Uber COO: "Harder to Justify" AI Token Spend, The Tokenmaxxing Backlash Begins

Hey everyone, welcome to the Builder's Briefing for May twenty-sixth, twenty twenty-six. I'm Alex.

And I'm Sam. We've got a packed one today — the big theme is cost discipline meeting quality discipline in AI, and honestly it's about time.

Absolutely. So let's jump right into our big story. Uber's COO Andrew Macdonald came out and said publicly that it's getting, quote, harder to justify the money they're spending on AI token consumption. This is not some indie hacker complaining on Twitter — this is the COO of a hundred-and-seventy-billion-dollar company.

Right, and what's wild is the Hacker News discussion had over a hundred and twenty comments, and it's full of teams saying the exact same thing. They scaled up agentic workflows, watched their inference bills balloon, and didn't see proportional business impact. There's even a term for it now — tokenmaxxing.

Tokenmaxxing — just throwing more tokens at problems hoping the quality improves. And the takeaway for builders is pretty clear: your unit economics around inference costs need to be a first-class design constraint, not an afterthought.

Yeah, things like prompt caching, intelligent routing between small and large models, aggressive output length constraints — these aren't nice-to-have optimizations anymore. They're requirements. And honestly, if you're on the vendor side, building a Datadog-for-inference-costs tool? That's suddenly a very hot product category.

And this connects directly to our next story. George Hotz — geohot — published an essay called Eternal Sloptember that's also trending. He argues we've entered a permanent September where AI-generated slop just floods every channel.

That's interesting because it's really two sides of the same coin, right? Uber's saying more tokens doesn't equal more value, and geohot's saying more AI output doesn't equal better output. If you're building anything content-dependent — search, aggregation, social — you need a quality signal layer or your product just becomes a slop pipe.

Also in the AI bucket — the Vatican's AI encyclical, Magnifica Humanitas from Pope Leo the Fourteenth, hit nearly eight hundred points on Hacker News. Highest engagement of the day.

I did not have 'Pope tops Hacker News' on my bingo card, but regardless of your views on the Vatican, if you ship AI products in Europe, you should at least skim it. That moral framing is going to show up in regulatory conversations.

Good call. Alright, shifting over to dev tools — there's a comprehensive Go-to-Rust migration guide from Corrode that's getting massive traction. Over two hundred points, two hundred comments.

This one is genuinely practical. It covers concurrency pattern translation, error handling, incremental migration via FFI — it's not just a 'Rust is better' argument. If you've got Go services hitting performance ceilings or you need stronger memory safety guarantees, this is the best resource I've seen for actually making that move.

And I love that it's trending alongside the AI cost stories, because it tells you builders are rethinking runtime costs at the systems level too. Not just inference costs — total cost of compute.

Oh, and the fun one — someone proved Jira's workflow engine is Turing-complete. Which is hilarious, but the real takeaway is: if your project management tool is Turing-complete, your workflow configs are probably way more complex than they need to be.

That explains so much about every Jira instance I've ever used. Okay, on the launches side — Audiomass is a Show HN with over three hundred points. It's a full multitrack audio editor running entirely in the browser, no backend needed.

That's really impressive. If you're building audio features into a web app, it's both a usable tool and a reference architecture for the WebAudio API. And speaking of browsers — Firefox is now integrating with Adafruit for in-browser hardware project development. Browser-as-IDE for hardware is actually getting real.

And a quick infrastructure note — Microsoft killed a two-hundred-and-forty-four-acre data center project after community pushback. This is becoming a real pattern. If you're relying on hyperscaler capacity expansion, community resistance is now a genuine constraint on available compute.

Yeah, plan your capacity needs accordingly, especially in the US Midwest where a lot of this opposition is growing.

On security — the Netherlands seized eight hundred servers and arrested two people running bulletproof hosting infrastructure. Major takedown reported by Krebs.

And related — CrowdSec is trending on GitHub again. It's an open-source, community-driven alternative to commercial WAFs. Basically a drop-in replacement for fail2ban but with a much broader crowdsourced threat intelligence layer. Worth checking out if you haven't already.

Quick hits before we wrap up. There's a developer essay called Leave Me Behind about tech industry burnout that resonated hard — nearly three hundred points. A deep dive on the Gnutella protocol with P2P architecture lessons still relevant today. And my personal favorite — a resurfaced paper showing didgeridoo playing as a treatment for sleep apnea.

I mean, that's the kind of content that keeps me coming back to Hacker News. Also, PhotoPrism is trending on GitHub — self-hosted, AI-powered photo management for privacy-conscious folks. Link in the briefing for all of these.

So here's your takeaway for the week. Today's through-line is clear: throwing more compute at AI doesn't automatically create more value. The next wave of winners won't be who uses the most AI — it'll be who uses it most efficiently.

Invest this week in inference cost monitoring, model routing — send simple queries to small models — and output quality measurement. That's where the edge is going to be.

That's the Builder's Briefing for May twenty-sixth. All links are in the show notes. Thanks for listening, everyone.

See you tomorrow. Go build something efficient.