The hardest part of building an AI app isn’t launch day. It’s day 180, when the hype’s gone, bugs pile up, and no one’s watching.

Everyone’s debating “agents vs workflows”, here’s the truth: they’re not competitors. They’re complements in different layers of abstraction. 1. Agents ≠ Workflows. An agent is an autonomous entity that decides what to do next. A workflow is a deterministic blueprint that

As society we need to channelize "Us and tomorrow" instead of "me and now"

When Android was brand new, I spent my weekends in a dorm room building an app called "Toronto Now". It was simple, just local news, weather, and tech updates for my city. But it hit 10K downloads, made solid ad revenue, and helped pay my tuition. It was my first real builder

It’s easy to frame this as Amazon bad, Perplexity good. But this moment is deeper than that. Amazon’s trying to protect its ecosystem. Perplexity’s trying to empower users. Both are right in different ways, but this fight shows how quickly we’ve moved into a new layer of the

The best use cases I have found so far using comet👇 Competitor analysis: open 3–4 sites, ask it to compare features, pricing, and messaging, export full report to Google Docs. Research agent: groups all open tabs into a workspace and summarizes key insights and common

A lot of people are panicking that “OpenAI will just build my app.” But their strategy is clear: Enable billions of use cases rather than owning them. The opportunity is in the last-mile: deep workflows, distribution, and solving a painful, boring problem better than anyone else.

So Google just forked the Windsurf codebase and they even forgot to remove the Cascade branding in some places

I’m running Gemini 3 against my internal evaluations and the early signal is strong: Better reasoning, better planning, fewer derails. Not perfect, but definitely “builder-ready.” Feels like Google finally shipped something with teeth.

Everyone’s talking benchmarks, but here's where I think Gemini 3 shines It handles multi-file reasoning, UI generation, and debugging with far fewer interventions than any model I’ve used lately.

let's take a moment of gratitude for how useful & incredible today's LLMs really are it's easy to take for granted just how much more productive these models make us big respect to anyone working on & inventing cutting-edge technology. this is the closest thing to magic we

The best analyses of the AI technology ever done, TLDR 🧵 :