tl;dr: Summary not available

tl;dr: OpenAI's new Decisions API (public beta) offers a dedicated `/v1/decisions` endpoint that returns typed answers—predicates (probabilities), choices (categories), or scores (ordered levels)—roughly 10x faster than the Responses API. It currently runs only on gpt-6-luna at $0.10 per 1M input tokens, with no output-token charges, and accepts text or inline base64 images as input. It's aimed at classification, routing, and prioritization tasks, as opposed to Structured Outputs or function calling.

tl;dr: Mistral is previewing Large 4 ("ML4"), a 1T-parameter (52B active) natively multimodal MoE model trained on 3,800 Grace Blackwell GPUs in European datacenters, with open weights promised by month's end. It claims competitive performance with top open-source models and leads in cybersecurity (notably because closed models like Claude Opus 5.5 and GPT-6 refuse such tasks), visual grounding, and legal/financial workloads. The release is backed by Mistral's €3B Series D and positioned as a sovereign, self-deployable alternative to US and Chinese frontier models.

tl;dr: Google released EmbeddingGemma 2, a 740M-parameter multimodal embedding model that maps text, code, images, audio, and video into a unified space for on-device use. Built on Gemma 4 and licensed under Apache 2.0, it features a modular design (270M text-only, with optional 170M vision and 300M audio encoders), an 8K context window, and Matryoshka Representation Learning for up to 6x storage reduction. It claims best-in-class performance for sub-1B models on MTEB Code and MAEB, running in ~191MB RAM for text-only on a Pixel 11 Pro.

tl;dr: The developer of Photopea, a browser-based photo editor, reports that users are stripping ads from his JavaScript code and republishing it on GitHub as their own projects, causing reputation damage when users confuse the clones with the original. He filed a DMCA takedown a month ago, but GitHub rejected it citing no violation of 17 U.S.C. § 1201 (anti-circumvention), likely via an automated response. He's asking HN whether to pursue legal action.

tl;dr: Summary not available

tl;dr: OpenTPU is an open-source AI accelerator designed largely by AI agents, running on a Xilinx Kintex-7 FPGA card with the entire stack (SystemVerilog RTL, ISA, simulator, compiler, and host software) in a single readable monorepo. It executes 10+ modern LLMs (Qwen3, LFM2, Gemma 4, Phi-4-mini, SmolLM3) with real weights, achieving bit-exact parity with its simulator and hitting 82-94% of DDR3 peak bandwidth during decode. Larger MoE models like Qwen3.5-35B run via PCIe expert streaming from the host.

tl;dr: By biomass, cattle dominate animals at 650M tonnes, beating humans (350M) and Antarctic krill (379M), while sugar cane leads crops at 2B tonnes annually. But plants crush everything—the Black Spruce alone (~11B tonnes in Canada) and Russia's Dahurian Larch (8.4B+ tonnes) each contain roughly 3x more carbon than all animals on Earth combined. Bacteria (70B tons of carbon) also massively outweigh animals (2B tons), though an undiscovered deep-sea bacterium could potentially top the list.

tl;dr: AnyPS5 is a tool that ports PS5 executables to Linux and Windows by relinking them to the target system's native format and reimplementing PS5 system PRX libraries for dynamic linking—no emulation or separate runtime involved. It currently maps about 87% of referenced system library functions, includes a shader recompiler that outputs SPIR-V, and supports SDL controllers plus configurable keyboard/mouse input. Verified games are limited so far (e.g., Dreaming Sarah runs at a stable 60fps on a GTX 1050 Ti), and the project is GPLv2-licensed with no bundled proprietary code.

tl;dr: Francis Halzen was awarded the 2026 Nobel Prize in Physics for his decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin. He receives the full prize share.

tl;dr: Paramount Skydance closed its $111B merger with Warner Bros. Discovery, forming a new company called Skydance that combines Paramount+ and HBO Max, CBS and CNN, and major sports properties including CBS Sports and TNT Sports. The deal cleared after Justice Kagan rejected a last-minute block and a California-led multistate antitrust suit was settled, with concessions requiring minimum film investment/releases and separate negotiations for the two entities' basic cable distribution. Free speech and media groups criticized the settlement as inadequate.

