Start your morning scandalously informed.

A daily curated digest with the tech news that matter + community vibes, delivered daily, in tabloid style. Like you always wanted.

← Back to Home

Sunday, February 22, 2026

AI Chips Explode, Password Vaults Crack Open!

AI Chips Explode, Password Vaults Crack Open!

AI races ahead while trust in security tools cracks... Chips get faster, gadgets get smarter, and strange new agents whisper from tiny boards... Ancient-style glass promises to hold our memories for ages, as cloud hiccups still shake the internet... Ads run themselves, bosses swear the robots are better, and some workers quietly eye the exit... Young voters dream of a different Europe, gamers revive dead worlds, and privacy fans watch the dark corners of the net crumble... Tonight we stare at the glowing screens and wonder who is really in charge.

AI Boom Speeds Up And Melts Nerves

  • New AI Chip Spits Out Words Like Water

    Taalas shows off a custom ASIC that runs Llama 3.1 8B at wild speed, around 17k tokens per second. It feels like a cheat code for AI hardware, but there is a nagging worry about closed chips, power use and what this does to NVIDIA’s grip on the market.

  • Hacker Runs Giant AI Model On One GPU

    A C++ NTransformer engine streams Llama 3.1 70B through a single RTX 3090, pulling weights from NVMe and even bypassing the CPU. It feels like home‑lab science fiction, but raises doubts about latency, reliability and how far hobbyists will push their aging gaming cards.

  • Tiny Nintendo 64 Becomes Talkative AI Dungeon

    A project called Legend of Elya squeezes a nano GPT model onto real N64 hardware with 4MB RAM and a 93MHz CPU. It turns a retro console into a chatty story machine and quietly mocks how bloated modern AI stacks are compared to careful old‑school engineering.

  • Pocket Chip Runs Its Own Mini AI Butler

    zclaw packs a personal AI assistant into under 888KB on an ESP32, handling schedules and notes while talking to Anthropic or OpenAI in the background. It feels delightfully small and efficient, yet it also suggests a future full of whispering gadgets we barely control.

  • Public Eyes AI Gold Rush With Growing Suspicion

    A New York Times piece says the current AI boom feels more like an exhausting grind than the fun dot‑com bubble. Between job fears, privacy shocks and hype fatigue, it is hard not to see the mood turning from excitement to a wary, tired side‑eye at big AI brands.

Security Shocks And The Hunt For Trust

  • Top Password Vaults Caught With Their Doors Open

    Researchers at ETH Zurich and USI tear into three big cloud password managers and manage to read and even edit stored secrets. The glossy promise of zero‑knowledge vaults suddenly looks thin, and keeping logins safe in a browser tab feels a lot less comforting than it did yesterday.

  • Rampaging Botnet Accidentally Wipes Out Privacy Network

    The I2P anonymity network is flooded by a 700,000‑node Sybil attack, likely from a stray botnet, leaving the system barely usable. It is a grim lesson that one sloppy swarm of infected machines can flatten years of work on privacy, and it makes other dark‑net projects look painfully fragile.

  • Internet Turns Dark Forest As AI Starts Hunting

    An essay argues constant AI scraping, spam and scanners are turning the web into a dark forest, where smart players hide behind logins and blocks. It matches the uneasy feeling that open sites are being drained dry by bots, forcing real people into closed gardens and private chats.

  • New Rulebook Demands Every AI Serve A Human

    The Human Root of Trust framework lays out a public‑domain scheme to tie every autonomous agent action back to a real person using cryptographic chains. It sounds refreshingly human‑centric, yet it is hard to imagine messy corporations and hobby projects cleanly wiring every bot to a named owner.

  • Startup Claims It Can Prove Your AI Is Honest

    Tinfoil proposes cryptographic checks so an inference provider can prove they run the full model, not a sneaky quantized downgrade. The idea scratches a deep trust itch for customers who hate black boxes, but it is easy to suspect the biggest clouds will dodge anything that adds friction.

Old Tech Fights Back With Glass And Batteries

  • Microsoft Etches Data In Glass For 10,000 Years

    Microsoft’s Project Silica uses borosilicate glass slabs to store data that could survive heat, floods and time for at least ten millennia. It feels like a time capsule for streaming ages, yet it raises awkward questions about cost, ownership and which stories deserve to be frozen for future civilizations.

  • EU Orders Phones To Get Easy-Change Batteries

    A sweeping EU rule will require replaceable batteries and strict recycling targets by 2027, tracked with digital passports and QR codes. It sounds like a win for repair fans and the planet, but there is a nagging suspicion phone makers will fight back with thicker cases and sneaky exceptions.

  • Meta’s New Ad Robot Eats Its Own Agencies

    A major buyer says Meta’s AI‑driven ads now work like a black box that swallows their playbook and delivers worse results. It feels like the platform is quietly replacing human expertise with sliders and promises, leaving agencies holding the risk while the algorithm keeps all the control and data.

  • Palantir’s Secret Map For All The World’s Data

    A deep look at Palantir Ontology and open‑source clones shows how modeling real‑world entities lets AI act on company operations like a live digital twin. It is an impressive vision of joined‑up data, but impossible to ignore how neatly it also packages surveillance, lock‑in and enormous power for whoever runs it.

