tl;dr: Amazon replaced Claude Sonnet 4.6 with Strands Decider E4B (a Gemma-based decision model that outputs probabilities over fixed options rather than text) for labeling internal tickets, cutting costs ~98% (~$22,200 to ~$500 per million tickets) while improving accuracy and latency by 78%. The team ran both models in parallel via an AppConfig switch to validate before cutover. The approach works when classification is a closed-choice problem with a stable catalog and context that fits in a few thousand tokens; one gotcha was that JSON key order affected output probabilities by up to 0.19.

tl;dr: Anthropic released Claude Fable 5.2 and Mythos 5.2 (same underlying model, with Mythos lacking certain cyber/bio safeguards and restricted to vetted orgs), claiming state-of-the-art performance on coding, reasoning, and vision benchmarks—beating Opus 5.5 and competitors like GPT-6 and Gemini 4 on most tests. Notable demonstrated feats include formalizing Fermat's Last Theorem in 30 hours via 1,000 agents, solving two open black hole physics problems, designing working electronic devices, and migrating a production service from Python to Rust. Pricing is $8/$40 per million input/output tokens, with a 2.5x "fast mode" at double the cost.

tl;dr: USPS is now selling official ZIP Code boundary polygon data derived from its delivery address information, filling a long-standing gap since ZIP Codes are delivery routes rather than true geographic areas. The dataset is offered in three licensing tiers (Internal, Enterprise, Publication), each with quarterly updates, targeting industries like mapping, insurance, real estate, and logistics.

tl;dr: Summary not available.

tl;dr: E-means replaces Lloyd's full-batch k-means updates with mini-batch exponential moving average updates (borrowed from VQ-VAE training), reaching Lloyd's 25-epoch loss in 1.2–3 epochs, yielding 10–45× GPU and 8–15× CPU wall-clock speedups. It requires no hyperparameter tuning—batch size and step size are set via fitted scaling laws—and adds cosine-decayed step sizes, improved dead-cluster revival via splitting large clusters, and a finalization stage. The authors attribute the quality gains to noisier mini-batch updates better escaping local minima; code ships in Faiss as `faiss.contrib.e_means`.

tl;dr: Foldable Flight is a catalog site featuring 109+ paper airplane designs, each rated by speed, glide, and accuracy, with difficulty levels from Very Easy to Expert. Notable entries include record-holders like Unite (31.2-second airtime world record) and Firefly (300+ ft flight), alongside community-submitted designs and origami-style jets from designers like Michael LaFosse. Users can filter by paper type (A4, US Letter, Square) and click any plane for folding instructions.

tl;dr: Researchers tested semi-transparent cadmium telluride solar panels that let red and some blue wavelengths through (casting magenta light) while converting the rest to electricity, growing broccoli beneath them. Yields matched unshaded controls despite taking 25 extra days to mature, and the panels actually improved the microclimate by reducing vapor pressure deficit, boosting biomass conversion efficiency 2.8–4.5x. Macronutrient levels held up, though some micronutrients like iron and zinc dipped slightly. The study was small-scale and preliminary.

tl;dr: Summary not available

tl;dr: Google has quietly reversed its 2023 promise that personal Gmail, Docs, and Drive content connected to Gemini wouldn't be seen by human reviewers, now explicitly warning users not to connect apps containing confidential data. Testing shows Gemini probabilistically ingests Workspace content—ranging from metadata summaries to full unredacted file contents including PII—triggered autonomously by casual conversation without explicit @Drive or @Gmail tags. Google's automated scrubbers reportedly fail to deidentify narrative data like account numbers and transaction details before contractor review.

tl;dr: Meta introduced Muse Ember, a small, efficient multimodal model aimed at high-volume agent tasks like tool calls, coding subtasks, extraction, and document parsing, with a 1M-token context and 512K output. It reportedly beats Claude Haiku 5.5 and GPT-6 Luna on intelligence-per-dollar, matching Haiku's top score at roughly a third of the cost. Available now via Meta Model API at $0.10/M input and $0.50/M output tokens, it's the first model trained using Muse Wonder Full.

