Trump denies offering US weapons to China, contradicts ambassador’s account

Jeff: Lightweight 0.8B Decision Models Trained Locally with 30ms Inference

A developer has released Jeff, an open-source project featuring small 0.8-billion-parameter decision-focused language models. The models are designed to be compatible with the Jev specification and can be trained on consumer hardware at home. Inference runs at approximately 30 milliseconds, making the models suitable for fast, lightweight decision-making tasks. The project is publicly available on GitHub under the repository firelex/jeff. It has garnered early attention on Hacker News, though community discussion is still limited.

BitNet b1.58 Brings 100B-Parameter AI Inference to Ordinary CPUs

A ternary neural network architecture called BitNet b1.58 reached broad scientific adoption in late September 2026, marking a significant shift in how large AI models are run. The approach encodes neural network weights using only three discrete values — −1, 0, and +1 — requiring roughly 1.58 bits per parameter, which eliminates the need for floating-point matrix multiplication during inference. This means computations reduce to simple integer additions and subtractions, drastically cutting hardware demands. The open-source framework bitnet.cpp has since been stabilized to support mainstream CPUs including Intel Xeon, AMD EPYC, Apple Silicon, AWS Graviton, and edge chips like Snapdragon and RISC-V. Models with up to 100 billion parameters can now run on commercial servers without expensive GPU clusters, potentially democratizing access to large-scale AI inference.

FlashAttention-4 targets memory bottlenecks in NVIDIA Blackwell B200 GPUs

FlashAttention-4 (FA4), developed by Tri Dao's lab in close collaboration with NVIDIA, has been released as a production specification optimised for the Blackwell B200 and GB200 accelerator architecture. The update addresses critical performance bottlenecks that emerged as large language models such as DeepSeek 4.1 and GPT-6 Astra began operating with context windows exceeding one million tokens. Earlier versions of the algorithm suffered synchronisation stalls and memory bandwidth saturation on Hopper-generation hardware, wasting up to 32% of available clock cycles. FA4 resolves this by introducing Asymmetric Kernel Pipelining via a hardware Tensor Memory Accelerator, splitting workloads between dedicated Producer and Consumer Warps running asynchronously. The release also adds native support for 4-bit quantisation formats FP4 and MXFP4, enabling fuller utilisation of the B200's 20 PFLOPS FP4 compute capacity and 8 TB/s memory bandwidth.

Unity Beginners Guide: Build Educational Games by Prioritizing Lessons Over Mechanics

A guide published on DEV Community outlines a pedagogy-first approach to building educational games in Unity, aimed at beginner developers. Lead Systems Architect Abdulrahman Maslmany advises starting with a single, measurable learning objective before designing any gameplay mechanics. The recommended core loop follows four steps: presenting a question, capturing player input, delivering immediate feedback, and offering an explanation. Curriculum content should be stored in decoupled C# ScriptableObjects rather than hardcoded into scripts, making it easier to update. Early playtesting with real students is emphasized to identify confusion before any visual polish or complex systems are added.

Web Project Showcases Highlights of British Design History

A new web project titled 'Best of British Design' has been published online, curating notable examples from British design history. The project is hosted on Vercel and appears to be an independent or personal initiative rather than an official publication. It was shared on Hacker News, where it received a small number of points and comments. The submission attracted limited but present community engagement on the platform. Further details about the project's scope, authors, or methodology are not available from the source provided.

I Built ResolveIQ to Remember What Customer Support Already Tried

╔══════════════════════════════════════════════════════════════╗ https://resolveiq-zrybe6cvazztu3cnobbayk.streamlit.app/ ║ ║ ║ ║ The example customer case demonstrates the central idea: ║ ║ remember what already happened, understand the recurrence, ║ ║ and give the support agent context before the next ║ ║ troubleshooting step. ║ ║ ║ ║ The most important lesson I took from building ResolveIQ is ║ ║ that memory is valuable when it changes what an agent does ║ ║ next. Persistent context is not the destination; it is the ║ ║ missing input that makes the next support decision more ║ ║ informed. ║

World Labs Is Joining AMD

Article URL: https://www.worldlabs.ai/blog/amd-announcement Comments URL: https://news.ycombinator.com/item?id=49883760 Points: 9 # Comments: 1

FBI lists gangster Goldy Brar among 10 most wanted fugitives, declares $1 million bounty

SupportMind AI Agent Uses Per-Customer Memory to Skip Repetitive Troubleshooting

A developer named JayaLakshmi built SupportMind, an AI-powered customer support agent that retains individual conversation histories to avoid making users repeat themselves. The system assigns each customer a separate memory bank identified by their customer ID, ensuring one person's issue history never appears in another's session. Built using Flask, Groq's language model, and a memory tool called Hindsight, the agent recalls relevant past interactions before generating a reply and saves each new exchange afterward. A returning customer with a known fix receives a tailored response, while a first-time customer with no history gets a standard troubleshooting checklist — same model, same prompt, different memory. The project was developed on the Code.in platform and demonstrates how persistent, per-user memory can meaningfully improve automated support interactions.