  • Cloudflare Slip Knocks Big Chunks Of Internet Offline

    A Cloudflare post‑mortem dissects a Feb 20 outage tied to BGP routing and BYOIP changes that briefly broke loads of sites. It reinforces the uneasy truth that one company sits under a huge slice of the web, and every misstep sends a sharp reminder of that hidden dependency.

Top Stories

Startup "prints" giant AI model onto a chip

Artificial Intelligence

A new ASIC from Taalas runs Llama 3.1 8B at extreme speed, hinting at a future where powerful AI lives on cheap dedicated chips instead of big GPUs.

Read more →Comments

Password managers caught far less secure than promised

Security

Swiss researchers broke into major cloud password managers, undermining the marketing story around “zero-knowledge” vaults that many everyday users rely on.

Read more →Comments

Botnet swamps and cripples the I2P anonymity network

Cybersecurity

A 700,000-node Sybil attack effectively wrecked the I2P privacy network, showing how fragile alternative anonymous infrastructures can be under botnet pressure.

Read more →Comments

Microsoft claims data storage that survives 10,000 years

Technology

Project Silica’s glass platters aim to preserve data for millennia, raising big questions about what our civilization chooses to save—and who controls it.

Read more →Comments

Public grows tired and wary of the AI boom

Artificial Intelligence

A NYT piece argues today’s AI gold rush feels more exhausting and threatening than the dot‑com era, mirroring mounting backlash in jobs, culture and politics.

Read more →Comments

Meta’s ad automation starts to crush agencies

Business

A big media buyer says Meta’s new AI-driven ad tools are gutting their control and margins, a loud early warning for white‑collar workers in ad tech.

Read more →Comments

The open internet turns into a dark forest

Technology

An essay argues constant AI scraping and automated abuse are forcing people to hide behind walls and logins, making the web feel hostile and closed.

Read more →Comments
All stories (67)
Complete list of news articles from this day

Microsoft team creates 'revolutionary' data storage system that lasts millennia

by gnabgib

Microsoft researchers presented a glass-based archival storage system that encodes data as microscopic deformations inside borosilicate glass using high-energy, ultrashort laser pulses. The informatio...

Key Points

  • Stores 4.8 TB on a borosilicate glass square with millennia-scale durability.
  • Encodes via ultrashort laser pulses; reads via microscope-based optical detection.
  • Immutable, maintenance-free medium designed for long-term archival backup.

Meta Deployed AI and It Is Killing Our Agency

by zenincognito

A digital advertising agency that manages millions of dollars in yearly Meta ad spend reports a persistent problem: newly created work accounts for experienced staff are being banned minutes to hours ...

Key Points

  • Immediate bans of newly created, fully verified Meta work accounts are reported.
  • Appeal via the Account Quality dashboard is blocked by login lockouts.
  • The agency seeks manual onboarding and human support overrides for erroneous bans.

Colorado moves age checks from websites to operating systems

by iamnothere

Colorado’s Senate Bill 26-051 proposes moving online age verification from websites to operating systems and app stores. Under the measure, operating system providers would collect a user’s age or bir...

Key Points

  • SB 26-051 shifts age verification duties to operating systems and app stores via an API-delivered age bracket signal.
  • Previous bills (SB 25-201, SB 25-086) failed or were vetoed due to privacy, feasibility, and constitutional concerns.
  • The new approach seeks a defined compliance point by leveraging vertically integrated mobile platforms.

Lean 4: How the theorem prover works and why it's the new competitive edge in AI

by tesserato

The article examines Lean4 as a means to bring mathematical rigor to AI systems that otherwise exhibit probabilistic behavior and hallucinations. It outlines how Lean4 operates as both a programming l...

Key Points

  • Lean4 enforces mathematical correctness via a trusted kernel and strict type-checking.
  • Its determinism and transparency address LLM unpredictability and hallucinations.
  • Research is combining LLMs with Lean4 (e.g., Safe) to validate AI outputs through formal proofs.

Trunk Based Development

by handfuloflight

The article outlines Trunk-Based Development (TBD), a source-control approach where teams work from a single trunk branch to minimize merge conflicts and keep software always releasable. It positions ...

Key Points

  • TBD keeps a single trunk branch, avoiding long-lived branches to reduce merge risks.
  • CI/CD is enabled by frequent commits and pre-integrate builds on developer environments.
  • Release strategies include JIT release branches or trunk-based releases with fix-forward.

Acme Weather

by cryptoz

Acme Weather, created by the team behind Dark Sky, launches with a mission to better communicate the inherent uncertainty in weather forecasting. Reflecting on Dark Sky’s evolution and its acquisition...

Key Points

  • Acme Weather focuses on visualizing forecast uncertainty with alternate predictions.
  • Forecasts are built from multiple data sources: models, satellite, ground stations, and radar.
  • Community Reporting enables real-time local condition reports to complement forecasts.

Gitas – A tool for Git account switching

by letmutex

Gitas is a cross-platform command-line utility designed to simplify managing multiple Git identities. It offers an interactive TUI for switching, editing, and removing accounts and supports adding new...

