Published on August 09, 2026 at 06:01 CEST (UTC+2)
My server is a phone now (166 points by seg6)
The author describes replacing a paid Hetzner VPS with their existing CMF Phone 1, which has 8 ARM cores, 8 GB of RAM, 128 GB of flash, Wi-Fi 6, a 5G modem, and a built-in battery backup. The phone now runs their personal infrastructure: a remote browser called Surf, a finance tracker, a screen sharing service, and several smaller web apps. These applications survive reboots, deploy from Git, and remain reachable as the phone roams between networks. The article reflects a growing pragmatic trend of repurposing underused consumer hardware for self-hosting.
Os8088: A powerful Mac-like OS for the IBM XT, 286, 386 (82 points by jggonz)
Os8088 is a graphical operating system for the Intel 8086 (and 286/386) that mimics a 1984 Macintosh System 1 desktop, written entirely in real-mode assembly and booted from a floppy disk. It features overlapping windows, pull-down menus, a serial mouse, loadable programs, and pre-emptive multitasking — something the actual 1984 Mac never had. The entire system runs in 256 KB of RAM at 640×480 resolution with 16 colors, using only a 78,950-byte kernel. It can be booted in a browser or from actual floppies, supporting twelve task slots with 18.2 Hz pre-emption.
Improving Heuristics for A* Pathfinding (114 points by bobbiechen)
This article from Red Blob Games examines how to improve A* pathfinding by refining its heuristic function rather than just optimizing the priority queue or map representation. It demonstrates with interactive examples that standard distance-based heuristics often push the search in the wrong direction when walls are present, wasting time exploring unnecessary nodes. The key idea is that a "perfect" heuristic that knows about walls can dramatically shrink the search space, though it must be recomputed for each goal and wall configuration. The page ends with real-game map examples showing practical savings.
Fastmail offers EU data region (339 points by groomlake)
Fastmail now offers an EU data region, giving users the choice to have their primary email data stored on Fastmail's own secure servers in Amsterdam instead of the US. The company emphasizes that this is not rented cloud infrastructure but their own hardware and software, co-located in a secure facility and managed by their in-house team. Data is encrypted at rest inside locked racks across all locations, with two primary copies plus a geographically separate backup. This gives European users options for local law compliance, data proximity, and regulatory alignment.
The phone book that led us to Assad's spy chief in hiding (28 points by cwwc)
This BBC investigation describes how reporters found Syria's spy chief in hiding after the Assad regime fell. The trail began in an abandoned luxury apartment in Damascus, where armed men sifted through designer clothes, documents, and dozens of photos of the same slender man under a crystal chandelier. A phone book scattered among the debris turned out to be the crucial clue that led the investigation to the intelligence chief. The article is an example of open-source investigative journalism using physical artifacts and document trails.
Shopify replaced Redis with MySQL for inventory reservations–and it scaled (86 points by adletbalzhanov)
Shopify engineering describes migrating their inventory reservation system from Redis to MySQL after initially failing with earlier MySQL attempts. The solution relies on MySQL 8's SKIP LOCKED feature, composite primary keys, and a design using one row per inventory unit rather than one row per item with a count. This approach handled the contention problem that plagued a naive single-row quantity column design. On Black Friday 2025, they hit a record $5.1 million in sales per minute at peak, proving the MySQL-based system scaled successfully as part of a unified database strategy.
Dithered QR Codes (66 points by jmusall)
This page explains how to create error-diffused dithered QR codes. QR codes are built from function patterns (seen as bold shapes) that scanners use to locate the code, and data modules that carry the encoded information; the function patterns must stay clear, while data modules can be modified. Brands frequently shrink or restyle data modules for visual distinction, and scanning still works because scanners locate module centers using the finder patterns. The article walks through how error diffusion can be used to embed images into a QR code while keeping it scannable.
Unexpected events and prosocial behavior: the Batman effect (20 points by davidbarker)
This scientific paper describes a quasi-experimental field study on the Milan metro testing whether an unexpected event — the presence of a person dressed as Batman — increases prosocial behavior. Passengers were significantly more likely to offer their seat to a pregnant woman when Batman was present (67.21% vs 37.66%, OR = 3.393, p < 0.001). Notably, 44% of those who offered a seat in the Batman condition appeared to notice him before acting. The authors suggest unexpected events disrupt automatic routines and enhance attention to the present moment, triggering helping behavior.
_for-sale DNS records (366 points by shaunpud)
This specification describes _for-sale, a reserved DNS leaf node name defined by RFC 10023 and registered with IANA. A TXT record published at _for-sale.example.com signals that a domain is available for purchase while remaining fully operational — the homepage serves, mail flows, and the record can be added or removed at will. This is explicitly not domain parking, which costs visitors by replacing the site with a sales page. Tags include ftxt, furi, fval, and fcod for free-text notes, contact URIs, asking prices, and proprietary codes, filling a gap in WHOIS/RDAP data.
