Vibe coding—the art of guiding LLMs to write code on your behalf, according to stated requirements—is shifting from a quirky, experimental approach to something that, in many cases, is expected even in large, bureaucratic organizations. It can be a huge productivity booster, but it also comes with real risks: even a single misplaced character can lead to SQL injection, bypass logins, or expose sensitive user data.

As AI tools have become more widespread, and generally better, I’ve seen the rift between those who incorporate these tools into their development process and those who don’t, widen.

Large Language Models (LLMs) are transforming the way we build software. Vibe Coding — popularized by tools like Cursor and Windsurf — refers to deeply embedding LLMs into the development workflow. Like all tools, however, it takes time to learn how to use them effectively.

Most developers start their journey with language models by prompting them: “Explain this code”, “Summarize this thread”, “Write a SQL query”. But eventually, you hit a wall. You want your model to do something dynamic — like call an API, look something up, or trigger a workflow. This is not data that you can provide in advance, because it is constantly changing, and yet LLMs are trained on a static dataset.

Companies are rushing to integrate AI into their customer support workflows, generating a lot of interest in the latest models. However, the power of the model is only truly unlocked if the if system architecture around it is right — how you deliver context, handle private knowledge, and adapt over time are all key factors in building a successful AI customer support system.

Every organization today is drowning in unstructured data: support messages, internal chats, user feedback, call transcripts. As our ability to collect more unstructured data within organizations grows, the need to understand that data in a scalable, cost-effective way is outpacing our ability to handle it.

François Chollet recently introduced a new version of the ARC AGI benchmark during his talk at Y Combinator’s Startup School. It’s one of the more thoughtful discussions on evaluating AI systems — and well worth watching.

As smart pet devices become more common, many of us have come to depend on them without a second thought. Unfortunately, that reliance may be premature, as we’re beginning to understand their limitations over time.