Lightbulbs, Rowhammer, and the Hidden Costs of Making Things Last

The physics of tungsten filament bulbs offers one of engineering's cleanest illustrations of inescapable tradeoff: running a filament hotter makes it brighter and more energy-efficient but shortens its life; running it cooler extends lifespan but shifts the emission spectrum toward infrared, producing more heat and less visible light per watt. 'Increasing the Lifespan of a Bulb Makes It Worse in Every Other Way' earned 98 points and 103 comments, with the thread extending the observation into a broader principle — longer-lasting tires are harder and grip less, longer-lasting consumer batteries often demand heavier or larger form factors — before arriving at a discussion of planned obsolescence.

That dark undercurrent connects to 'Ten Ways NAS Is Getting Enshittified' from NASCompares, which has 106 points and 93 comments. Network Attached Storage devices — sold as the privacy-respecting, self-hosted alternative to cloud storage — are increasingly requiring cloud accounts, phoning home, pushing subscription services, and stripping features from local interfaces to steer users toward vendor cloud offerings. The structural pressure is real: NAS manufacturers historically made money on the hardware sale, and as margins compress, monetizing the software layer through subscriptions becomes tempting. The problem is that buyers choose NAS specifically to avoid subscription models — the same dynamic that has undermined the smart home and router markets.

The Rowhammer and RowPress DRAM memory paper on arXiv — 52 points, 33 comments — documents a phenomenon with significant security implications: by repeatedly reading from one row of DRAM cells at high frequency, an attacker can induce bit flips in adjacent rows without requiring any software vulnerability. RowHammer has been known since around 2014 and has been leveraged in real privilege escalation exploits; RowPress is a more recently characterized variant in which even a modest number of reads sustained over time can produce similar effects. Modern cloud infrastructure runs workloads from different customers on the same physical hardware, and mitigations such as targeted row refresh and error-correcting memory are not uniformly deployed across commodity infrastructure.

Jeff Geerling's documentation of achieving 25 Gbps Thunderbolt Ethernet on a Mac Studio earned 192 points and 98 comments, with the interest centering less on the speed figure than on the methodology: Thunderbolt is nominally a peripheral interface, and reliably carrying 25 Gigabit Ethernet over it requires specific adapter hardware and careful system configuration. Posts like this routinely become reference material for engineers building high-throughput local infrastructure. At the more specialized end, the AMD MI450 GPU attention decode kernel optimization guide from the ROCm team — 44 points, no comments at time of recording — addresses kernel-level optimization for LLM inference on AMD hardware using their Gluon framework, an important frontier given NVIDIA's dominance of the AI training market and AMD's active effort to close the gap in inference workloads.