Welcome to the newsletter designed for those with a technical curiosity. Here, you’ll find updates, tool reviews, and insights from an enthusiastic founder-turned-investor and lifelong tinkerer.
Hello everyone,
OpenAI recently hosted a town hall where they addressed inquiries from builders. Utilizing YouTube’s “Ask” feature, I gathered that OpenAI aims to reduce costs by 100x in the coming two years. Their focus remains on general models, with GPT-5.x expected to introduce improved writing capabilities.
Andrej Karpathy shared his insights regarding the current landscape of coding agents and vibe coding. I resonate with his thoughts. In summary, the trend is shifting towards agent-driven work with minimal human involvement—essentially, “no more IDEs.” While agent swarms are gaining excessive attention, these agents tirelessly tackle tasks (think of them as super-intelligent teammates). This leads to increased productivity, and rather than lounging on the beach, we’re engaging more, mastering the art of guiding models. Crafting failing tests for subsequent success or using a loop with a browser for verification makes working with agents genuinely enjoyable. Looking ahead, 2026 is predicted to be the year of experimentation—embracing ‘slop’ as a pathway to learning and creating high-quality work remains a valid approach.
I participated in Every’s Vibe Code Camp (watch the full recording here) alongside other prominent creators. We discussed my experiences with reverse-engineering tools, building projects, and my explorations with agents and coding—despite my non-technical background.
Signals is an impressive development where a Droid learns from its mistakes and proposes actionable items for team improvement. Currently, human operators review and merge changes, indicating early signs of self-improving agents.
Claude has also introduced interactive interfaces for applications like Slack, Asana, Figma, and others, which closely parallel ChatGPT Apps and are built on the MCP framework.
Claude Code is replacing Todos with Tasks, which is more suited for lengthy projects, and these tasks can be stored on your device for easy access by multiple agents.
ChatGPT is enhancing functionality by allowing users to pip/npm install packages, run bash commands, and download files within the Code Interpreter.
Vercel has launched skills.sh, offering a directory for agent skills and a straightforward way to install them. Context7 is undertaking a similar initiative. Here are some intriguing skills I discovered over the weekend:
Kimi K2.5 has emerged as a new open-weights model from China, outperforming Opus 4.5 and GPT-5.2 in all benchmarks except for coding. It also excels in visual tasks like Gemini 3 Pro and is competitively priced against Gemini 3 Flash. I’m eager to try it out, particularly since they’re also developing tools like Kimi Code (CLI) and their web app for generating slides, among other tasks. Additionally, check out Qwen 3 Max Thinking, which offers comparable performance but lacks open weights.
Enterprises often struggle with valuable data locked within their content due to manual processes. Box Extract securely & accurately retrieves essential data at scale, facilitating quicker decision-making and automated workflows. Discover more from Box on transforming enterprise content into actionable data.*
-
Scroll.ai transforms any knowledge base into an enterprise-grade AI agent. Use code BENSBITES26 for two free months ($158 value).*
-
Feynman 3 by Opennote – A thought partner for learning.
-
Text to diagram in Excalidraw has improved, now streaming output with increased speed and smarts.
-
Company Search on Exa allows semantic searches across 60M+ companies, providing structured data on each (including web traffic, headcount, financials, and more).
-
Pencil – An infinite design canvas for Claude Code.
-
Cavo merges Excalidraw with a webcam and screen recording features. (read more)
-
Swing by Cartwheel – Create 3D movements for characters.
-
gcombinator – Change ‘y’ to ‘g’ to access the full article and discussions from HN.
-
Monitor by Parallel AI keeps web searches active, notifying you when relevant new information appears in your preferred format.
-
Sprites provides a stateful sandbox environment with checkpoint and restore features for safely running AI-generated or user-uploaded code.
-
Clawdbot has been rebranded to Molty. You can either deploy it on Railway or let Devin help set it up.
-
Cursor now employs subagents to manage different tasks, generate images, and ask for clarification while working in the background. They’ve also introduced the feature “Cursor Blame.”
That’s all for today! Please feel free to comment and share your thoughts. 👋
* This newsletter is brought to you by our sponsors 🙂
Interested in partnering with us in Q1?