Key Points

  • Cross-platform Git identity manager with interactive TUI and GitHub login support.
  • Per-command identity proxy using -c flags without modifying persistent configs.
  • Secure token storage via native OS keychains and clear install/uninstall paths.

EU mandates replaceable batteries by 2027 (2023)

by cyrusmg

The EU’s Batteries Regulation has entered into force, establishing end-to-end rules governing how batteries are sourced, manufactured, used, labeled, collected, reused, and recycled within the Europea...

Key Points

  • User-replaceable portable batteries mandated by 2027 across electronic products.
  • Phased 2025 carbon footprint and recycling targets for EV, light transport, and industrial batteries.
  • QR-coded digital battery passports and due diligence obligations to improve transparency and sustainability.

24 Hour Fitness Won't Let You Unsubscribe From Marketing Spam, So I Fixed It

by daem

The article reports that 24 Hour Fitness’s marketing email unsubscribe page is broken, returning a Spanish-language error tied to OneTrust token retrieval and preventing users from opting out. The aut...

Key Points

  • 24 Hour Fitness’s unsubscribe page is broken due to a JavaScript bug causing a OneTrust token error.
  • The author reported the issue in November 2025 and built an alternative unsubscribe tool.
  • CAN-SPAM requires functional opt-out; past fines show significant enforcement risk.

LibreOffice blasts OnlyOffice for working with Microsoft to lock users in

by XzetaU8

The Document Foundation (TDF), steward of LibreOffice, has criticized OnlyOffice for aligning closely with Microsoft’s ecosystem. TDF describes OnlyOffice as “fake open-source,” arguing that by defaul...

Key Points

  • TDF calls OnlyOffice “fake open-source” for defaulting to Microsoft formats.
  • TDF criticizes ISO’s OOXML standardization and cites ongoing interoperability issues.
  • OnlyOffice promotes compatibility with Microsoft formats and offers community and enterprise editions.

I Verified My LinkedIn Identity. Here's What I Handed Over

by ColinWright

The article details a user’s experience with LinkedIn’s identity verification, revealing that the process is handled by Persona Identities, Inc., a company based in San Francisco. After submitting a p...

Key Points

  • LinkedIn verification is processed by Persona, not LinkedIn.
  • Persona collects extensive biometric, device, and behavioral data.
  • Identity document images are used to train AI under “legitimate interests.”

Understanding Std:Shared_mutex from C++17

by ibobev

This article explains why std::mutex can be overly restrictive for read-heavy data access patterns and how std::shared_mutex, introduced in C++17, addresses this by allowing multiple concurrent reader...

Key Points

  • std::shared_mutex enables concurrent reads and exclusive writes for read-heavy workloads.
  • Replace std::lock_guard<std::mutex> with std::shared_lock/std::unique_lock on std::shared_mutex as appropriate.
  • Observed output illustrates timing and scheduling effects in concurrent programs.

When etcd crashes, check your disks first

by _ananos_

A team built a cloud–edge testbed to demonstrate MLSysOps, a framework for applying telemetry-driven policies that adapt application placement and runtime behavior without manual intervention. The dem...

Key Points

  • Recurring Karmada pod crashes were traced to etcd timing out from slow, inconsistent storage I/O.
  • etcd’s dependency on timely write-ahead log and fsync operations makes it highly sensitive to disk latency.
  • Adjusting timeouts offered limited mitigation; the underlying storage issue remained unresolved.

Andrej Karpathy talks about "Claws"

by helloplanets

Simon Willison’s link post spotlights Andrej Karpathy’s recent tweet-thread coalescing the term “Claws” to describe a new layer atop AI agents. Karpathy explains that while he is cautious about runnin...

Key Points

  • Karpathy defines “Claws” as a coordination layer above AI agents, emphasizing orchestration and persistence.
  • He spotlights NanoClaw’s small codebase and containerized design and lists several other early projects.
  • Willison notes the term “Claw” is gaining adoption for OpenClaw-like, personal-hardware agent systems.

Large Language Model Reasoning Failures

by T-A

This survey examines why Large Language Models (LLMs), despite strong performance across many tasks, still exhibit notable reasoning failures. It introduces a structured, two-axis framework: one axis ...

Key Points

  • Introduces a two-axis framework for LLM reasoning and failure classification.
  • Details definitions, root causes, and mitigations for each failure type.
  • Provides a curated GitHub repository of research on LLM reasoning failures.

Coccinelle: The Linux kernel's source-to-source transformation tool

by anon111332142

The article introduces Coccinelle, a tool designed for style-preserving source-to-source transformations on C code, enabling complex refactoring through semantic patches. After installation, users run...

Key Points

  • Coccinelle performs style-preserving source transformations on C and is run via the spatch script.
  • Users can run Coccinelle without installation by sourcing env scripts; precompiled builds are available.
  • Contributions require DCO 1.1 Signed-off-by lines per Linux kernel guidelines.

The bare minimum for syncing Git repos

by speckx

The article outlines a minimalist approach to synchronizing personal Git repositories across multiple devices without relying on cloud platforms like GitHub. The author notes that alternatives such as...

Key Points

  • A Git remote can be a local or SSH path, not just a hosted service.
  • Bare repositories are the correct targets for pushes; non-bare repos should not receive them.
  • Simple commands (git init --bare, git remote add) implement a local, low-dependency sync setup.

