Published on August 28, 2026 at 06:01 CEST (UTC+2)
Saving 100 terabytes of memory by optimizing 1.1.1.1's DNS cache (589 points by TangerineDream)
Cloudflare’s Big Pineapple DNS cache stores over 250 billion entries, so even a one-byte waste per entry costs 250 GB of memory across the fleet. Five memory-layout optimizations cut the per-entry footprint by more than 50%, saving roughly 100 terabytes of RAM overall. The changes also improved performance: insertion throughput rose 43% and lookup latency dropped 19%. This was especially impactful for locations using EDNS Client Subnet, where multiple versions of the same query must be cached.
Small Models Have Arrived (518 points by tosh)
Calvin French-Owen argues that small AI models have reached a surprising level of capability, speed, and affordability. He highlights models like gpt-5.6-luna that run at roughly 100 tokens per second and cost only cents even when searching across thousands of emails. The article explains that high inference token costs have been a major barrier to building consumer AI companies. As small models improve, they could finally make AI-powered consumer products economically viable.
Show HN: OpenTIE and OpenXWA, Modern Ports of Tie Fighter and X-Wing Alliance (96 points by elyosh)
OpenTIE is an open-source reimplementation of Star Wars: TIE Fighter that runs the original game data natively on modern Windows, macOS, and Linux systems. It supports both the 1995 Collector’s CD-ROM and the 1998 Windows release, and can combine the best parts of each into one experience. The repository does not include any game content, so users must provide a copy of the original game. The project also includes OpenXWA, a modern port of X-Wing Alliance.
507 Mechanical Movements (506 points by helloplanets)
The website 507movements.com presents Henry T. Brown’s classic technical reference “507 Mechanical Movements” with animated diagrams. Many of the 507 mechanisms are already animated, with color thumbnails marking the completed ones. Users can browse the collection using prev/next links, and more animations are expected to be added over time. It serves as both a historical archive and an interactive educational resource for mechanical design.
Terminal-Bench-Science: Evaluating AI agents on scientific research workflows (51 points by matt_d)
Terminal-Bench-Science is a new benchmark that evaluates AI agents on expert-curated scientific research workflows. It was led by Stanford researchers and built with domain experts from the life, physical, Earth, mathematical, and engineering sciences. The first release includes 70 tasks, and the strongest model tested, Claude Opus 5, only achieved a 30% resolution rate. The benchmark is designed to evolve continuously alongside frontier AI, creating a feedback loop between scientific needs and AI development.
GoGoGrandparent (YC S16) is hiring back end engineers (1 points by davidchl)
GoGoGrandparent, a YC S16 company, is hiring a senior backend engineer for a fully remote role. The company provides a concierge service that lets older adults use on-demand apps for rides, meals, groceries, medication, and home management over the phone. It has been profitable since 2016 and has not raised venture capital financing. The position involves Node.js and TypeScript, with close collaboration with the founders and real ownership over architecture and product direction.
Gemini-3.5-Transcribe (181 points by k9294)
Google announced Gemini 3.5 Transcribe, a new speech-to-text model designed for precise, real-time transcription. Unlike conventional speech recognition, it handles background noise, complex jargon, and disfluency cleanup, producing clean and formatted text from raw audio. It is already used in products like the Gemini app and Android, and is now available to developers through the Gemini API and AI Studio. Intended use cases include voice agents, real-time captioning, and post-call analytics.
Judge Rules Trump Administration’s Blacklisting of Anthropic Was Illegal (111 points by jbegley)
According to the article title, a judge ruled that the Trump administration’s blacklisting of Anthropic was illegal. The ruling is a legal setback for the government’s action against the AI company. It highlights how AI companies can challenge executive branch decisions through the courts. The full NYT article content was not available in the preview.
Microduck (548 points by robotswantdata)
Microduck is a 25 cm open-source biped robot from Pollen Robotics, available for pre-order at $399. It is trained using reinforcement learning in simulation and then deployed on real hardware through a sim-to-real pipeline. Users can run it out of the box or teach it new tricks, since every behavior is a policy that can be retrained on their own machine. This brings advanced robot learning techniques into a low-cost consumer product.
Show HN: We built open OpenRouter that turns usage into a better model (128 points by SilenN)
Experiential is an open-source model gateway and router that provides one OpenAI-compatible API across hosted, BYOK, local, and custom models. It lets developers control which users and agents can use which models, for which use cases, and how much they can spend. The project also aims to turn production traffic into a custom router or model optimized for quality, speed, and cost. It is positioned as an open alternative to services like OpenRouter.
Small, fast, cheap models are shifting the economics of AI product development.
The rise of models like gpt-5.6-luna shows that small models can now handle real workloads at high speed and near-zero marginal cost. This matters because inference cost has been a major blocker for consumer AI companies, especially those that rely on free or freemium products. Actionable takeaway: developers should build applications that default to small models and escalate to frontier models only when necessary, using a router or gateway to manage cost and quality dynamically.
Benchmarks are expanding beyond software engineering into expert scientific workflows.
Terminal-Bench-Science evaluates AI agents on real research tasks from multiple scientific domains, with top models still failing 70% of the time. This matters because AI’s next big unlock may be accelerating scientific discovery, which requires long-horizon planning, tool use, and deep domain knowledge. Actionable takeaway: AI teams should invest in domain-specific evaluation sets and work directly with experts, rather than relying on generic coding benchmarks, to find real capability gaps.
Open-source model gateways are becoming the control plane for multi-model AI infrastructure.
Projects like Experiential show that teams increasingly want one API to manage hosted, open-source, local, and custom models while enforcing spending and access policies. This matters because no single model provider dominates every task, and organizations need flexibility, privacy, and cost control. Actionable takeaway: adopt an open gateway layer to avoid vendor lock-in, enable model routing, and turn production usage into a signal for optimizing model selection or even training custom models.
Reinforcement learning and sim-to-real transfer are entering consumer robotics.
Microduck demonstrates that a $399 open-source robot can be trained in simulation and deployed on real hardware, with users able to retrain policies themselves. This matters because it makes embodied AI research accessible far beyond big labs, and it could lead to rapid community-driven progress in robot learning. Actionable takeaway: robotics builders should lean on sim-to-real pipelines and open platforms to lower cost and accelerate iteration on new behaviors.
Speech-to-text is moving from raw transcription to polished, structured, real-time understanding.
Gemini 3.5 Transcribe highlights that modern speech models must handle noise, jargon, and disfluencies while outputting clean, formatted text. This matters because high-quality transcription is a foundation for reliable voice agents, captioning, and analytics pipelines. Actionable takeaway: product teams should use purpose-built audio models that clean up conversational speech, reducing the need for fragile post-processing cascades.
Memory efficiency remains a critical, underappreciated performance lever at scale.
Cloudflare’s DNS cache optimization cut per-entry memory usage by over 50% and improved throughput and latency at the same time. This matters because AI workloads increasingly depend on massive caches, vector stores, and key-value systems where single-byte savings translate into terabytes of reduced infrastructure cost. Actionable takeaway: apply data-oriented design, packed data structures, and improved memory locality when building or tuning large-scale AI infrastructure.
Regulatory and legal risk is becoming a first-class concern for AI companies.
The court ruling against the blacklisting of Anthropic shows that government actions can have serious consequences for AI firms, but those actions can also be successfully challenged. This matters because legal and procurement decisions can shape which models get deployed, who builds frontier AI, and how public institutions interact with AI vendors. Actionable takeaway: AI companies should invest in compliance, policy engagement, and legal readiness, not just model capabilities, to navigate an increasingly contested regulatory environment.
Analysis generated by deepseek-reasoner