@vfarcic.bsky.social
@vfarcic.bsky.socialPen testing happens once a year (if at all), costs a fortune, and the report arrives after you've shipped 12 more releases. What if an AI did it on every release for pennies? 👇 https://youtu.be/dJv_f4w8ZAA
@vfarcic.bsky.socialHeroku shipped your entire platform team's roadmap in 2011. Finished. Priced. Documented. You can still sign up tonight in under a minute. Your platform team is going to build it anyway.
🧵 New video:
@vfarcic.bsky.socialAutoscaling LLM inference isn't reacting to traffic; it's forecasting it. Your replica takes 10+ min to arrive. By then, the spike is history. https://youtu.be/wwrPwU-JCMw
@vfarcic.bsky.socialStack Overflow taught a generation to copy, paste, and take the credit. Now it's fading. The story of how it rose, why it worked, and what killed it. New video 👇
@vfarcic.bsky.socialInference is a memory management problem, not a compute one. Your GPU almost never runs out of arithmetic first. It runs out of room to hold conversations. And your engine already tells you the exact ceiling at startup — before a single request arrives. 🧵 https://youtu.be/g_5g1hBmAzA
@vfarcic.bsky.socialJenkins never said no. Not once. That's why it's still running in your company right now, and why nobody can switch it off. The nightmare isn't that it died. It's that it didn't. 🧵🕯️
@vfarcic.bsky.socialContinuous batching. Paged attention. Prefix caching. Speculative decoding. 12 inference ideas, a couple of minutes each, one pass. 🧵👇 https://youtu.be/MVo3P-D7eSc