AI uBlock Blacklist

by rdmuser

This article presents a personal uBlock Origin blacklist created to block AI-generated content farms that frequently surface in search results. The author argues these sites often provide mediocre, ad...

Key Points

  • Install the blacklist via subscription or import the GitHub raw URL into uBlock Origin.
  • Contribute sites through GitHub issues or pull requests, with guidance on precise filter scoping.
  • Use provided heuristics and filter syntax to identify and block AI-generated content farms effectively.

Over 80% of 16 to 24-year-olds would vote to rejoin the EU

by saubeidl

An exclusive ITV-commissioned poll by Savanta indicates that 83% of UK residents aged 16 to 24 would vote to rejoin the European Union if a new referendum were held, while 17% would opt to remain outs...

Key Points

  • 83% of 16–24-year-olds would vote to rejoin the EU; 17% would remain outside.
  • PM Keir Starmer seeks deeper economic integration and closer single market alignment in more sectors.
  • Political parties diverge: Reform blames Conservatives; Lib Dems push customs union; Greens campaign to rejoin.

Choose Your Fictions Well (2010)

by 1970-01-01

This 2010 essay by Henry Jenkins and Peter Ludlow examines how competing “fictions” and goals collide in open-ended virtual worlds. Using the 2004 Alphaville elections in The Sims Online as a case stu...

Key Points

  • Open-ended design fosters divergent, often conflicting player narratives.
  • Historical and recent cases show recurring collisions in virtual worlds.
  • Concepts from Gee and Benkler frame why these conflicts persist.

JWasm: Masm Compatible Assembler

by doener

JWasm is a MASM-compatible assembler with broad platform support and explicit build instructions for multiple operating systems. The article outlines how to compile JWasm across Windows, Linux, DOS, a...

Key Points

  • Cross-platform assembler: JWasm runs on Windows, Linux, DOS, and OS/2.
  • Platform-specific makefiles support Visual C++, Open Watcom, GCC, and Clang.
  • DOS support includes a limited 16-bit build for 8088 CPUs, with possible HX requirements for older VC++.

macOS's Little-Known Command-Line Sandboxing Tool

by Igor_Wiwi

This article introduces sandbox-exec, a built-in macOS command-line tool that runs applications in a constrained sandbox, limiting what they can access. It explains why sandboxing is valuable: it prot...

Key Points

  • sandbox-exec runs apps in a restricted environment using profiles
  • Profiles use Scheme/LISP-like syntax with default policy and rules
  • Two approaches: deny-by-default (secure) vs allow-by-default (permissive)

CXMT has been offering DDR4 chips at about half the prevailing market rate

by phront

China’s memory makers are intensifying competition in the global market. CXMT is selling older-generation DDR4 DRAM at roughly half the prevailing price, leveraging a supply shortage that has pushed c...

Key Points

  • CXMT halves DDR4 pricing amid a supply-driven price surge, drawing OEM interest.
  • Korean leaders focus on HBM4 while maintaining large legacy DRAM capacity, creating exposure.
  • YMTC hits 10% NAND share and expands with a new Wuhan fab, half allocated to DRAM.

DJB's Cryptographic Odyssey: From Code Hero to Standards Gadfly

by crypto_watchdog

The article recounts Daniel J. Bernstein’s influence on applied cryptography, beginning with the 2006 debut of Curve25519 and complementary designs ChaCha and Poly1305, consolidated in the NaCl librar...

Key Points

  • TLS 1.3 recommendations and broad platform support propelled X25519/Ed25519 to dominance.
  • ChaCha20‑Poly1305 was standardized for TLS 1.2 (RFC 7905) after advocacy by major deployers.
  • X25519’s permissive input handling and clamping reflect intentional design trade‑offs documented by IETF RFC 7748.

How far back in time can you understand English?

by spzb

This article from The Dead Language Society by linguist Colin Gorrie presents a crafted experiment in readability: a fictional travel blog about a visit to the coastal town of Wulfleet that gradually ...

Key Points

  • The blog post’s language shifts back by centuries, demonstrating English’s evolution.
  • The passages are constructed to be realistic for each historical period.
  • The newsletter provides guides and book clubs for reading historical English texts.

What Not to Write on Your Security Clearance Form

by wizardforhire

Les Earnest recounts an incident from 1943 in which a homemade cipher key led to unexpected scrutiny. After reading Fletcher Pratt’s 1942 book “Secret and Urgent,” which explained letter-frequency ana...

Key Points

  • A lost glasses case containing a cipher key prompted an FBI inquiry in 1943.
  • The code was inspired by Fletcher Pratt’s book on cryptography and letter-frequency analysis.
  • The resulting FBI record later affected Earnest’s security clearance process.

Show HN: Iron-Wolf – Wolfenstein 3D source port in Rust

by ragnaroekX

Iron Wolf is a Rust-based source port focused on delivering a pixel-perfect and mod-friendly recreation of Wolfenstein 3D. The project provides a practical, ready-to-run experience: users can clone th...

