In a new analysis, Sequoia Capital argues that the next trillion dollar company will likely emerge from AI. Their thesis is that AI is becoming a foundational infrastructure layer similar to the internet or mobile. Every industry, from software to robotics, will increasingly run on top of it.

This shift changes how companies scale. Traditional software companies grow through headcount, sales teams, and distribution. AI companies can scale capability directly through models, compute, and data loops, which means they can expand much faster once product-market fit is reached.

Sequoia also points out that the largest opportunities may not come from model providers themselves, but from companies that build dominant applications on top of AI infrastructure.

Caitlin Kalinowski, the head of robotics at OpenAI, has left the company only days after OpenAI signed a cooperation agreement with the U.S. Pentagon. The departure quickly sparked debate across the AI community as some researchers worry that advanced AI systems and robotics are being pulled deeper into defense programs.

OpenAI has long emphasized responsible AI development, but closer relationships with military institutions blur the line between a research lab and a defense contractor.

Cursor, one of the fastest growing AI coding tools right now, is aggressively competing with GitHub Copilot and other AI powered IDEs. The company is shipping frequent updates focused on developer workflow such as code generation, debugging, and refactoring.

The rapid rise of AI coding assistants is turning the IDE into one of the most important battlefields in AI. As more software is written alongside AI, the tool that controls the developer workflow gains a massive advantage.

Developers on GitHub are starting to experiment with DeerFlow 2.0, an open agent framework reportedly built by ByteDance. Instead of acting like a chatbot, the system takes a goal and automatically researches the problem, writes code, generates files, launches sub agents, and attempts to deliver a finished output.

A developer benchmarked a database engine rewritten entirely by an LLM and found it ran more than 20,000 times slower than SQLite for a simple query. The code compiled, passed tests, and looked architecturally sound, yet a flawed query planner caused every lookup to become a full table scan.

-> AI often produces plausible code, not necessarily correct systems. Developers still need to deeply understand the architecture they build.

Meta is reportedly experimenting with a new AI powered editor called Vibes, designed to compete with emerging AI writing and editing tools.

-> Big tech is targeting the productivity layer where AI can become a co‑pilot for everyday creative and knowledge workflows.

A new chip design approach that prioritizes SRAM instead of traditional DRAM could significantly reduce latency when running AI models.

A Google product manager has open sourced an agent capable of persistent memory without relying on a vector database. This could simplify how developers build AI agents compared with the typical RAG heavy stack used today.

One common mistake when working with AI is writing extremely long prompts. Many people write hundreds of words describing tone and style, yet the output still feels obviously AI generated.

A simpler trick shared by Ruben Hassid is to create a file that lists common patterns of AI generated writing and instruct the model to avoid them.

The prompt is only 29 words:

Here’s how to set it up:

→ Instead of writing longer prompts, give the model a style filter so it learns what patterns to avoid.

They all run their businesses on beehiiv. Newsletters, websites, digital products, and more. beehiiv is the only platform you need to take your content business to the next level.

🚨Limited time offer: Get 30% off your first 3 months on beehiiv. Just use code PLATFORM30 at checkout.

Start building for 30% off today.

If you find yourself repeating the same instructions to Claude, you can package those workflows into something called a "Skill". A skill is essentially a small folder that contains instructions and optional scripts that teach Claude how to perform a specific task consistently.

Skills work best for repeatable workflows such as research processes, document creation, or multi‑step automations. Instead of explaining the process every time, you define it once and Claude can reuse it across conversations.

Here’s how to try it:

→ Think of MCP as the tools in the kitchen. Skills are the recipes that tell the AI how to actually cook

A16z argues that AI will likely disrupt much of today’s application software layer. But that does not necessarily mean SaaS disappears.

Instead, the real moat of software may shift from interface design to data, workflow integration, and the ecosystem around the product.

Read more.

👉 Forward it to someone trying to keep up with AI.

👉 Read online: techzip.beehiiv.com

Tech.zip Newsletter
| Zipping what truly matters in the AI era.