The newsletter for the technically curious. Updates, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks,
GPT-5-Codex has a mini variant now.88 It’s not that far from the big brother on SWE-Bench Verified. The model is only available in Codex CLI and IDE extensions. API is coming soon, though. Codex-mini gives you 4x the mileage of Codex. Also, all paid ChatGPT plans now have double the rate limits for Codex.
Kimi K2 Thinking192 is an open-weights model that is now #1 on many benchmarks like HLE & BrowseComp across all models (not just open ones). Kimi models seem to have really great writing taste, while being great at coding tasks and being cheaper.
So the trend of models that go further continues with two options: closed-source RL-ed models (Composer, SWE-1.5) and huge MoE open models (GLM, MimiMax, etc.)
Replit now lets you add AI features to your apps156 with zero extra setup. No need to make developer accounts, get API keys, etc. You can use 300+ models, including Gemini, OpenAI, Claude or even open-weights models from OpenRouter. The AI cost is billed against your credits. Smart move tbh.
There’s a new stealth model on OpenRouter named Polaris Alpha, which looks a lot like GPT-5.174. I also caught a few leaks for Nano Banana 2120, mostly creating realistic images of news channels.
Eliminate AI anxiety. Build with confidence. AI agents are the future—but centralized systems create costly risks. With Airia, prototype and orchestrate AI agents securely while creating a resilient, decentralized AI ecosystem for your organization. Get started today at airia.com102*
Droid updates - I’ve been on a pretty steep slope learning to be a bit of a developer 😬. I shipped our new docs site111, run the changelog45 process, started making guides55 (with help from Ian Nuttall) and lots more. We’ve got a new (super simple) MCP feature33 to add/remove/enable and disable MCPs, plus some big new things coming out the next couple of weeks. You can join the discord35 to hear about early access 👀.
Teach your AI to think like a senior engineer731.
AI progress and recommendations451 from OpenAI.
Building Cursor Composer169 with Sasha Rush.
Highlights from USV’s 2025 Annual Meeting124 (captured with AI).
From words to worlds204: Spatial intelligence is AI’s next frontier.
What today’s AI (memory) products204 can learn from Chrome’s early design explorations of web history.

Voices has the best voices for AI-assistants. BMW found their voice with Voices–join the product launch125 on Nov 13, 1pm ET to see how.*
DeepWiki201 - Understand any GitHub repo with navigable and comprehensive docs generated on demand.
Voice Conversion by Hume130 - Generate the same pacing, pronunciation, and intonation with one recording across any voice you choose.
Papiers204 - A new interface for arXiv papers with mindmaps, children/parents of a paper, X discussion surrounding it and more.
Parse349† - Build an API for any website in minutes, 100x faster than browser using automations.
Swipe by vectors.fun184 - A new way to steer image models. Start with a prompt, swipe left/right to steer the model towards what you’re thinking of. (read more34)

Parallel’s new Search API101 claims the highest accuracy with the lowest cost across a range of benchmarks—built specifically for agents.
Gemini API now has a fully managed RAG file search168 with generous free storage, embedding generation and more.
Vercel Sandbox CLI55 - Create, run, and manage secure, ephemeral Sandbox environments right from your terminal.
MCPorter45 - Generate a self-contained bun CLI out of any MCP server. (an implementation for the code execution for MCP thesis)
Kimi Writer192 - An autonomous agent powered by the kimi-k2-thinking model for creating novels, books, and short story collections.
cc-trace96 - A claude skill that lets claude code reverse-engineer itself.
jsxtool87† - An in-browser IDE for building React UIs.
A unified diff viewer60 that parses standard unified diffs and renders them as minimally as possible for React.
Quora has a new mission127 - grow the world’s collective intelligence.
SoftBank is repeating history285 - it has sold its entire stake in Nvidia worth ~$5.8B.
Inception AI61 is building diffusion-based LLMs for code and text generation. They just raised $50M18 led by Menlo Ventures.
Gamma157 has raised its Series B at a $2.1B valuation while being profitable. (I’m an investor)
That’s it for today. Feel free to comment and share your thoughts. 👋
Read about me21 and Ben’s Bites
📷 thumbnail creds: @keshavatearth12,
Thanks to today’s sponsors who made this newsletter possible :)
Airia and Voices.com.
Wanna partner with us13? Last few slots left for the rest of the year.
Models with more mileage
early stages of becoming a developer.
2,454 readers clicked at least once. 43 links, 43 with clicks. Heat is relative to the most clicked link in this issue.
Substack · essay