The AI bill came due, Uber torched its budget in 4 months, and the Pope wrote 42,000 words on AI

Why pay tokens to re-send style rules the model forgets anyway? Andrey Sitnik's team moves them into a pre-commit hook with nano-staged and lets a dumb script enforce them for free. Old rule, new bill: if a script can do it, keep it out of the docs.

New flagship. My favorite detail is in the footnotes, where they admit a competitor's cheaper model beats Opus on one finance eval. Too sad nobody reads footnotes. Nobody re-runs the "trust me" numbers either.

Google's best Flash yet, built for agents and coding, soon the default everywhere. Also roughly 3x the price of the last Flash and creeping toward Pro. In the good old days "Flash" used to mean cheap and fast. Now it just means fast.

Nolan Lawson pushes back on the idea that LLMs are only good as slop cannons for unread PRs. His move is to point the models at the code instead, running an ensemble (Claude, Codex, Bugbot) over each PR so three reviewers converge on real bugs and the false positives drop away. Velocity doesn't go up, it goes down, because the process keeps surfacing pre-existing bugs you now have to fix. That's the point.

GeoHot's six-month verdict: agents front-load progress, never finish, and big orgs will quietly drown in confident broken code. It's like handing an over-eager intern named George your search algorithm. Fast, fearless, and you find out what he did much later.

Callstack trained a domain-specific coding model for React Native on top of Gemma 4, frontier-ish results for mobile work at a fraction of the cost, because the framework knowledge lives in the weights instead of in tool calls and context.

Languages with strong, boring conventions get better agent output. Go beats Rust and C here because there's mostly one way to do things, so the model guesses right more often. When you're gambling tokens on every run, bet on predictable.

The new garbage collector in 3.14.0–3.14.4 can blow up memory and trigger OOM kills in the cloud. Wait for 3.14.5+ or skip to 3.15. Consider yourself warned.

CISA, the agency whose entire job is telling everyone not to do this, left AWS GovCloud admin keys and a plaintext-password CSV in a public repo since November. The kicker: someone manually switched off GitHub's secret scanning to pull it off. Some grandma's password spreadsheet is more secure. At least she never pushed it.

Aurora Scharff walks the arc from useEffect to React Query to loaders to RSCs. Every couple of years React rediscovers that prop-drilling data from the top hurts, then ships a new fix. This one composes nicely, at least.

Our friend Dominik Dorfmeister on getting Router and Query to stop fighting over who owns the loading state. Saves you the trial and error.

Recovered from paper printouts and OCR'd by hand. Best part: it started life as "QDOS", the Quick and Dirty Operating System, before running the entire PC era. The original careless MVP, shipped to a few hundred million machines.

Microsoft is reportedly pulling most direct Claude Code licenses and pushing people to Copilot CLI – six months after it got too popular to afford. Uber burned its whole 2026 AI coding budget in four months at $500–$2,000 per engineer, and its COO admits he can't connect the spend to anything shippable. That last part is the real story: everyone measures tokens, and tokens tell you nothing about whether you're actually shipping.

Not to brag, but Roman spent his whole DEVWorld Amsterdam talk on exactly this: why token count is the wrong metric, and how to measure AI adoption across people and outcomes instead. https://ralabs.org/blog/devworld-amsterdam-measuring-ai-adoption/.

Anthropic is rumored to be heading for its first profitable quarter while companies keep getting shocked by their LLM bills. Simon Willison's take, which I share: those are the same fact. Surprise bills are what PMF looks like when you sell intelligence by the token.

We've officially reached "the Pope has an AI take." 42,000 words on protecting the human person in the age of AI, arguing for slower, more deliberate development. One line lands regardless of faith: you can speed up execution, but responsibility doesn't move that fast.

How a Japanese toilet maker became a major supplier of a semiconductor part for AI data centers. The thesis: the Japanese firm exists to keep existing, not to please shareholders, so it reinvents itself endlessly: Nintendo from playing cards to consoles, Fujifilm from film to chemicals. In a year when every company is "pivoting to AI," the long-game version is worth understanding.

Armin Ronacher wants "clanker" instead of "agent", it keeps the thing a tool, not a colleague or a little spirit in your terminal. His real point: "the agent decided" is how responsibility quietly disappears. He's right. He's also not getting invited to the cookout after the robot uprising.

For the people who copy a question out of Slack, paste it into AI, and paste the answer straight back as if they wrote it. You know who you are.

If you've ever heard a confident "coding is dead" take and felt a flicker of unease, I built a small project that lays every one of these predictions since 1952 against what actually happened. The obituaries have a bad track record. Fun detail: the timeline was generated with Google AI Studio.

Is it nepotism to tell everyone to watch my wife's first webinar? Probably. I'm doing it anyway, because the conversation is good on its own merits. Stasia and Stephan Moerman get into when a vibe-coded prototype becomes a product and when it becomes a trap. Best line: the moment nobody can explain why the code works, you've stopped paying technical debt and started paying technical interest – and the rate climbs fast.

A friend of mine is working at a startup where the main productivity metric is number of PRs. So developers are PRing like crazy: out of priority, features nobody asked for, UI changing daily, clients confused. No quality metrics, only quantity. The old coding bottleneck was doing real work: it forced teams to prioritize, talk to actual clients, and make sure what they shipped served a purpose. Remove the bottleneck without replacing the filter, and you get volume without value.

We run strategy sessions on a schedule: every six months, not only when something's on fire. To lock the projects we'll carry for the next year. Two of this round's bets are AI, and one, against the current of everything above, has nothing to do with AI at all. We ran the session in Monaco, swam in a not-yet-warm sea, and over-toured a few museums. Lesson learned: renting a villa up in the mountains looks great until you're driving down to Menton for breakfast every morning and losing two hours a day to it.

I picked this one up for something I'm building – more on that another time. If you've ever wanted to understand how radio and radar actually work, down to the level where you could build one yourself, this is the reference to start with.

We continue to grow and regularly open new roles across engineering and delivery. You will find the current list of open positions above.

https://ralabs.org/hiring/