tl;dr: Polars 2.0 ships with streaming/out-of-core execution enabled by default (spilling to disk at ~80% RAM), first-class SQL support, and a new Map dtype backed by Arrow's MapType. In TPC-H and TPC-DS benchmarks on AWS c7a instances, Polars outperforms DuckDB 1.5.6, DuckDB 2.0 alpha, and DataFusion on nearly all queries, though it has scaling overhead at 192 threads. The release also tightens strictness around dtypes and schema validation to catch errors earlier, which the team pitches as useful for AI-driven iteration.

tl;dr: A large meta-analysis combining 423 studies and 222,829 data points across 23 ecosystem services found that "functional redundancy"—the assumption that ecosystems can absorb species losses because remaining species pick up the slack—is largely wrong, with most benefits scaling continuously with biodiversity. Ocean carbon sequestration showed the strongest dependence on biodiversity of any service measured, though it's also among the least-studied, raising concerns for blue carbon climate strategies. Exceptions exist for services like coastal flood protection, which often rely on a few foundational species rather than overall diversity.

tl;dr: Gleam v1.19.0 now compiles to Erlang abstract forms (a binary intermediate representation) instead of Erlang source code, bypassing the front-half of the Erlang compiler for significantly faster build times and accurate stack traces mapped to Gleam source. The team opted against targeting BEAM bytecode directly due to its unstable nature across VM versions, following the same approach as Elixir. The release also includes JavaScript codegen optimizations for pattern matching and list literals, better TypeScript interop, and language server improvements.

tl;dr: Reflection is releasing Beam, a 501B-parameter (23B active) open-weight Mixture-of-Experts model targeted at coding, reasoning, and agentic tasks, pretrained on 23.8T tokens and trained with large-scale RL (100M+ rollouts on 10.5K GB300 GPUs over 4 weeks). It claims competitive performance with larger open models like GLM 5.2 while using 3–4× less inference compute, with weights to be released under Apache 2.0 later this month.

tl;dr: Claude Opus 5.5 agents, running DFT simulations, identified two candidate Luttinger-compensated magnets—semiconductors with zero net spin but spin-sorted electron bands useful for spintronics. One, YBaMnFeO₅, was newly designed but likely hard to synthesize due to atomic disorder; the other, KV[Cr(CN)₆], was actually first made in 1999 as a Prussian blue analog and remains magnetically ordered up to 376 K, but its spin-sorting properties were never recognized. The author suggests remaking and measuring KV[Cr(CN)₆] directly, and has published all computations on GitHub.

tl;dr: Summary not available

tl;dr: IANA updated example.com in September 2026 to rotate through six languages via JavaScript with a per-character opacity fade, then days later reverted to showing all languages statically. The article traces the site's history since 2002, covering copy tweaks, infrastructure moves (Apache → EdgeCast → Cloudflare), and the recent addition of a warning telling developers not to use the domain for testing or monitoring.

tl;dr: Cloudflare launched Web Search API in beta, letting AI agents query live internet data through three providers: Ceramic.ai, Exa, and Linkup, all supporting Zero Data Retention. Requests route through AI Gateway for logging and are billed at each provider's list price with no markup, with the option to bring your own API key. It's accessible via REST API or the Workers AI binding.

tl;dr: Dust is a zeroth-order optimization method that trains transformers by perturbing activations independently per token, turning each token into a virtual population member evaluated in a single forward pass. It matches or exceeds backprop on pretraining in several settings and is 1000-10,000x more compute-efficient than weight-space ES methods like EGGROLL. Counterintuitively, larger models (up to 243M params tested) are more population-efficient than smaller ones, and Dust's gradient estimates increasingly align with backprop's as population grows—though it's not yet compute-competitive with backprop at today's scales.