Sam Feldt

My neighbours are hearing this non-stop right now😂 Click the link to stream KEEP ON RISING🎶 link.to/_KeepOnRising

Goal for July: Make more obscure things nobody asked for. Watching my son grow up, I’ve realized just how much kids build things 'just because'. And how little we do it as adults. Most times, if there’s no clear ROI for a project, we don’t bother. But experimentation is how we’ve gotten some of the world's biggest solutions. Google Street View. Twitch. Gmail. The list goes on. The real question isn’t "why waste time building projects without a clear path forward?" It's "why did we ever stop?"

Weekends like this >> #djlife #weekends #restday

Played 200+ shows before I wrote a single line of production code. Everyone assumed the DJ couldn't build the AI. But reading a crowd and reading a user are the same muscle — sensing what lands a half-second before the data confirms it. The stage was my training set.

example from healify: scared user asks about an abnormal bloodwork result. single LLM call: technically correct answer, clinically flat tone, user leaves forever. orchestrated agent: "this is unusual, here's what i can say with confidence, here's where to take this next, want me to flag it for a clinician." user stays. same model. different system. different product.

the reason teams ship orchestrators isn't accuracy. it's that orchestrators let you build *judgment* into the system — escalation paths, tone branches, "i'm not sure, let me check" responses. these are the moments that make or break trust. a single-call LLM can't do them.

The framing is backwards. The problem isn't that Claude Code is too good — it's that big orgs reason about AI cost at the CFO level, not the call-site level. I run a multi-product shop solo-ish. A single mis-configured EventBridge cron firing a metered LLM call billed me $6.9K/mo of Anthropic credits for two days before I caught it. The fix wasn't a budget meeting — it was a process-global rate-floor + a per-service daily alarm that pages in hours, not days. Uber doesn't have that at the call site. Most don't.

Uber's CTO told @LauraBratton5 that AI coding tools—particularly Anthropic’s Claude Code—has already maxed out its 2026 AI budget 📈 “I'm back to the drawing board, because the budget I thought I would need is blown away already,” Neppalli Naga said. theinformation.com/newsletters/ap…

everyone's debating model benchmarks. operators are asking one question: does it ship faster than the last version? i've been building with claude code daily. the thing that changes the game isn't the evals — it's the edit loop. how fast can you go from "this is broken" to "this is working" without leaving the conversation. that feedback loop is where AI actually compounds. not the benchmark number.

Signed artist who also builds AI infra, so the Spotify/UMG AI-covers deal hits both ways. The real tell: independent artists are locked out because they don't have the rights-management layer the majors do. The breakthrough here was never the model — it's the consent + attribution + royalty pipeline underneath it. That layer is buildable without UMG. Whoever ships it for independent artists owns the next wave.

Enterprises blow their AI budget because they put the guardrail in the finance review, not the code. I learned this the hard way, solo: a single mis-configured cron on multi-leader infra leaked $6.9K/mo of LLM calls for two days before I caught it. The fix wasn't a budget meeting. It was a process-global rate-floor at the call site plus a per-service alarm that pages in hours, not days. Token discipline is an infra problem, not a spreadsheet problem.

What failed in that report is AI bolted onto an engagement chatbot. That architecture optimizes for session length — exactly the wrong objective for a teen in distress. Building health AI the other way (which I do) means the agent's job is sometimes to end the session and escalate to a human. Different objective, different eval, different conscience. The category isn't the problem. The incentive design is.

10 years ago i was playing to 300 people in clubs in the netherlands. 10 years later: 10+ billion streams, mainstage at the world's biggest festivals — and a health app i built myself with AI. the music career gave me something i didn't expect: obsessive product sense. every show is a product launch. setlist = feature roadmap. the crowd's energy = real-time analytics. you learn to read what people actually want vs what they say they want. turns out that's exactly what building software requires.

karpathy moving to pretraining at anthropic = a signal. the model layer just became the new artist's canvas. founders shipping AI products in 2026 won't win on model access. they'll win on knowing what to point the model at. same shift the music industry went through in 2008 — when access to the tools got free, the moat moved to taste.

my entire stack for building Healify: — claude code (the actual builder) — cursor (for when i want to stay in the code) — expo + eas (ship the mobile app) — aws agentcore (the AI agents under the hood) that's it. no devops team. no platform engineer. one person, four tools, a live app in the App Store. vibe-coding isn't a shortcut. it's removing the gatekeeping between an idea and a real thing.

what i'm actually building on Healify right now: — AI that reads your blood panel and explains every marker in plain language — it cross-references your wearable + HealthKit data, not one snapshot — a "view source" on every recommendation so you see the study behind the advice — predictive insights that flag a trend before it becomes a problem every feature starts from a real frustration — mine, or someone who messaged me after a show. building in public from here.

the lab founders are training models. the artist-founders spent 10 years training audiences. when the model finally ships, it still has to go find users. the artist already has them.

If you're a creative sitting on a product idea, telling yourself you're not technical enough — that excuse expired in 2024. I'm building in public here. Follow along and I'll keep sharing what actually works.

Building on AI agents made one thing obvious: the models are a commodity. taste isn't. knowing what to build for a scared human > knowing how to wire 5 LLMs together. the next great AI founders will be creatives, not engineers.

I did 10 billion streams before I wrote my first line of code. Now I'm building an AI health platform. Here's what nobody tells you about going from creative → founder 🧵