StepFun 3.5 Flash Is #1 Cost-Effective Model for OpenClaw Tasks (300 Battles)
StepFun 3.5 Flash emerged as the top‑ranked model on the OpenClaw gaming benchmark, achieving a score of 1,327 ± 88 over 98 battles. It outperformed Grok 4.1 Fast (1,274 ± 70) and Minimax M2.7 (1,182 ± 75), securing the most cost‑effective position among 15 contenders. The ranking is provisional, with tight confidence intervals suggesting stability, but additional battles could shift positions. The table also highlights a wide spread of scores across models from major AI providers.
The OpenAI Graveyard: All the Deals and Products That Haven't Happened
OpenAI has abruptly halted several high‑profile initiatives, most notably the Sora video‑generation app and its $1 billion Disney licensing deal, as it pivots toward tighter financial discipline ahead of a potential IPO. Other projects such as the NSFW ChatGPT mode, Instant...
7,655 Ransomware Claims in One Year: Group, Sector, and Country Breakdown
From March 2025 to March 2026 ransomware groups posted 7,655 victim claims on public leak sites, averaging about 20 claims per day. The five most active groups—Qilin, Akira, INC Ransom, Play and Safepay—account for 40% of all claims, with Qilin alone responsible...
Google's 200M-Parameter Time-Series Foundation Model with 16k Context
Google Research released TimesFM 2.5, a decoder‑only time‑series foundation model with 200 million parameters, down from 500 million in version 2.0. The new model supports a 16 k context window, far exceeding the prior 2 048 limit, and adds an optional 30 million‑parameter quantile head for continuous...
Ollama Is Now Powered by MLX on Apple Silicon in Preview
Ollama announced a preview that leverages Apple’s MLX framework to deliver the fastest local inference on Apple Silicon Macs. The integration taps the GPU Neural Accelerators of M5‑series chips, boosting both time‑to‑first‑token and token‑per‑second rates, especially with the Qwen3.5‑35B‑A3B model...
Axios Compromised on NPM – Malicious Versions Drop Remote Access Trojan
On March 30‑31 2026, attackers compromised the npm maintainer account for the popular HTTP client library axios and published two malicious releases, axios@1.14.1 and axios@0.30.4. Both versions added a fake dependency, plain‑crypto‑js@4.2.1, which runs a post‑install script that drops a...
Philly Courts Will Ban All Smart Eyeglasses Starting Next Week
Philadelphia’s First Judicial District announced a ban on all smart or AI‑integrated eyewear effective Monday. Any glasses with video or audio recording capabilities will be prohibited throughout courthouses, even if prescribed, with violations potentially resulting in contempt charges. The policy...
ChatGPT Won't Let You Type Until Cloudflare Reads Your React State
Every ChatGPT request now triggers a Cloudflare Turnstile challenge that runs a virtual‑machine program inside the browser. The bytecode arrives encrypted as a 28,000‑character base64 string, but the decryption key – a server‑generated float – is embedded directly in the...
Midnight Train From GA: A View of America From the Tracks as Airports Struggle
A federal security‑staff walkout tied to a budget stalemate over immigration enforcement crippled Atlanta’s Hartsfield‑Jackson airport, forcing travelers to confront hours‑long TSA lines. Writer Donald Trump‑era politics prompted a 14½‑hour, 650‑mile overnight Amtrak Crescent trip from Atlanta to Washington as a...
The Cognitive Dark Forest
The essay frames today’s AI‑driven web as a "cognitive dark forest" where every prompt and code snippet becomes data that large platforms harvest. It argues that the early internet rewarded open sharing, but consolidation of services and cheap AI execution...
Voyager 1 Runs on 69 KB of Memory and an 8-Track Tape Recorder
Voyager 1, now over 15 billion miles from Earth and traveling 38,000 mph, remains the most distant human‑made object after 48 years in space. It operates on a modest 69 KB of memory and an 8‑track digital tape recorder, transmitting data at just 160 bits per...
