be really big for PMs and founders.

*sponsored

🥣 dev dish

• Mistral has a bigger version of its developer-focused model, Devstral¹³⁸. Devstral Medium is available on the Mistral API, and an update to Devstral Small is on HuggingFace.

• gemini-embedding-001¹⁹⁸ is now generally available. It’s not multimodal (= text only), but it has the best benchmark scores.

• diffs are a way to show deleted lines and newly added lines when a code files is added. Developers are familiar with them but I think there are other ways to show that difference (aka the "diff"). this tool diffsitter²²¹ focuses on creating semantic diffs ignoring basic formatting differences.

• Goose³⁰⁶, the open-source ai dev agent by twitter’s co-founder jack is getting a dedicated team¹⁴⁹ of engineers and designers. The github repo already has 16k+ stars.

• genai-processors¹⁸⁸ - A new python library by Google to build applications with AI. From the surface, looks like the focus is to enable AI apps that deal with audio/video input-output easily.

• Cactus²²⁸ - Ollama for smartphones.

• asyncmcp¹⁶⁸ - Give MCP servers the ability to queue up requests.

• Reaper¹⁹⁶ - An open-source SDK for finding dead code.

• How to build a multi-agent research assistant²⁴³ with Gemini and Llamaindex

🍦 Afters

• Product Talk¹³⁹ with Chris Pedregal, CEO of Granola - Tues 12 August in London

• H-Nets¹⁴⁷ - new architecture in state space models to get better tokenisation.

• The AI usage survey²⁰¹ for the State of AI report 2025 is open for inputs. Takes <10 mins.

• The US Department of Defense is giving $200M each^169† to Anthropic, Google, OpenAI, and xAI to build AI workflows for the department.

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

• Find me on X¹²⁴, Linkedin¹²⁹, or Instagram¹³⁶

• Read about me¹²⁸ and ben’s bites