HackerNews Digest

The study, led by researchers from the University of Exeter and Cardiff University and published in *Science*, investigated water transport in tropical Dipterocarp trees ranging from 7 m to 71 m in Malaysian Borneo. Measurements of hydraulic traits along the trunk revealed that taller trees compensate for height by widening basal water‑conducting vessels and by developing leaves with higher wilting thresholds, maintaining liquid water flow despite extreme negative pressures. Growth‑rate data before, during, and after the 2023‑2024 El Niño drought showed no height‑related decline in growth or drought‑induced mortality compared with shorter conspecifics. The findings contradict the prevailing hypothesis that vessel length and gravity limit hydraulic efficiency, photosynthesis, and growth in tall trees. Consequently, models predicting greater drought vulnerability for the tallest 1 % of forest carbon stock may overestimate risk. The research, funded by the Natural Environment Research Council, involved multiple international institutions and highlights the need to reassess hydraulic resilience in other tall tree species.

Leanstrad 1.5 is an Apache‑2.0‑licensed proof‑engineering model for Lean 4 with 119 B total parameters and 6 B active ones. It is trained in three stages—mid‑training, supervised fine‑tuning, and RL with CISPO—using two environments: a multiturn theorem‑proving loop that iteratively refines proofs via Lean compiler feedback, and a code‑agent loop that edits files, runs bash commands, and interacts with the Lean language server to complete long‑horizon proof tasks. Benchmarks show full coverage of miniF2F (100 % validation and test), 587/672 solved on PutnamBench, and state‑of‑the‑art results on abstract‑algebra suites (87 solved on FATE‑H, 34 on FATE‑X). Pass@8 scales smoothly with token budget, reaching 587 solved problems at 4 M tokens. Case studies demonstrate practical code verification: Leanstrad proved AVL‑tree insertion/deletion time bounds over 2.7 M tokens, and an automated Rust‑to‑Lean pipeline flagged 47 property violations across 57 repositories, uncovering 11 real bugs (5 previously unreported). The model, weights, and a free API endpoint are hosted on HuggingFace and usable via Mistral Vibe.

The article compares AMD Instinct MI355X GPUs to NVIDIA Blackwell (B300) GPUs for large‑language‑model inference. AMD hardware costs ≈2.75× less while offering comparable silicon specs, but NVIDIA’s software stack provides superior day‑0 support, making AMD inference slower and harder to set up. Wafer achieved a performance‑per‑dollar improvement on the MI355X by: * Quantizing GLM‑5.2 from bf16 to AMD’s MXFP4 format using AMD Quark, achieving lossless accuracy versus FP8. * Selecting the sglang inference engine (vLLM lacked MXFP4 support; ATOM degraded at long context) and fixing its ROCm image to enable speculative decode, including correcting module‑prefix mismatches for the MTP head and adding a ROCm guard to the fused metadata kernel. * Applying configuration flags (`--kv-cache-dtype fp8_e4m3`, `--enable-aiter-allreduce-fusion`) and tuning MoE kernel selection for fp4 shapes (model_dim 6144, moe_inter 2048, E 256, topk 8). * Optimizing tensor parallelism (TP4×DP2) for prefill‑bound workloads, raising throughput from ~1461 to 2626 tokens / s / node at 2.4 RPS, which is 80 % of Blackwell performance at >2× lower cost. The results demonstrate that with proper quantization, framework fixes, and kernel tuning, AMD GPUs can approach NVIDIA performance while delivering significantly better cost efficiency, indicating the CUDA advantage is diminishing.

The author examined MSI Center, preinstalled on MSI laptops and desktops, by extracting its offline installer (identified as Inno Setup) and decompiling the 170 bundled executables (mostly C#) with ilspycmd; a subset of C++ binaries were inspected in IDA. Searching for insecure patterns revealed that the “Notebook Foundation” service creates a named pipe (\\.\pipe\MSI_SERVICE_2) on boot with permissions allowing any authenticated user to connect. The pipe accepts commands that can read, write, or delete registry keys, manipulate WMI settings, execute or terminate processes, all under the LocalSystem account. Communication is encrypted with 3DES using a client‑provided string as the key; the service brute‑forces decryption against registered client names, enabling privilege escalation when a payload is successfully decrypted. A proof‑of‑concept launches cmd.exe as LocalSystem, and the exploit can be triggered remotely via SMB if valid credentials exist. MSI patched the issue within two days of report (09‑05‑2026 discovery, 01‑06‑2026 patch release) but did not issue a CVE; a CVE request is pending review.