Solar Is Winning the Energy Race
Since 2000, solar has outpaced all energy sources, rising from 228 GW in 2015 to an estimated 2,919 GW in 2025, now providing about 10% of global electricity. China leads with roughly 1,300 GW and dominates panel manufacturing, while the EU and the...
Show HN: Public Transit Systems as Data – Lines, Stations, Railcars, and History
An open‑source dataset compiles key metrics for nine major transit systems worldwide, covering stations, lines, track length, and daily ridership. Systems range from Baltimore’s modest Light RailLink with 33 stations to Beijing’s massive subway with 527 stations and 10.5 million daily...
Alzheimer's Disease Mortality Among Taxi and Ambulance Drivers (2024)
A population‑based analysis of U.S. death certificates from 2020‑2022 examined Alzheimer’s disease mortality across 443 occupations. Taxi drivers (1.03% of deaths) and ambulance drivers (0.91%) showed the lowest adjusted proportions of Alzheimer‑related deaths, well below the overall adjusted rate of...
Don't YOLO Your File System
Developers are increasingly seeing AI agents wipe files, empty directories, and corrupt home folders when given unrestricted system access. The new open‑source tool jai offers a single‑command sandbox that isolates an agent’s workspace while keeping the current working directory writable....
Chicago Artist Creates Tourism Posters for City's Neighborhoods
Chicago illustrator Steve Shanabruch has turned a personal side project into a thriving business by designing WPA‑inspired tourism posters for more than 60 Chicago neighborhoods. The $28 prints, sold online and in venues like the Willis Tower Skydeck shop, now...
Show HN: I Put an AI Agent on a $7/Month VPS with IRC as Its Transport Layer
A developer deployed an AI chatbot on a $7‑per‑month virtual private server and connected it to a self‑hosted IRC server, allowing visitors to query his GitHub repositories directly. The public bot, named nullclaw, answers questions by cloning code and extracting...
Updates to GitHub Copilot Interaction Data Usage Policy
GitHub announced that, starting April 24, interaction data from Copilot Free, Pro, and Pro+ users will be used to train its AI models unless users opt out. The policy excludes Copilot Business and Enterprise customers, whose data remains untouched. Users can...
Quantization From the Ground Up
Quantization compresses large language model weights from 16‑bit or 32‑bit floating point formats to lower‑precision representations such as 8‑bit, 4‑bit, or even 2‑bit integers. By applying blockwise scaling, outlier parameters are handled without overwhelming the compression, preserving most of the...
Ensu – Ente’s Local LLM App
Ente has launched Ensu, an offline large‑language‑model chat app that runs entirely on iOS, Android, macOS, Linux and Windows devices. Built with a Rust core and released as open‑source software, Ensu offers zero‑cost, privacy‑first interactions without sending data to cloud...
TurboQuant: Redefining AI Efficiency with Extreme Compression
TurboQuant, a new vector quantization framework slated for ICLR 2026, tackles the memory‑overhead problem of high‑dimensional AI vectors by combining PolarQuant’s rotation‑based compression with a 1‑bit Quantized Johnson‑Lindenstrauss (QJL) error‑correction step. The method compresses key‑value (KV) caches to as little as...
LiteLLM Python Package Compromised by Supply-Chain Attack
A malicious .pth file was discovered in the LiteLLM 1.82.8 wheel on PyPI, automatically executing a credential‑stealing script whenever the Python interpreter starts, without any import of the library. The payload gathers environment variables, SSH keys, cloud provider credentials, and...
DOOM Over DNS
The Doom‑Over‑DNS project streams the classic shareware DOOM game using only DNS TXT records hosted on Cloudflare. By compressing the 1 MB WAD file into roughly 2,000 TXT entries, a PowerShell 7 script fetches and reassembles the data entirely in memory, bypassing...
POSSE – Publish on Your Own Site, Syndicate Elsewhere
POSSE (Publish on Your Own Site, Syndicate Elsewhere) is a content strategy that prioritizes publishing first on a personal website before distributing copies or links to third‑party platforms such as Twitter, Facebook, Mastodon, and others. The model enhances ownership, canonical...