tl;dr: Using Claude Fable 5.2 with Claude Code, the author one-shotted a complete 12-level, 4-player beat 'em up game for his nephew based on hand-drawn character designs, generating ~175 images (via ChatGPT Images 2.5) and ~140 audio clips (via ElevenLabs) overnight. The agent rewrote most of the existing code, built a bot to playtest all levels in a headless browser, and produced cutscenes, music, and an announcer unprompted. Total cost: ~340M tokens across 600 requests over two days, eventually expanding to 18 unlockable characters based on family members.

tl;dr: Claude Fable 5.2 has achieved a score of 65 on Artificial Analysis's Intelligence Index v5.0.1, the highest recorded so far on the benchmark, which aggregates nine evaluations spanning agentic coding, scientific reasoning, knowledge reliability, and document reasoning. The model currently has no publicly available API. The index also tracks cost per task, token usage, context window, and latency metrics alongside raw intelligence scores.

tl;dr: Google expanded its 8th-gen TPU lineup with two new systems: TPU 8tx, which adds hardware-level confidential computing (keeping data encrypted even in memory) for regulated enterprises, and TPU 8i Max, which packages two TPU 8i chips together to double compute, SRAM, and HBM bandwidth for million-token context workloads. Alongside the hardware, Google is pushing broader TPU accessibility via open-sourced TorchTPU, native PyTorch/vLLM/SGLang support, an XLA Serving Engine, and tools like TPU Sync and GKE dynamic slicing to reduce latency and improve utilization.

tl;dr: A user claims that Claude Opus 5.5's reasoning behavior shifted noticeably on November 3, based on an analysis of 1,400 Claude Code sessions run through Gemini 4 Argon. While the model name remained unchanged, the analysis suggests an unannounced modification affected how the model reasons, fueling speculation that it has been "nerfed."

tl;dr: A Boulder woman is suing the Ethereum Foundation after being fired for organizing employee resistance to a 1,000-year "mandate" document that included a "Source Seppuku License" pledging authors would commit suicide by sword if they failed to keep the software free. Foundation leadership, including Vitalik Buterin, defended the clause as artistic/cultural reference not meant literally, but demanded employees commit to the mandate or resign. The lawsuit alleges disability discrimination and retaliation, and claims the mandate triggered an exodus, with most of the 53 employees who signed a protest letter later resigning or being terminated.

tl;dr: Security researchers chained several vulnerabilities in a major automaker's connected car platform to remotely unlock any of their vehicles in Australia (and later the US and Canada) using just a license plate number. The attack chain combined plate-to-VIN lookups, hardcoded OpenIDM admin credentials found in a JavaScript bundle, a GPT-assisted WAF path traversal bypass, and an MFA bypass via OAuth's Resource Owner Password Credentials flow. This allowed them to mint auth tokens for any user and issue commands like door-unlock and engine-start via the telematics API.

tl;dr: BenchUp (branded as "KetchUp" in the HN title) is a free, open-source 3D modeling app aimed at woodworkers and furniture designers, offering SketchUp-like tools such as Push/Pull, snapping, dimensions, groups, and components. Built in Rust with WebGPU/WebGL2, it runs in desktop browsers or as a native app on Windows, macOS, and Linux, with local autosave and no account required. It supports importing SketchUp (.skp) geometry and exporting OBJ, and is MIT-licensed on GitHub.

tl;dr: The article walks through building a toy malloc from scratch using sbrk and mmap, explaining chunk headers, free lists, splitting/coalescing, demand paging, and trimming the heap back to the kernel. Benchmarks show the toy allocators run 65-150x slower than glibc, which stays fast by using size-segregated bins, a thread-local cache (tcache) of recently freed chunks, and an adaptive mmap threshold—trading memory overhead to avoid syscalls on the hot path.

tl;dr: Summary not available

tl;dr: Summary not available.