Steam Controller Auto‑Charge is an open‑source web app that autonomously pilots a Steam Controller into its magnetic charging puck using computer‑vision and WebHID. It tracks user‑selected points on the controller and puck via OpenCV.js optical flow (Lucas‑Kanade) from an overhead webcam. The app communicates with the controller through WebHID, sending 70 Hz asymmetric haptic pulses to the dual LRAs for navigation and reducing pulse frequency by 50 % within 150 px of the puck (Proximity Creep Mode). Battery status is read by intercepting HID Report 0x79 (ID 121) for charging confirmation and Report 0x43 (ID 67) for live percentage and cell voltage. The front‑end is built with Vue 3; tracking logic includes a PID loop, object‑avoidance via an in‑browser Rust/WASM CNN, and off‑loaded detection in a Web Worker. The only dependency is Nix; the app runs on Windows, macOS, and Linux in Chromium‑based browsers with WebHID support. MIT‑licensed.

SearXNG is a free, open‑source metasearch engine that aggregates results from multiple search services and databases while ensuring users are not tracked or profiled. The repository provides an installation guide, a configuration guide for fine‑tuning, and links to additional documentation. Community involvement is encouraged through contribution guidelines, and users can connect via the project’s CONTRIBUTING resources. The software is released under the GNU Affero General Public License (AGPL‑3.0). The GitHub page also displays visual elements indicating organization, documentation, licensing, commit history, and translation status.

The article recounts the deletion of the Wikipedia page for the Odin programming language through an Articles for Deletion (AfD) process, noting a 5‑to‑7 vote for deletion versus a 1‑to‑4 vote for keeping. Creator “GingerBill” and others argue that the decision reflects Wikipedia’s reliance on traditional, independent sources, which are scarce in modern software development. Tweets from GingerBill cite perceived bias among moderators, the language’s commercial use (e.g., JangaFX), community adoption, and the inadequacy of primary sources such as the language’s own website or creator statements. Casey Muratori adds that programming‑related notability guidelines are mismatched to the field, where reliable coverage often appears in blogs, YouTube, or internal documentation rather than peer‑reviewed journals. Both emphasize that strict application of existing notability rules would eliminate most contemporary programming topics. The piece suggests adding subject‑specific guidelines—e.g., metrics based on repository counts or notable code‑hosting platforms—to better accommodate software‑industry sources while maintaining Wikipedia’s standards against self‑promotion.

None

The article contrasts differential and integral calculus to illustrate why synthesis (global reasoning) is harder than analysis (local reasoning). Differential calculus computes a function’s slope at a point using only nearby behavior, making it algorithmic and easily automated (e.g., automatic differentiation in machine learning). Integral calculus seeks area under a curve, requiring knowledge of the function over an entire interval; no general algorithm exists for arbitrary functions, and many integrals lack closed‑form solutions, needing series expansions or numerical methods. The Fundamental Theorem of Calculus links the two as inverse operations. The author extends this mathematical analogy to software engineering, arguing that analysis breaks problems into localized components (facilitated by encapsulation and separation of concerns), while synthesis must integrate multiple components, a global task that is intrinsically more difficult. In fields like incident response and site reliability engineering, synthesis expertise is undervalued despite being essential for understanding system interactions and resolving complex issues. Improving methods for learning system-wide behavior is therefore advocated.

In the late‑1600s Amsterdam, rapid economic growth created dense fire hazards in homes, workshops, and shipyards. The city first bought dozens of Hans Hautsch water‑pumping engines (horse‑drawn, bucket‑fed) and equipped each district with ladders, hooks, tarpaulins and 28 000 leather buckets, assigning firefighting duties to four guilds. Early incidents—such as the 1669 sugar‑refinery blaze and the 1672 Blaeu printing‑office fire—revealed that the engines could not deliver water to roofs or interior spaces, especially in cold weather. Jan van der Heyden, a painter‑inventor, introduced a new fire‑engine system (with his brother Nicolaas) featuring a suction hose, a flexible “snake” hose, and an air‑chamber for continuous high‑pressure flow. After a decisive 1672 harbor fire, the city retrofitted half its engines and later ordered lightweight models for each of the sixty districts, adding regulated alarms (trumpeters, night watchmen) and monetary incentives for rapid response. From 1682‑1687 fire losses fell to 18 353 guilders (≈ 1 % of the prior decade’s losses), demonstrating the effectiveness of the integrated technical and organizational reforms.