Chest Fridge (2009)
A chest freezer repurposed as a fridge can consume as little as 0.1 kWh per day, dramatically out‑performing conventional upright models. The author’s 20‑year campaign pushed manufacturers to add +6 °C temperature control, enabling chest freezers to serve as efficient fridges. Replacing...
Floci – A Free, Open-Source Local AWS Emulator
Floci is a free, open‑source local AWS emulator that runs with a single Docker Compose command, positioning itself as a no‑strings‑attached alternative to LocalStack’s community edition, which will sunset in March 2026. It eliminates authentication tokens, offers unlimited CI/CD support, and...
Tinybox – Offline AI Device 120B Parameters
Tinycorp announced that its tinybox line of on‑prem AI servers is now shipping, with the red v2 model priced at $12,000 and delivering 778 TFLOPS of FP16 performance. The green v2 blackwell version, equipped with four RTX PRO 6000 GPUs, provides 3.1 PFLOPS and...
OpenCode – The Open Source AI Coding Agent
OpenCode is an open‑source AI coding agent that runs in terminals, desktops, and IDEs, automatically loading the appropriate language‑server protocols for large language models. It supports over 75 LLM providers—including local models—and integrates with GitHub Copilot and ChatGPT Plus/Pro accounts....
Show HN: I Made an Email App Inspired by Arc Browser
João announced an early‑access email application built on design cues from the Arc browser. The Show HN post invites a select group to try a demo that includes an exclusive 3D digital welcome card. All messages in the demo are...
Tin Can, a 'Landline' For Kids
Tin Can, a $75 Wi‑Fi landline phone, lets children call approved contacts during set hours, offering a safe alternative to smartphones. Founded by three Seattle dads, the startup has raised $3.5 million and reports tens of thousands of units sold since...
France's Aircraft Carrier Located in Real Time by Le Monde Through Fitness App
On March 13, a French Navy officer logged a 7‑km run on the Strava app while aboard the aircraft carrier Charles de Gaulle. Because his profile was public, the GPS data instantly revealed the carrier’s exact position northwest of Cyprus,...
Nvidia NemoClaw
NVIDIA introduced NemoClaw, an open‑source stack that wraps the OpenShell runtime to run OpenClaw always‑on assistants in a sandboxed environment. The platform installs the OpenShell gateway, creates a container with Landlock, seccomp and network‑namespace policies, and routes inference calls to...
Federal Cyber Experts Called Microsoft's Cloud "A Pile of Shit", Yet Approved It
Federal reviewers harshly criticized Microsoft’s Government Community Cloud High (GCC High), calling it a “pile of shit” due to missing security documentation and unclear encryption practices. Despite these concerns, the FedRAMP program granted the cloud suite an authorization, effectively giving it...
OpenRocket
OpenRocket is a free, open‑source model‑rocket simulator that combines aerospace‑grade six‑degrees‑of‑freedom flight analysis with a user‑friendly CAD‑style interface. It lets designers create, modify, and instantly evaluate rockets, offering real‑time data on altitude, velocity, stability, and more. An AI assistant can...
IBM, Sonic Delay Lines, and the History of the 80×24 Display (2019)
IBM’s 3270 terminal, launched in 1971 with an 80×24 display, became the market‑dominant CRT terminal and forced competitors to adopt the same size. The later IBM PC’s monochrome display adapter added a fifth line, establishing the 80×25 format that persists...
A Visual Introduction to Machine Learning (2015)
The article uses a simple housing dataset to illustrate core machine‑learning concepts, focusing on a decision‑tree classifier that separates San Francisco from New York homes based on elevation and price per square foot. It walks through intuition, feature selection, split points, recursion,...
I Found 39 Algolia Admin Keys Exposed Across Open Source Documentation Sites
Security researcher Ben Zimmermann uncovered 39 Algolia admin API keys embedded in open‑source documentation sites that use Algolia’s free DocSearch service. The keys, found through large‑scale frontend scraping and GitHub code searches, grant full write, delete and settings permissions on...
