Business-agent trial ends with no growth and a costly lesson
Bottleneck Labs says its 24-hour experiment exposed a mix of useful engineering behavior, brittle tooling and increasingly risky growth tactics.
Read report »Technology
1106 reports.
Gemini Robotics 2 brings whole body intelligence to robots
Google DeepMind introduced Gemini Robotics 2 as an intelligence layer for adaptable machines. The company says the system supports whole-body movement, dexterous object handling, cooperation between robots, local execution and adaptation across different robot bodies.
Read report »
Read this before you buy that TV streaming stick
A security investigation described generic Android television boxes that allegedly misrepresented themselves as phones and generated fraudulent advertising activity. Researchers traced telemetry and common installed applications across many devices, reinforcing longstanding warnings about cheap streaming hardware.
Read report »
Stacked PRs are now live on GitHub
GitHub placed stacked pull requests in public preview, adding an ordered view of dependent changes and a way to merge a stack together. The feature works with existing checks and review requirements and is exposed across GitHub interfaces and tooling.
Read report »
UEFA and its national associations will not participate in FIFA competitions
UEFA and its 55 national associations jointly rejected a FIFA proposal involving private ownership interests in the World Cup and other competitions. Their statement said they would not participate in FIFA competitions under the proposed arrangement.
Read report »
Cyberattack disrupts more than 30 Minnesota community water systems
State and federal authorities investigated a coordinated intrusion whose methods resembled earlier Iran-aligned activity, while stressing that attribution remained preliminary and drinking water appeared safe.
Read report »
Researcher details self-propagating prompt injection in Copilot for Word
A coordinated disclosure describes how hidden instructions can pass from an attacker-controlled document into newly generated files and trigger again downstream.
Read report »
Keychron plans open-source firmware platform for gaming mice
The ZGM project is scheduled to debut with the G6 HE in early 2027 and proposes modular support for sensors, buttons, lighting and connectivity.
Read report »
Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
TurboFieldfare is a model-specific Swift and Metal runtime intended to run the instruction-tuned Gemma 4 26B-A4B on Apple-silicon Macs with about 2 GB of memory. It retains a shared core and cache in memory while streaming selected experts from solid-state storage.
Read report »
The coolest use for the Vision Pro
Developer Christian Selig demonstrated a Vision Pro use case centered on spatial computing. The account frames the headset’s value around an experience enabled by its display, tracking and mixed-reality presentation rather than ordinary flat-screen application use.
Read report »
Task-trained 9B model tops frontier systems in catalog-review test
FermiSense reports that a reinforcement-learning fine-tune exceeded its strongest prompted frontier baseline while sharply reducing per-decision inference costs.
Read report »
OpenAI releases Codex Security CLI and TypeScript toolkit
The open-source package can scan, validate and fix code vulnerabilities, with options for CI, containerized multi-repository work and persistent findings.
Read report »
Landmark distances offer a faster guide for A* pathfinding
An interactive technical guide shows how differential heuristics can use precomputed map structure to reduce wasted search without changing the A* algorithm itself.
Read report »
Kimi K3 scales a linear-model design to 2.8 trillion parameters
An architecture review highlights LatentMoE, delta attention, cross-layer residual weighting and a complete move away from positional embeddings.
Read report »
Backchannel Brings Multi-Network Discussions Alongside Web Pages
The single-file userscript combines conversations from selected online communities, adds in-page annotations and keeps personal notes and settings on the user's device.
Read report »
New HIV vaccine shows unprecedented success in preclinical study
La Jolla Institute for Immunology reported unusually strong preclinical results for an experimental HIV vaccine strategy. The work remains preclinical, meaning the evidence concerns laboratory or animal-stage research rather than demonstrated protection in people.
Read report »
Substack writers, you need a website
Writer Elizabeth Tai urged newsletter authors to maintain an independent website alongside Substack. The argument centers on control over identity, archives and audience access instead of relying entirely on a single publishing platform.
Read report »
Nova Scotia court acquits man after underscore error exposed mistaken identity
Brandon Klayme served an 18-month sentence after police sought records for a one-underscore Kik username instead of the two-underscore account linked to an investigation.
Read report »
Small SlopCodeBench trial finds low strict pass rates for Claude models
A 17-checkpoint experiment gave Opus 5 the best result among three tested models, but none completed any selected challenge without accumulated defects.
Read report »
Court narrows Google’s DMCA claims in search-scraping dispute
A judge ruled that anti-circumvention claims cannot proceed where Google’s SearchGuard controls access to search results containing no copyright-protected material.
Read report »
Former Eyeline executive alleges medical disclosure contributed to firing
Kevin Baillie claims in a lawsuit that a retreat discussion of prescribed ketamine treatment became part of an investigation that preceded his dismissal.
Read report »
Kimi K3 Model Arrives on Hugging Face
Moonshot AI published Kimi-K3 materials through Hugging Face, giving developers a distribution point for the model and its accompanying technical information. The listing is a primary project artifact rather than an independent evaluation of capability.
Read report »
Our position on open-weights models
Anthropic said it does not support blanket bans on open-weight AI models and described less-capable open models as a public good. It nevertheless argued for targeted measures around advanced capabilities, distillation and national-security risks.
Read report »
Southwest France wildfire creates country’s first reported pyrocumulonimbus
The fire-generated storm cloud is producing shifting winds and lightning, forcing crews into a defensive strategy, a national firefighters’ spokesman said.
Read report »