My least favorite part of GPT-5 as a developer API is the smart routing feature. The indeterminism of when it will route to "thinking" vs. "not thinking" makes it very unreliable when you are using it as part of infra. Am I missing a toggle/setting here?

Claude's plan mode (shift+tab) with opus 4.1 feels like god mode in vibe coding. Probably the biggest unlock for me since I first used it in research preview.

RIP Charlie Kirk. No one should be murdered in cold blood for speaking their mind in public, regardless of disagreement. Shocked to see some people justifying this tragic event. There is no sliver of reason here. Heart goes out to his family and friends.

Coding agents are bad at writing end-to-end tests because they can't see. Even when they successfully write them, the chosen selectors are brittle, leading to flaky tests that validate "how", not "what". Watch Playmatic self-heal tests & help Claude Code one shot tests: 0:23

how to build a billion dollar startup

Midnight flow is intoxicating. Love the feeling of being locked in on a weekend night.

We just built an AI agent that QAs your app in the background. Tag it with what you want to test in Slack, and it will: - come up with a plan - write & run playwright tests Working tests are committed in a PR, and bugs get reported back. LMK if you want to try it out!

How to Keep Winning amasad.me/keep-winning

luna Been in the startup world for a couple decades (lived in SF 15 years ago) and I have some bad news for you... my company never endorsed this and we did fine but even early social media days there were people glorifying this. Bleh.

x.com/i/article/2004…

Software is getting commoditized every day. It won’t be far when everyone will have niche apps for their own use cases. We will probably see the barbellification of software where you have heavy enterprise systems of records on one end and a plethora of personal apps on the

When we first launched Kernel, our paid plans *started* at $200 / mo. We wanted to ensure that all of our early customers had: (1) non-zero budgets and desire to pay for what we were building (2) production-level use cases so we could invest in the foundations to be successful in

Week 1 OpenAI has been incredible. The ambition and velocity here is unmatched & it truly feels like the sky is the limit. A small call for feedback: If you use smart browsers extensively in any part of your workflow, LMK, I’d love to chat with you!

We're back to shipping post holidays with a ChatGPT Atlas release last week, and another one today. - Tab Groups (they're slick and you can use emojis) - Updated search results layout (vertical stacked links) - "Auto" option for your default search engine to decide between

Being wrong is forgivable. Being slow is not. Years later, the primacy of speed is still the best lesson from FB.