i just spent the summer sleeping in the tenderloin. now sleeping in dogpatch. thank you @ycombinator and @amiklas for the upgrade. proof attached ↓

elite event met tons of goated founders

cheffing up something new @runmirrors at the yc office thinking of calling it dynamic context editing

sf is absolutely insane

turned 19 today first birthday where I'm actually happy with what I'm doing with my life. spent it with my family, which is all I wanted most of my friends are off at college and I'm not. feels weird but I wouldn't trade it grateful for another year 🙏

🪞🪞🪞

🪞🪞🪞

yc takes two technical co-founders and quietly turns one of them into a salesperson. happened to me. during the batch, i remember thinking there's no way i'm going to do sales. now it might be my favorite part of the job. if i had a kid going to college right now, i'd suggest they study computer science, then go into sales anyway. being technical and being able to sell is a rare combination, and it compounds like crazy.

these guys are building some pretty cool boxes

hosting purdue's first startup party! luma.com/2rlchpe2

I'm leaving college to build a startup focusing on robotics for hospitals. I'm still 18, and I just got backed by @speedrun. I can't believe how much has changed since I graduated from @TJHSST_Official. A year at @RiceUniversity studying Bioengineering on the pre-med track, research at the Rice Neuroengineering Initiative, work at Motif Neurotech, and time well spent with some of the best people I've met. I wanted to be a doctor my whole life. But it's time for a change. There's no one else I'd rather be joined by than the two smartest people I know, @adibalaguru and @rohith_yel. We've been best friends since middle school and have built things together ever since. Three teenagers, from one seven-year-old group chat. We're beyond grateful for our incredible partner, @JoshLu, for believing in our vision, and the endless support from @andrewchen, @Tocelot, @_CallMeMacy, @tabishgilani, @fromevanwang, @kenanhsaleh, @tkexpress11, @Chen, @tmhammer, @ndrewlee, @SamiraBehrouzan, @TheRioDeGennaro, and the rest of the speedrun team. If you're in hospital administration or work at a robotics company, we'd love to talk. Big things to come!

Your team is spending more on AI than you can explain. Today, we’re launching @Mentlio (YC S26). We turn AI coding spend into a measurable unit: cost per PR, reuse, and downstream quality, benchmarked against industry. We also cut waste without touching your prompts, via cache hits, less context bloat, and auto-routing. On Terminal-Bench 2.1, our local router kept 98.93% of Fable 5's accuracy at 24.9% lower cost on Claude Code. If you’re spending real money on tokens, see your real AI ROI: mentlio.com/intro

error analysis is super relevant and most teams we talk to actually still have that as a main blocker, but reproducing it after the world state changes is still underrated

Off Season II at @fdotinc last Friday: ~100 teams, six weeks of building. Our founder @SidraMiconi spent Friday night filming what people built during six weeks at @fdotinc. 10 demos from a room of ~100 teams. The full cuts are coming! @ai_singhal → @runmirrors @pedro_nagamine + @bellujrb → @usebido @autobrik + @alvax64 → @robonet_ @HeapSmasher → @reachpad @hasnainbaig06 + @Jahenderson05 → @mirdainsystems @001TMF + @MukherjeeArvaan → Proteus-Bio

We just hit SOTA on the standard physics-ML benchmark (Navier-Stokes 2D turbulence): Edviro research is building world models that understand and operate energy infrastructure. Buildings are our first deployment environment. We started with the NS2D benchmark because fluid dynamics underpins many energy systems, including airflow and industrial processes. The goal was to rethink fundamental architectural choices. Instead of asking a neural network to learn the entire PDE update, we analytically solve the linear component and train the network only on the nonlinear increment. The results: we hit 0.0305 (± 0.000079) across three independent seeds. Applying symmetry-averaged inference further reduces error to 0.02969 at 2x inference costs. Against code released claims, we beat them by 36.9%. If you’re working on operator learning, physics ML, controls, or operate large-scale energy systems, we’d love to talk.

your AI agent will do something expensive eventually so we built a staging environment for AI agents here's 90 seconds on it ↓