Data Debunks Seven Common AI Search Myths, Core SEO Still Matters

Search Engine Land has analysed seven widely circulated claims about AI and search, testing them against observed data rather than accepting them as fact. The report challenges assertions such as AI killing traditional search and the idea that adding an llms.txt file guarantees visibility in AI-generated results. Its key finding is that AI is reshaping how content is discovered, but it has not made fundamental SEO signals like quality content, brand authority, and credible sourcing irrelevant. Businesses are advised to treat AI-related recommendations as testable hypotheses and connect visibility efforts to tangible outcomes like traffic and conversions. The report cautions against both ignoring AI search developments and overhauling entire content strategies based on unproven claims.

Rahul Gandhi Raises Alarm Over 727 Voter Deletions in Delhi Rolls

Congress leader Rahul Gandhi has raised concerns over the alleged exclusion of 727 voters from Delhi's electoral rolls, posting his objections on social media platform X. He directed criticism at Chief Election Commissioner Gyanesh Kumar, warning that the broader system enabling such actions would also be held accountable. Gandhi's remarks suggest he views the deletions as part of a systemic issue rather than an isolated incident. The Election Commission of India has responded to the allegations, though its detailed explanation remains under scrutiny. The controversy adds to ongoing tensions between opposition parties and the poll body over electoral roll integrity.

How Tailscale Solves the Laravel Herd Local API Access Problem on Real Phones

Developers using Laravel Herd locally often find that physical mobile devices cannot reach their .test API endpoints due to three distinct issues: DNS resolution failure, loopback-only routing, and untrusted TLS certificates. Common workarounds like LAN IP binding or public tunnels such as ngrok introduce their own drawbacks, including network dependency, changing URLs, and unintended public exposure of development environments. Tailscale, a WireGuard-based mesh VPN, addresses all three problems by assigning each enrolled device a stable private address and DNS name while keeping traffic off the public internet. It can also terminate TLS using a publicly trusted certificate, eliminating the need to manually install custom CA profiles on test devices. The approach is specifically designed for physical devices connecting to a Laravel Herd site, as emulators and simulators operate under different networking models.

SEBI Clears Adani Group Companies of Minimum Public Shareholding Violations

India's markets regulator SEBI has cleared several Adani Group companies of charges related to minimum public shareholding violations. The clearance applies to Adani Enterprises Ltd, Adani Power Ltd, Adani Ports and Special Economic Zone Ltd, and Adani Transmission Ltd, now rebranded as Adani Energy Solutions Ltd. The directors of these companies have also been absolved in the ruling. The decision marks a significant regulatory relief for the conglomerate, which had faced scrutiny over its shareholding structure.

Tutorial: Build a Multiplayer Browser Game with TypeScript, Phaser, and MongoDB

A new developer tutorial published on DEV Community walks through building a real-time multiplayer browser game using TypeScript, Phaser, Socket.io, and MongoDB. The game places two players on a canvas for 30 seconds, where they compete to collect the most randomly spawned coins, with all game logic managed server-side via WebSockets. MongoDB stores player accounts, match records, and leaderboard data, leveraging its flexible document model and atomic operators to handle concurrent updates without complex application-level coordination. The backend is built with Express, Socket.io, Zod, and the official MongoDB Node.js driver, while the frontend uses Phaser and Vite, with the two sides communicating only at runtime over HTTP and WebSockets. The full source code is available on GitHub, and readers need Node.js 18+, a MongoDB instance, and two browser tabs to follow along.

Fire at Russian Fireworks Factory in Yaroslavl Kills 14

A fire at a fireworks factory in Russia's Yaroslavl region has killed 14 people and left one person injured. Russia's emergency ministry confirmed the death toll, which was reported by the state-run TASS news agency. The incident occurred at a facility involved in fireworks manufacturing, making it particularly dangerous due to the flammable materials on site. Authorities have not yet publicly disclosed the cause of the blaze or whether any investigation has been launched.

How a Single @Transactional Annotation Can Crash Your Spring Boot App Under Load

A common but dangerous pattern in Spring Boot applications involves placing the @Transactional annotation on methods that also make external HTTP calls, such as payment gateway requests. When this happens, a database connection is borrowed from the HikariCP pool at the start of the transaction and held open for the entire duration, including the time spent waiting for the third-party API to respond. Since external calls can take seconds, even moderate traffic can exhaust all available connections in the pool, causing new requests to time out and the application to become unresponsive. The outage can appear misleading because CPU, memory, and database metrics all look healthy, masking the real bottleneck. Developers can fix this by using programmatic transaction boundaries, domain events with @TransactionalEventListener, or the Transactional Outbox pattern to ensure database connections are not held during external I/O.

Gisting: How AI Agents Can Compress Long Contexts Without Losing Key Information

AI coding agents often accumulate tens of thousands of tokens across many reasoning turns, yet only a fraction of that history is relevant to completing the task at hand. This inefficiency prompted researchers at Stanford and Meta to introduce 'gist tokens' in a 2023 NeurIPS paper, a technique that compresses lengthy prompts into a small set of learned internal representations. Unlike standard summarization, which produces natural-language text, gisting trains the model to encode essential information into special bottleneck tokens that future reasoning can draw upon. Applied to agents, the approach shifts context management from preserving raw conversational history to maintaining compact, task-relevant state. The core engineering goal is to discard redundant tokens while retaining exactly the information needed for future computation.

14 killed as fire tears through Russian fireworks factory

Virus Stole a Human Gene and Won't Let Go of It

Article URL: https://www.nytimes.com/2026/09/28/science/virus-molluscum-human-gene.html Comments URL: https://news.ycombinator.com/item?id=49883635 Points: 4 # Comments: 0