Open-source interactive map for the Aug 12 total solar eclipse (100 points by MarcoDewey)
This is an open-source interactive map showing the total solar eclipse of August 12, 2026. The tool includes layers for eclipse geometry, Besselian elements, live umbra tracking, a three-dimensional shadow view, and cloud projections. Users can position themselves at different geographic coordinates to explore the eclipse path at any time point. It is a community-built alternative to commercial eclipse visualization tools.
Edge AI and resource-squeezed compute are becoming the norm. Articles 1 and 2 both demonstrate serious software running on hardware most would dismiss as underpowered: a phone acting as a full personal server, and a Mac-like GUI OS in 256 KB of RAM. This parallels the broader AI shift toward small quantized models (Llama.cpp, Qwen, Phi) that run on phones and laptops. Why it matters: AI workloads no longer assume a datacenter; edge devices can serve inference, fine-tune locally, and preserve privacy. Implication: Teams should prioritize model compression, quantization, and on-device inference to reduce cloud costs and enable offline AI experiences.
Algorithmic cleverness beats brute-force infrastructure. Article 3's "perfect heuristic" for A and Article 6's MySQL SKIP LOCKED design both show that a smarter algorithm or data layout can outperform throwing more hardware at a problem. Why it matters: AI is hitting a similar inflection — FlashAttention, speculative decoding, and distillation yield large speedups without larger GPU clusters. Implication:* Before scaling ML infrastructure, audit your algorithms and data structures; optimizing the loss landscape, sampling strategy, or attention pattern often delivers bigger wins than adding nodes.
Data sovereignty is reshaping AI infrastructure. Article 4's EU data region reflects a demand for regulatory compliance, local law adherence, and data proximity that now extends into AI. The EU AI Act and GDPR push model training and inference into regional boundaries, and we're seeing the rise of sovereign/regional LLM initiatives. Why it matters: AI models trained on data in one jurisdiction may not legally operate on data in another; multi-region data residency is becoming a first-class design requirement. Implication: Build AI pipelines with data residency in mind from day one — regional model deployment, local vector databases, and encryption-at-rest in each region will be expected, not optional.
Contextual triggers and behavioral influence require ethical guardrails in AI agents. Article 8's Batman effect shows that subtle environmental disruptions (a costumed person entering a metro car) can dramatically increase prosocial behavior — from 38% to 67% seat-offering. This has direct implications for AI assistants, recommender systems, and autonomous agents that nudge human decisions. Why it matters: AI systems increasingly shape user behavior through prompts, notifications, and defaults; understanding how unexpected events alter decision-making can make AI more effective — or more manipulative. Implication: Deploy behavioral AI with transparency and consent safeguards; measuring AI-driven behavior change demands identical rigor to the Batman field study, including control groups and statistical reporting.
Robustness to degraded input is a shared design goal for vision AI. Article 7's dithered QR codes show how error-diffusion can embed images into QR codes while preserving scannability — meaning computational systems can tolerate heavily modified visual input when the underlying encoding has redundancy. Why it matters: Real-world AI vision must handle blurry photos, occluded objects, and adversarial perturbations; the QR code's robustness-versus-aesthetics tradeoff maps directly to training vision models that survive distribution shift. Implication: Design vision systems with explicit error-correcting capacity and test against corrupted/adversarial inputs; techniques like data augmentation, contrastive learning, and robust loss functions are the AI analogues of QR error correction.
Machine-readable standards will unlock the agentic web. Article 9's _for-sale DNS record and Article 10's open-source eclipse map both encode structured, machine-readable metadata that automated agents can discover and act upon. Such standards — DNS TXT records, schema.org, structured map layers — are the foundation for AI agents that navigate, negotiate, and transact autonomously. Why it matters: LLM-driven agents today rely on scraping and parsing unstructured pages; as standards emerge, agents can query intent, prices, and availability directly, making automation cheaper and more reliable. Implication: Publish structured metadata for your services and APIs; design data with agent accessibility in mind, because the next wave of AI systems will consume machine-readable signals preferentially over prose.
Human-in-the-loop intelligence is the near-term frontier for investigative AI. Article 5's BBC investigation shows that turning scattered physical artifacts (photos, phone books, documents) into a tracked-down fugitive remains a deeply human cognitive achievement. Multimodal AI can assist — transcribing documents, matching faces in photos, clustering names — but the integration of messy, ambiguous, real-world evidence still needs human judgment. Why it matters: AI-assisted OSINT is exploding across journalism, defense, and law enforcement, yet its failure modes (hallucinated connections, false positives) are dangerous in high-stakes contexts. Implication: Build human-in-the-loop pipelines where AI does the heavy lifting on retrieval and correlation, and humans do final reasoning; invest in uncertainty estimation and provenance tracking for every AI-derived lead.
Analysis generated by deepseek-reasoner