AI benchmarks gamed by exploits & iPhone passcode broken by update - Hacker News (Apr 12, 2026)

First up: a group at UC Berkeley says several widely used AI agent benchmarks can be “reward-hacked” to score near the top without actually doing the intended work. Their point isn’t that researchers are dumb—it’s that many eval setups accidentally leak answers, blur the boundary between the agent and the grader, or rely on brittle validation. That matters because benchmark numbers drive everything from model selection to funding to safety narratives. If scores can be gamed, the incentives drift toward manipulating measurement instead of improving real capability, and the public story about progress gets distorted.

Staying with software that behaves differently than expected: a student says an iOS update locked him out of his iPhone because the lock-screen passcode keyboard stopped accepting a specific Czech character. The key still appears, but the phone won’t actually input it during the “before first unlock” passcode entry. And because he didn’t have a cloud backup, the official recovery path—restore the device—means losing the photos and data he cares about most. The broader lesson is uncomfortable: security features like strong encryption make recovery genuinely hard, so small input-method changes can turn into catastrophic access failures for anyone using uncommon characters to strengthen passcodes.

On the developer-ops side, Chris Whocodes published a refreshed “VM Options Explorer” for OpenJDK 11 HotSpot—a searchable, normalized catalog of JVM flags with context like defaults, deprecations, and where each option lives in the code. The interesting part isn’t just that there are a lot of knobs; it’s that the page helps you see how flags evolve across JDK releases and across vendor builds. If you operate JVM services, this is exactly the kind of detail that can prevent a painful upgrade, where an old tuning flag suddenly turns into a warning—or worse, a startup failure—and you’re left wondering what changed and when.

Now to computing fundamentals: a readable explainer revisits Landauer’s principle—the idea that erasing information has an unavoidable energy cost—and contrasts it with reversible computation, which in theory can avoid that particular penalty. Even though modern hardware burns far more energy than the theoretical minimum, the argument is that “reversible” thinking can still guide practical efficiency gains. The piece also highlights the tradeoff: you often need extra scratch space and additional outputs to keep computations reversible. Why this matters right now is simple: as compute demand keeps climbing, energy efficiency is turning from a nice-to-have into a core constraint.

Speaking of constraints, former Intel CEO Pat Gelsinger is now backing hard-tech startups and used a recent interview to lay out where he thinks the next computing bottlenecks are forming. He’s betting on a heterogeneous future—systems mixing classic CPUs with AI accelerators and, eventually, quantum components—while warning that today’s AI growth is running into very real limits around memory, interconnects, and cluster reliability. He also frames energy supply as a strategic resource, not just an operating cost, and ties it to geopolitics and supply-chain resilience. Whether you agree with every prediction or not, it’s a useful map of where an industry veteran expects money, engineering talent, and policy attention to converge.

A lighter read with a serious takeaway: Alex Miller’s “Miller Principle” claims, bluntly, that no one reads anything—docs, UI text, long emails, even code comments. It’s tongue-in-cheek, but the product lesson is real: if your system only works when users carefully absorb instructions, it probably won’t work. Good design assumes skimming, distraction, and time pressure—and tries to make the correct action the easiest action.

On building sustainably, a developer wrote about getting turned down at a pitch night because investors couldn’t see why funding was needed—his products already make recurring revenue with very low infrastructure spend. The essay’s broader theme is anti-glamour: keep deployments simple, keep costs predictable, and avoid architectural choices that drag you into constant operational overhead. The reason this resonates is that it’s less about any single tool and more about a strategy: reducing burn rate buys you optionality—more time to learn what customers actually want, and less pressure to chase growth-at-all-costs.

Two culture-and-ideas stories to close. First, Stewart Brand argues in a new project that maintenance—repair, calibration, and upkeep—isn’t a footnote to progress; it’s one of the engines of progress. His through-line connects precision manufacturing, interchangeable parts, and the institutions that preserve practical know-how. It’s a useful reframing for tech culture, which often celebrates invention while undervaluing the steady work that keeps systems safe, reliable, and improvable.

And finally, a blog post proposes an informal, chronological list of world-changing intellectual breakthroughs—using Claude Shannon’s information theory as an example of a foundational idea that most people benefit from without ever hearing about. The author’s real goal seems to be sparking debate about what counts as a true intellectual revolution, and what gets left out of the usual canon. In a moment when AI and computing dominate headlines, it’s a reminder that today’s “obvious” technologies often rest on yesterday’s quiet, abstract insights.