Key Points

  • Rust-based source port aiming for pixel-perfect, mod-friendly Wolfenstein 3D recreation
  • Immediate local run using included shareware data and run-sdl-shareware command
  • Web version available; upload full game files to play the complete version in-browser

Permacomputing

by tosh

The article introduces permacomputing, a sustainability-focused approach that extends hardware lifespans, minimizes energy consumption, and emphasizes using already available computational resources. ...

Key Points

  • Prioritize planned longevity and maintenance to reduce energy use and extend hardware life.
  • Design products for disassembly using fasteners, material labeling, and common tools.
  • Build malleable, recombinable systems and favor computation only when necessary.

Claws are now a new layer on top of LLM agents

by Cyphase

The article presents “Claws” as an emerging layer above LLM agents that advances orchestration, scheduling, context handling, tool invocation, and persistence, positioning them as a new tier in the AI...

Key Points

  • Claws are framed as a new orchestration and persistence layer over LLM agents.
  • Security concerns are emphasized for large implementations like OpenClaw.
  • NanoClaw is highlighted for small codebase, containerization, and skill-based configuration.

DialUp95 – A 90s inspired nostalgia hit

by robputt

DialUp95 is presented as a software experience designed to evoke the look and feel of 1990s dial-up Internet connections. The interface closely mirrors classic Windows 95 connection dialogs, featuring...

Key Points

  • DialUp95 simulates 1990s dial-up interfaces.
  • Experience Switcher includes Windows 95 default and AOL 4 modes.
  • UI replicates standard fields and actions from legacy connection dialogs.

Loon: A functional lang with invisible types, safe ownership, and alg. effects

by surprisetalk

The article introduces Loon, a functional programming language designed around uniform, expression-only syntax where every construct yields a value. Loon’s brackets-based notation aims to keep code co...

Key Points

  • Expression-only design with uniform bracket syntax and pipe-based transformations.
  • Complete Hindley–Milner inference (let-polymorphism, row types) with no annotations required.
  • Algebraic effects for composable side-effect handling, replacing exceptions and async/await.

Denonomicon: The Dark Arts of Deno Foreign Function Interface Programming

by enz

“Denonomicon” is presented as an unofficial, third-party guide dedicated to Deno’s Foreign Function Interface (FFI). The introduction emphasizes that it is not affiliated with the Deno team and may be...

Key Points

  • Unofficial guide focused specifically on Deno’s FFI.
  • Targets interfacing Deno with native C APIs and warns of pitfalls.
  • Directs general Deno programming queries to the official Deno Manual.

Cloudflare outage on February 20, 2026

by nomaxx117

Cloudflare disclosed a service outage on February 20, 2026, lasting 6 hours and 7 minutes, that affected a subset of customers using its Bring Your Own IP (BYOIP) service. The incident began at 17:48 ...

Key Points

  • Outage caused by BYOIP configuration change, not a cyberattack.
  • ~1,100 prefixes withdrawn; partial and manual restorations completed by 23:03 UTC.
  • Guidance enabled customer self-remediation via Cloudflare dashboard at 19:19 UTC.

Canvas_ity: A tiny, single-header <canvas>-like 2D rasterizer for C++

by PaulHoule

Canvas_ity is a compact, single-header C++ library for immediate‑mode 2D vector graphics that closely follows the W3C HTML5 2D canvas specification. Designed with a focus on high-quality rendering, ea...

Key Points

  • Single-header C++ library modeled on W3C HTML5 2D canvas
  • Implements most features; omits hit regions and some property getters
  • Example shows full pipeline from path ops to TGA output with gamma-correct blending

Personal Statement of a CIA Analyst

by grubbs

This first-person statement from 4 October 2018 recounts a CIA analyst’s initial polygraph experience during the Agency’s applicant screening. Already holding CIA TS/SCI clearances as a defense contra...

Key Points

  • The author prepared via Lykken’s polygraph research book but aimed to be truthful, not use countermeasures.
  • The CIA polygraph occurred over two days in Northern Virginia, with consent and non-disclosure instructions.
  • She failed after scrutiny of a “lied to your boss” question and was told to return the next day.

Inputlag.science – Repository of knowledge about input lag in gaming

by akyuu

Inputlag.science introduces itself as a centralized resource dedicated to understanding and addressing input lag in gaming. The site defines input lag as the delay between a user’s input and the resul...

Key Points

  • Defines input lag and frames it as a growing industry issue.
  • Attributes increased latency to system complexity and knowledge gaps.
  • Focuses on a three-part lag chain and precise measurement methods.

Parse, Don't Validate and Type-Driven Design in Rust

by todsacerdoti

This article presents the “parse, don’t validate” approach through a Rust-focused lens. Instead of relying on runtime checks or returning fallible results for invalid inputs, it recommends encoding in...

Key Points

  • Encode invariants in Rust types instead of relying on runtime validation.
  • Use a NonZeroF32 newtype with a fallible constructor to forbid zero values.
  • Strengthen parameter types rather than weakening return types to Option/Result.

Toyota Mirai hydrogen car depreciation: 65% value loss in a year

by iancmceachern

The article analyzes the steep depreciation of Toyota’s hydrogen fuel cell sedan, the Mirai, in the U.S. used market. It traces the model’s evolution from its first generation, launched for the 2015 m...