Mouser: An Open Source Alternative to Logi-Plus Mouse Software
Mouser is an open‑source, fully local utility that replaces Logitech Options+ for the MX Master 3S mouse. It lets users remap all six programmable buttons, adjust DPI from 200 to 8,000, and invert scroll directions without any cloud telemetry. The tool supports...
1M Context Is Now Generally Available for Opus 4.6 and Sonnet 4.6
Claude announced that Opus 4.6 and Sonnet 4.6 models now support a full 1 million‑token context window at regular pricing. The pricing remains $5/$25 per million tokens for Opus 4.6 and $3/$15 for Sonnet 4.6, eliminating any long‑context surcharge. Media limits have been expanded sixfold...
Can I Run AI Locally?
The article provides a detailed matrix of open‑source LLMs, ranking them by size, memory usage, context window, and token throughput to show which can run on a typical consumer machine. Models under 2 B parameters (e.g., Meta 1B, Google 1B) run smoothly, while...
1B Identity Records Exposed in ID Verification Data Leak
Researchers uncovered an unprotected MongoDB database belonging to IDMerit that exposed roughly 1 billion identity records across 26 countries, including more than 203 million records in the United States. The data set contained full names, addresses, dates of birth, national ID numbers...
3D-Knitting: The Ultimate Guide
3D‑knitting, driven by Wholegarment machines, creates fully seamless garments from a digital model, using up to 99% of yarn and eliminating waste. Originating from Shima Seiki’s 1995 invention, the technology offers superior durability, comfort, and on‑demand customization while reducing labor...
Agents that Run While I Sleep
Developers are using Claude‑powered agents to generate code autonomously, but lack reliable verification. Traditional code reviews are overwhelmed as agents produce dozens of pull requests weekly, prompting a need for automated testing. The author proposes a TDD‑style workflow: write precise...
Launch HN: RunAnywhere (YC W26) – Faster AI Inference on Apple Silicon
RunAnywhere’s RCLI brings a full speech‑to‑text, large‑language‑model, and text‑to‑speech pipeline to macOS, running natively on Apple Silicon via the proprietary MetalRT GPU engine. The tool supports 43 voice‑controlled macOS actions, local retrieval‑augmented generation over user documents, and delivers sub‑200 ms end‑to‑end...
Debian Decides Not to Decide on AI-Generated Contributions
Debian’s developers opened a discussion in February on whether to accept AI‑assisted contributions, drafting a general resolution that would require explicit disclosure and machine‑readable tags for any LLM‑generated code. The debate highlighted divergent views on terminology, licensing, onboarding, and the...
FreeBSD Capsicum Vs. Linux Seccomp Process Sandboxing
FreeBSD’s Capsicum and Linux’s seccomp‑bpf both aim to curb the ambient authority of compromised processes, but they take opposite approaches. Capsicum adopts a subtraction model: a single irreversible cap_enter() call strips the process of all global namespaces, leaving only explicitly granted...
Reverse-Engineering the UniFi Inform Protocol
A former UniFi hosting provider discovered that the UniFi inform protocol includes an unencrypted MAC address in the first 40‑byte header. By extracting this MAC, he built a lightweight Go proxy that routes device inform packets to the correct tenant...
Ireland Shuts Last Coal Plant, Becomes 15th Coal-Free Country in Europe (2025)
Ireland retired its final coal‑fired power plant in 2025, making it the 15th European nation to operate without coal. The closure ends a decades‑long reliance on coal for baseload electricity and follows a broader EU push toward net‑zero emissions. Ireland’s...
Launch HN: Cekura (YC F24) – Testing and Monitoring for Voice and Chat AI Agents
Cekura, a YC F24 startup, launches a simulation platform that lets developers test voice and chat AI agents at scale. The service creates synthetic user conversations, imports real production dialogs, and uses a mock‑tool environment to evaluate agent behavior without hitting...
Show HN: Omni – Open-Source Workplace Search and Chat, Built on Postgres
Omni is an open‑source AI‑powered workplace search and chat platform that unifies full‑text (BM25) and semantic (pgvector) queries across Google Workspace, Slack, Confluence, Jira and other tools. It runs entirely on a single Postgres instance (ParadeDB), eliminating the need for...