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,

Reporting live from sunny Mallorca (again)! Kids are running me ragged in the pool and AI doing the same via X as usual. I’m trying to have a little break from fundraising, building, investing and kids - but it’s not quite how this trip is shaking out 😅.

Multiple new launches are possible this week. The rumourland’s buzzing that openai’s open-weight model is likely going to come in two sizes, 20b and 120b. Plus, expect news about opus 4.1, gpt-5 and new gemini stuff this week.

In the meantime, Google quietly dropped Gemini 2.5 Deep Think¹²⁶. You can only get it if you pay $250/mo for the Gemini Ultra plan, so I haven’t tried it yet. I read Shawn’s post¹⁶⁹ that Deep Think’s performance improvement over the base 2.5 Pro seems much larger than o3 Pro’s improvement over base o3.

Google is launching a Game Arena¹⁸³ (RL environments) via Kaggle to pit models against each other in games, starting with chess. These environments are designed for the models to use text representation of chess moves and test their logical reasoning (instead of just using a code library for predicting chess moves).

Grok 4 subscribers are getting access to video and image generation in beta. Here’s a review of it⁵⁰⁶ without needing to subject yourself to NSFW content.

Want to save time and avoid the AI subscription trap? Sōkosumi solves this with pay-per-task agents: research, content, design. Use DeepSeek or VEO3 with one login. EU-based and GDPR-ready. Sokosumi.com²⁴⁹ + BENSB100*

Claude Artifacts²⁰¹ can now accept file uploads from a user²⁰¹. i.e. you can now build artifacts that create charts based on a document. Or maybe a study guide where you can add your class notes. The file storage looks temporary, but you can now make an MVP with Claude Artifacts.

Some portfolio updates:

• Supermemory¹⁷⁴ and Smithery⁸⁶ both now support OAuth for MCPs.

• Solve Intelligence^107† is hiring for founding engineers in London.

• Modelence¹²⁴ - Ship production AI apps 10x faster with built-in auth, database and LLM integration.

• Comp AI raised $2.6M⁹⁷ pre-seed to speed up the SOC-2 approval process.

*sponsored

We’ve got a few open ad slots over the summer. Wanna partner with us⁴²?

🌐 What I’m consuming

• Chris Paik’s letter on why they bet on Cline⁵⁰¹ (and its way of building in AI). Cline just raised a $32M Series A.

• Best practices to stop Claude Code from being "dumb"⁷⁵⁰ and ship features with less mistakes.

• AI is polytheistic³⁹², not monotheistic, and 10 more thoughts on AI from Balaji.

• Don’t read this startup slop⁵¹⁷ - Peter got banned on a forum because he uses AI to assist with writing.

• The Master Builder is the most valuable type of DevRel³⁰⁰ right now.

• Building an index of 15 years of evening news³⁷⁷ - 150k+ stories indexed in 39 mins, costing just $153.80 with Gemini 2.5 Flash Lite.

• 6 weeks of Claude Code⁶³⁴.

⚙️ Tools I’m looking into

• IMG.LY AI Editor²⁴⁴ – AI-first visual editing for devs. Add AI-powered design, image, and video creation to your app. Works with any AI model.*

• Cloudy¹³⁸ - Rent GPUs that always continue exactly where you left off.

• OpenBench¹⁰⁰ by Groq - One official implementation per benchmark with MIT license.

• Producer²⁸⁶ - your personal AI music agent. Powered by FUZZ-2.0 from the Riffusion team.

• Action Agent by Writer¹⁵⁴ - General-purpose agent to do tasks with a secure environment of its own. Kinda like ChatGPT Agent.

• Lindy¹⁴² also gave its automations similar capability (a browser of their own) under the name “Autopilot”, and their new “Agent Builder” builds a Lindy workflow from plain text descriptions.

• Wide Research by Manus¹⁹⁹ - Multiple Manus instances running in parallel to make high-volume research effortless.

*sponsored

🍦 Afters

• A new research from Anthropic finds “persona vectors²¹⁰” in LLMs that can be used to promote or mitigate specific behaviour (like evil or good).

• ChatGPT now asks you to take a break²⁴² if you’ve been chatting with it for a long time.

• CloudFlare is accusing Perplexity¹¹² of using stealth, undeclared crawlers to evade website no-crawl directives.

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

• Find me on X⁵⁴, Linkedin⁵², or Instagram⁴³

• Read about me⁴⁴ and ben’s bites

📷 thumbnail creds: @keshavatearth³⁶