Key Points

  • Second-gen Mirai improved range by increasing hydrogen capacity and added ADAS/safety and a 2023 infotainment update.
  • Used second-gen Mirai prices now range from under $10,000 to about $22,000; last year’s models list for $15,000–$18,000.
  • Limited California-only availability and insufficient hydrogen infrastructure are key drivers of the Mirai’s depreciation.

How an inference provider can prove they're not serving a quantized model

by FrasiertheLion

The article addresses a common concern in AI inference: clients cannot easily verify that an API provider is serving the exact model and configuration advertised. It cites variability in evaluations a...

Key Points

  • Modelwrap cryptographically binds runtime model weights to a public commitment.
  • Boot-time attestation is extended with dm-verity to enforce integrity on every read.
  • Merkle trees provide a compact commitment for large weight files.

I Don't Like Magic

by edent

A front-end developer outlines a philosophy of minimizing “magic”—opaque abstractions presented as effortless—arguing that such approaches reduce agency and can harm users through bloat and performanc...

Key Points

  • Avoids frameworks like React due to added abstraction and performance concerns.
  • Uses minimal, understood dependencies (two JS libraries) and avoids npm for client-side.
  • Urges review of AI-generated code before deployment.

zclaw: personal AI assistant in under 888 KB, running on an ESP32

by tosh

zclaw is an open-source (MIT) personal AI assistant firmware written in C for ESP32 microcontrollers, designed to fit within a ≤888 KB default build. It provides timezone-aware task scheduling, GPIO r...

Key Points

  • ≤888 KB C firmware for ESP32 with scheduling, GPIO guardrails, and persistent memory
  • Chat via Telegram or hosted web relay; supports Anthropic, OpenAI, OpenRouter
  • Tested on ESP32-C3/S3/C6; scripts for install, secure flash, provisioning, emulation, and benchmarking

EDuke32 – Duke Nukem 3D (Open-Source)

by reconnecting

EDuke32 is a free, open-source engine and source port for Duke Nukem 3D that brings the classic FPS to modern systems with extensive enhancements. It runs natively on Windows, Linux, macOS, and FreeBS...

Key Points

  • Open-source, multi-platform engine and source port for Duke Nukem 3D.
  • Advanced Polymer renderer with real-time lighting and shadow mapping.
  • Includes VoidSW (Shadow Warrior port) and powers Ion Fury by Voidpoint.

The Software Development Lifecycle Is Dead

by zenon_paradox

The piece argues that AI coding agents are transforming software development by collapsing the traditional, stage-based software development lifecycle into an integrated, iterative process. Instead of...

Key Points

  • AI agents merge SDLC stages into a continuous, iterative workflow.
  • Requirements become iterative artifacts, reducing reliance on stage-based tools like Jira.
  • System design shifts to real-time discovery with agent collaboration.

MeshTNC is a tool for turning consumer grade LoRa radios into KISS TNC compatib

by todsacerdoti

MeshTNC is an open-source firmware and tool that connects LoRa data to consumer-grade radios while adding KISS TNC compatibility. Hosted on GitHub, it offers precompiled releases and supports flashing...

Key Points

  • Adds KISS TNC compatibility to consumer-grade LoRa radios via an open-source firmware.
  • Provides a serial CLI for raw LoRa transmission, radio configuration, and LoRa/BLE logging.
  • Supports APRS over LoRa and works with clients like xastir and APRSISCE32.

Online Pebble Development

by teekert

The article introduces a web-based environment for creating Pebble applications that removes the need for traditional local setup. Developers can write apps directly in their browser without installin...

Key Points

  • Zero-install, browser-based Pebble app development
  • No need for Linux, virtual machines, compilers, or Python
  • Resource management via a user-friendly online interface

Beagle CRDT SCM outer interface

by gritzko

This article outlines Beagle SCM, a source control system designed to reduce interface complexity and enhance code management by treating code as a database. It critiques Git’s layered model and expan...

Key Points

  • Beagle stores ASTs to enable fine-grained operations beyond file-level blobs.
  • CRDT merges are deterministic and non-intrusive, supporting flexible workflows.
  • Interface uses HTTP verbs and URI addressing to simplify commands and references.

Mathematics in the Library of Babel

by robinhouston

This essay charts the evolving role of large language models (LLMs) in mathematical research from early trials in 2020 to present-day tools capable of producing useful proofs. Starting with GPT-3 via ...

Key Points

  • Reasoning models and ChatGPT 5.2 Pro now often produce useful, routine mathematical proofs, though errors persist.
  • Empirical evaluations like “First Proof” test current model utility on real lemmas.
  • Updated timelines reflect faster-than-expected progress, with a public bet likely to be lost.

When Code is Free, Why is Claude is an Electron app?

by dbreunig

The article explores why Electron remains a prevalent choice for desktop applications, using Anthropic’s Claude desktop client as a case study despite the rise of AI coding agents. It explains Electro...

Key Points

  • Electron offers a single codebase and reuse of web tech, aiding cross‑platform delivery.
  • Agentic coding showed promise but struggled with the last 10% in an Anthropic Rust–C compiler project.
  • Maintaining multiple native apps increases support burden; Electron remains practical for now.

