Article URL: https://blog.xkeeper.net/the-cutting-room-floor/tcrf-2026-ddos-postmortem/ Comments URL: https://news.ycombinator.com/item?id=49835913 Points: 3 # Comments: 0
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A July 2026 study by Zhu et al. (arXiv:2607.19331) has provided mathematical proof that Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR) alter AI model weights through fundamentally different mechanisms. Using Singular Value Decomposition analysis, researchers found that SFT modifies a model's singular value spectrum to inject new factual knowledge, vocabulary, and formatting into its representations. In contrast, RLVR leaves the singular value spectrum nearly unchanged from the base model, instead adapting the model by rotating its input and output coordinate frames. This distinction means RL does not add new facts but rather reorganizes how the model routes and applies its existing pre-trained capabilities for multi-step reasoning. The findings challenge the common assumption that SFT and RL are interchangeable steps in post-training pipelines.
A developer building an Android task app called Wumilo discovered that delivering a simple reminder involves a complex chain of systems, not a single feature. Factors such as exact alarms, notification permissions, battery optimization, and device-specific manufacturer restrictions all affect whether a reminder reaches the user on time. Android's power-management features like Doze and App Standby can delay background activity, and different phone brands may behave inconsistently even when running identical code. The developer found that treating reminders as a multi-step delivery chain — rather than a single scheduling action — made debugging significantly more effective. A key takeaway is that precise alarm scheduling and successful notification delivery are related but distinct problems that must each be addressed separately.
A developer has detailed a system that lets users pay for an app via a web funnel and then open the iOS app already authenticated, even before it is installed. The approach avoids Apple's in-app purchase commission by processing payments entirely on the web, then using a one-time login token to authenticate the user on first app launch. The token is passed to the app using AppsFlyer's deferred deep linking, which survives the App Store installation barrier. A key technical hurdle on iOS was that since iOS 14.5, only a restricted set of deep link parameters are passed to fresh installs, causing custom token fields to be silently dropped. The developer's fix was to place the token inside one of AppsFlyer's explicitly permitted parameter slots, such as deep_link_sub1, rather than a custom query parameter.
A routine public IP change on a cloud server running minimized Ubuntu caused complete SSH loss and zero ping response, because the OS lacked tools like cloud-init or NetworkManager to apply the update automatically. The data center had assigned the new IP at the hypervisor switch level, but the server's network card still held the old, defunct address. Emergency access was restored via an out-of-band VNC console, where temporary ip and route commands brought the network back online instantly. A permanent fix was achieved using systemd-networkd with the GatewayOnLink flag correctly placed under the Route section, alongside cleanup of legacy IP references in configuration files. The incident also revealed that absent ICMP responses do not confirm a server is down, as the data center's anti-DDoS firewall was silently dropping ping packets while TCP connections remained fully functional.
Discussion (0)
Log in to join the discussion and vote.
Log in