AI goal mode benchmark & GoPro searches for options - Hacker News (Jul 18, 2026)

What would you do if your cloud dashboard suddenly claimed you owed hundreds of millions of dollars? That happened to some AWS users, and it is one of those stories that gets attention fast because billing is the one place nobody wants surprises. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is July 18th, 2026. In today's roundup: an AI benchmark with a surprising lesson about persistence, new signs that GoPro may be running out of room, a Windows hardware install controversy, some beautifully nerdy posts on JPEGs and semiconductor measurements, and a quick look at both retro computing history and Stack Overflow's early explosion.

AI goal mode benchmark

Let's start with AI. A new test pitted Claude Fable 5 against GPT-5.6 Sol on an unpublished NP-hard network optimization problem, so this was not the kind of benchmark either model could have memorized. Fable 5 came out ahead overall, and the most interesting part was not just the win, but the consistency. Its plain mode delivered the safest and strongest average performance, while Sol was more erratic and generally weaker. The other takeaway is a useful reality check on so-called persistence features. Native /goal modes did win more individual runs, but they were not a universal upgrade. On hard search problems, persistence can help a model stay on a promising path, but it can also make it cling to a bad one for longer. That matters because a lot of AI tooling is now being sold around autonomy and staying power, when in practice the real question is whether the system knows when to change course.

GoPro searches for options

Next, GoPro appears to be in genuine trouble. Reports say founder Nicholas Woodman has lent the company 20 million dollars as a temporary bridge while a buyer is being sought. That follows a rough first quarter, with revenue down 26 percent year over year and camera unit sales down 29 percent. GoPro has tried to branch out beyond its core action camera business, including higher-end gear and even aerospace and defense efforts, but the pressure from rivals like Insta360 has not eased. Add layoffs and a lost patent fight, and the picture is hard to ignore. This matters because GoPro was once one of the defining gadget brands of its era. Now it looks like a case study in how difficult it is to stay distinctive once a hardware category matures and competitors catch up.

Windows trust and billing scares

A pair of stories today also touched the same nerve: trust. First, some LG monitor owners found that simply connecting certain displays to a Windows PC could trigger automatic software installation through Windows Update, with no clear consent prompt. In testing, that included LG-related installer packages and even repeated McAfee promotions. It is the kind of behavior users really dislike because a monitor is supposed to be a display, not a delivery channel for surprise software. Then there was AWS, which had a billing incident that showed some customers estimated charges in the hundreds of millions, and in some cases even billions, of dollars. AWS says the issue affected estimates, not actual charges, but that distinction does not help much when a customer opens the console and sees a number that looks catastrophic. Put together, these stories are reminders that user trust is built on predictable behavior. Whether it is hardware installing apps or a cloud dashboard inventing a giant bill, the damage comes from the same place: systems acting in ways people did not expect.

JPEG tricks and real electronics

There were also two great technical posts that show why engineers keep poking at old assumptions. One explored a quirk in progressive JPEGs. Because those images arrive in multiple scans, a cleverly crafted file can appear to switch between pictures as it loads, and with a few more tricks, it can even fake a tiny animation or video-like sequence inside a single image. It is playful, but it also shows how much behavior can hide inside formats most of us think are simple. The other post was about measuring real voltage-current curves for semiconductor devices. Using careful instrumentation and pulsed testing, the author captured diode and MOSFET behavior in a way that makes the textbook versions look a bit too clean. The broader point in both cases is the same: real systems are messier and more interesting than the diagrams we use to explain them, and that gap is often where the best learning happens.

Retro computing, old and new

On the retro side, one story looked back at Rekursiv, a remarkable custom computer from the 1980s built by Scottish hi-fi company Linn. It enforced memory safety in hardware, handled garbage collection in silicon, and treated storage more like one persistent object space than separate memory and disk. In other words, it was chasing ideas that still feel modern. It failed commercially, largely because general-purpose chips improved too quickly, but many of its concepts are resurfacing now in capability hardware, memory tagging, and domain-specific design. Right next to that was a smaller, more personal revival: someone brought a 2009 ASUS Eee PC back to life with Arch Linux 32. It worked, but only up to a point. Even after a RAM upgrade, the old Atom processor and hard drive remained major bottlenecks. Together, these two posts make a nice contrast. Sometimes old ideas were simply early. Sometimes old hardware is just old.

Stack Overflow's breakout years

And finally, a quick data point from Stack Overflow's early history. A query tracking monthly question volume shows just how quickly the site scaled after launch. It started with only a handful of questions in mid-2008, climbed into the tens of thousands per month not long after, and peaked in the visible data at just over 100,000 questions in March 2011. That kind of growth helps explain why Stack Overflow became such a defining part of programming culture. It was not just useful; it reached critical mass at remarkable speed, and once it did, it became the default place where developers compared notes, solved problems, and occasionally argued about semicolons.

That's it for today. If one theme tied this episode together, it was that tools and systems are only as good as the trust and judgment behind them. Thanks for listening, and remember: links to all stories can be found in the episode notes.

AI goal mode benchmark & GoPro searches for options - Hacker News (Jul 18, 2026)

Our Sponsors

Today's Hacker News Topics