Password managers less secure than promised

by mono-bob

ETH Zurich’s Applied Cryptography Group uncovered serious security weaknesses in three major cloud-based password managers—Bitwarden, LastPass, and Dashlane—services used by an estimated 60 million us...

Key Points

  • Researchers showed that malicious servers could view and modify passwords in Bitwarden, LastPass, and Dashlane.
  • Twelve, seven, and six attacks were demonstrated against Bitwarden, LastPass, and Dashlane respectively.
  • Findings undermine providers’ “zero-knowledge encryption” assurances and highlight gaps in commercial end-to-end encryption.

Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU

by xaskasdf

NTransformer is a C++/CUDA LLM inference engine built to run large GGUF-format models on consumer GPUs by streaming weights across a three-tier memory hierarchy. Demonstrated on an NVIDIA RTX 3090, it...

Key Points

  • Runs Llama 3.1 70B on a single RTX 3090 by tiering and streaming layers, with optional NVMe-to-GPU direct I/O.
  • Achieves ~0.2 tok/s for 70B (33× over mmap baseline) and ~49 tok/s for 8B fully resident.
  • PCIe bandwidth is the limiting factor; Gen4 x16 is expected to increase throughput to ~0.5 tok/s.

Happy Zelda's 40th first LLM running on N64 hardware (4MB RAM, 93MHz)

by AutoJanitor

Legend of Elya is an original Nintendo 64 homebrew title that runs a tiny character-level language model entirely on the console’s 93.75 MHz NEC VR4300 (MIPS III) CPU with 4 MB of RDRAM. To bypass the...

Key Points

  • First-of-its-kind N64 homebrew running a live nano-GPT model fully on-console with fixed-point math
  • Compact Q4-quantized 232 KB weight file enables ~427k-parameter transformer to run within 4 MB RDRAM
  • Roadmap plans RSP vector acceleration for an estimated 4–8× performance boost

Declarative, Inquisitive, then Imperative (2017) [pdf]

by tosh

This 2017 note by Samuel A. Falvo II explores Rich Hickey’s “Simple Made Easy” distinctions—simple vs. complex and easy vs. hard—and applies them to the Forth programming language. The author argues F...

Key Points

  • Hickey’s simple vs. easy framework is used to analyze Forth.
  • Forth’s minimalism is simple but can hide coupled concerns.
  • Forth is easy to use yet has hard aspects like global state and typed operators.

Are compilers deterministic?

by fragmede

The article analyzes whether compilers are deterministic by separating theoretical and practical viewpoints. The theoretical stance asserts determinism when considering the compiler as a function of i...

Key Points

  • Deterministic in theory (full inputs), nondeterministic in practice due to uncontrolled variables.
  • Compiler contract preserves semantics, not byte-for-byte identity.
  • Multiple entropy sources in build processes undermine reproducibility.

How I use Claude Code: Separation of planning and execution

by vinhnx

This article presents a disciplined approach to using Claude Code that prioritizes planning before coding. The author structures tasks into a repeatable pipeline: conducting deep research on the codeb...

Key Points

  • Require a reviewed plan (plan.md) before any code is generated.
  • Conduct deep research and record findings in research.md as a review artifact.
  • Base plans on the actual codebase with explicit file paths, snippets, and trade-offs.

Forward propagation of errors through time

by iNic

This article explores whether the backward-in-time computation of gradients in recurrent neural networks (RNNs) is strictly necessary. It derives an exact gradient-based method that propagates error s...

Key Points

  • An exact forward-in-time error propagation algorithm for RNNs is derived and demonstrated.
  • A warm-up phase enables reconstruction of exact gradients without reverse-time traversal.
  • Severe numerical instability in forgetting regimes limits practical adoption despite theoretical soundness.

A Botnet Accidentally Destroyed I2P

by Cider9986

On February 3, 2026, the I2P anonymity network was overwhelmed by a large-scale Sybil attack that introduced roughly 700,000 hostile nodes, dwarfing its usual 15,000–20,000 active devices. While simil...

Key Points

  • Kimwolf IoT botnet flooded I2P with ~700,000 nodes, overwhelming its 15k–20k usual size.
  • I2P shipped v2.11.0 within six days, enabling default hybrid post-quantum encryption (ML-KEM + X25519).
  • Past February attacks (2023–2024) used malicious floodfill routers and remain unattributed.

Red Robin Died by Spreadsheet. Don't Make the Same Mistake

by mooreds

A post by Garry Tan uses Red Robin and Chili’s as a case study in operational strategy and market outcomes. The article states Red Robin’s stock fell roughly 96% (from around $92 to $3.61) after manag...

Key Points

  • Red Robin’s stock reportedly fell ~96% after service staff cuts and operational changes.
  • Operational metrics worsened (walkaways up 85%, peak-hour losses, slower ticket times).
  • Chili’s saw growth after menu simplification, operational investment, and a $10.99 offer amplified on TikTok.

Palantir's secret weapon isn't AI – it's Ontology. An open-source deep dive

by leading-AI

This open-source book project examines Palantir’s core concept of “Ontology” as the operational layer that connects data and AI. Rather than framing Ontology as a narrow IT component, the work positio...

Key Points

  • Ontology is presented as an operational digital twin unifying objects and actions.
  • Governance of real-world changes is handled via branching and review.
  • Open-source project by Satoshi Yamauchi (Leading.AI) invites contributions under CC BY 4.0.

The Internet Is Becoming a Dark Forest – and AI Is the Hunter

by windcbf

The article contends that AI has transformed cybersecurity by enabling autonomous, machine-speed offensive and defensive operations. It spotlights PentAGI, an open-source agent that automates full pen...

Key Points

  • PentAGI democratizes autonomous penetration testing with broad tool orchestration and multi-LLM support.
  • Anthropic’s Claude Opus 4.6 uncovered 500+ high-severity bugs in widely used open-source projects.
  • The article proposes “Zero Visibility” to eliminate pre-authentication exposure beyond traditional Zero Trust.

Japanese Woodblock Print Search

by curmudgeon22

Ukiyo-e Search is an online platform for discovering Japanese woodblock prints across multiple museum and university collections. It offers a search-by-image feature that lets users upload a photo or ...

Key Points

  • Search-by-image across multiple collections (upload or paste URL).
  • Catalog of 223,891 prints organized by historical periods and artists.
  • Planned enhancements: better data, more images, improved search; sign-up available.

Two Bits Are Better Than One: making bloom filters 2x more accurate

by matheusalmeida

Floe describes how it employs Bloom filters to accelerate SQL queries in its lakehouse engine, asserting that its approach halves false positives. The article reviews Bloom filter fundamentals—fast me...

Key Points

  • Bloom filters are pushed down to the storage engine for first-pass filtering before decompression.
  • Filtering adaptively toggles based on observed skip statistics to balance overhead and benefit.
  • In low-match joins, early filtering yields large decompression savings on the probe side.

People Loved the Dot-Com Boom. The A.I. Boom, Not So Much

by 1vuio0pswjnm7

This article reports a growing gap between Silicon Valley’s sweeping promises for artificial intelligence and a lukewarm public response, a contrast to the dot‑com era’s enthusiasm. Industry leaders h...

Key Points

  • Public enthusiasm for A.I. lags behind the dot‑com era’s excitement.
  • Sam Altman says A.I. adoption is slower than expected.
  • Leaders’ sweeping claims contrast with uncertainty about near-term utility.

Gamedate – A site to revive dead multiplayer games

by msuniverse2026

GameDate is a lightweight web platform aimed at reviving multiplayer activity for legacy and niche titles. It enables players to find and join sessions for “dead games,” niche mods, and retro netplay—...

Key Points

  • Account-free platform to find or host sessions for dead games, niche mods, and retro netplay
  • Sections for “Trending,” “In Demand,” and “All Sessions,” with a snapshot showing zero active sessions
  • Credits MIT-licensed assets vgui.css and cursor-effects, plus standard site links (privacy, rules, catalog)

“Playmakers,” reviewed: The race to give every child a toy

by fortran77

This review examines Michael Kimmel’s “Playmakers,” which explores Jewish-American entrepreneurs’ role in building the U.S. toy industry. It centers on the origin of the Teddy bear, created by Brookly...

Key Points

  • Morris Michtom’s 1902 Teddy bear, inspired by Theodore Roosevelt’s hunting incident, sparked a nationwide craze.
  • Clifford Berryman’s cartoon amplified the event, driving demand and cultural impact.
  • Michael Kimmel’s “Playmakers” argues the Teddy bear helped create the modern U.S. toy market, highlighting Jewish-American entrepreneurs.

Carelessness versus Craftsmanship in Cryptography

by ingve

Trail of Bits details a critical misuse pattern in two widely adopted AES libraries, aes-js (JavaScript) and pyaes (Python). Both libraries default to a fixed initialization vector (IV) when AES is us...

Key Points

  • aes-js and pyaes default to a fixed IV in AES-CTR, enabling key/IV reuse risks.
  • Reused key/IV in CTR or GCM exposes XOR of plaintexts, undermining security.
  • strongSwan responded constructively; the libraries lack AES-GCM/GCM-SIV support.

How Taalas "prints" LLM onto a chip?

by beAroundHere

Taalas, a 2.5‑year‑old startup, has built a fixed‑function ASIC that runs Llama 3.1 8B by hardwiring the model’s weights into silicon. The article reports a throughput of roughly 17,000 tokens per sec...

Key Points

  • Hardwired Llama 3.1 8B ASIC reports ~17,000 tokens/sec and ~10× gains in cost, power, and speed.
  • Streaming architecture avoids external DRAM/HBM; on‑chip SRAM holds KV cache and LoRA adapters.
  • Model mapping via top‑mask customization took ~2 months; the chip is single‑model and non‑rewritable.

The Human Root of Trust – public domain framework for agent accountability

by 3du4rd0v3g4

The Human Root of Trust is a public-domain framework proposing cryptographic accountability for autonomous agent systems. It centers on the principle that every agent must trace to a human, addressing...

Key Points

  • Principle: every autonomous agent must trace to a human.
  • Whitepaper v1.0 (Feb 2026) details the trust chain and dual-path architecture and is public-domain.
  • Open invitation for community to formalize and implement the framework.
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.