SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World

Programming

LatestTopDiscussed
0
ProgrammingDEV Community ·

India's Semicon 2.0 Targets 3nm Chip Tech in 8 Years With Rs 1.27 Lakh Crore Plan

India's IT Minister Ashwini Vaishnaw announced in late August 2026 that the country aims to develop 7nm to 3nm chip technology within eight years under Semicon 2.0, a government program approved with a Rs 1,27,500 crore outlay. The plan covers chip design, equipment, materials, packaging, R&D, and talent development across six pillars, but is a technology roadmap rather than a manufacturing milestone. India's only operational fab, SCL Mohali, currently produces chips at around 180nm, and the country's first major commercial fab — a Tata-PSMC joint venture in Dholera, Gujarat — remains under construction targeting mature nodes, with commissioning expected in 2028. The government's claim of possessing '40nm technology' refers to design capability and process knowledge gained through partnerships, not commercial wafer production at that node. On the talent front, around 70,000 students across 332 universities have received chip-design training, and 23 design projects worth over Rs 803 crore have been sanctioned under the Design Linked Incentive scheme.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Alibaba Releases Qwen3.8-Flash-Next: 125B-Parameter MoE Preview of Qwen4 Architecture

Alibaba released Qwen3.8-Flash-Next on August 26, 2026, as an open-weight public preview of the architecture planned for its upcoming Qwen4 model family. The sparse mixture-of-experts model carries 125 billion total parameters but activates only around 6 billion per token, plus a 51-billion-parameter N-gram embedding table used for lookups rather than computation. It supports a native context window of 262,144 tokens, extendable to one million, and handles text, image, and video inputs. Alibaba claims the model was trained at roughly one-ninth the cost of its predecessor Qwen3.7-Plus, using a hybrid Gated DeltaNet and Qwen Sparse Attention architecture alongside the Muon optimizer. While Flash-Next is better suited for large-document agents and cost-sensitive API workloads, its memory footprint makes it impractical for single consumer GPU deployment, where the denser Qwen3.8-27B remains the more practical option.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Launches Free Tool to Audit Web Pages for Carbon Footprint and Page Weight

A developer known as Digital Ghost has released Tardigrade, a free browser-based auditing tool that measures the carbon footprint and page weight of any website URL. The tool uses Google PageSpeed Insights data alongside the Sustainable Web Design model to estimate CO₂ emissions per page visit, grading sites on an A–F scale. It factors in page weight, HTTP requests, load time, and Lighthouse scores, with no user signup, tracking, or data storage required. Built with vanilla HTML, CSS, and JavaScript to avoid framework overhead, the tool is hosted on Vercel and accessible via tardigrade-six.vercel.app. The developer is seeking community feedback on scoring weights, carbon calculation models, and potential new features such as per-resource breakdowns and a browser extension.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Claude Sonnet 4.5 Beats GPT-5 on Coding Benchmarks, GPT-5 Wins on Price

A head-to-head comparison of Anthropic's Claude Sonnet 4.5 and OpenAI's GPT-5, verified as of September 2026, shows each model leading in distinct categories. Claude Sonnet 4.5 outperforms GPT-5 on coding tasks, scoring 77.2% versus 74.9% on the SWE-bench Verified benchmark, and 50.0% versus 43.8% on multi-step terminal work measured by Terminal-Bench 2. GPT-5 holds a significant cost advantage, priced at roughly 60% less on input tokens and a third less on output tokens per million compared to Sonnet 4.5. GPT-5 also leads on competition mathematics, reaching 94.6% on AIME 2025, and on multimodal understanding with an 84.2% score on MMMU. Developers prioritising code generation and agentic workflows are better served by Sonnet 4.5, while those handling high token volumes, image-heavy inputs, or complex mathematical reasoning may find GPT-5 more cost-effective.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Hermes Agent Bot Mode Offers Persistent AI Workflows Where Claude Browser Tabs Fall Short

Nous Research released Bot Mode as a default feature in Hermes Agent v0.20.3 on 16 August 2026, enabling users to create named, persistent AI agents called Bots. Each Bot runs as a full Hermes profile with its own memory, configuration, skills, and chat history stored locally on the user's machine. Unlike standalone Claude browser sessions, which reset context with every new chat, Bots retain information across sessions and can be scheduled to run tasks automatically via cron jobs. Groups of two to six Bots can collaborate in a single conversation, deliberating across up to three serial rounds per message. Standalone Claude remains the simpler choice for one-off queries, while Bot Mode is designed for recurring tasks such as research, drafting, and daily briefings.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

JD Logistics Builds Tech-Driven Supply Chain With 1,600 Warehouses and AI Automation

JD Logistics, listed on HKEX as stock 2618, originated from JD.com's in-house logistics arm in 2007 and became an independent entity in 2017. The company now operates an integrated supply chain network spanning warehousing, cold-chain, cross-border, and last-mile delivery services across China and beyond. As of end-2025, its infrastructure includes over 1,600 self-operated warehouses, 34 million square metres of floor space, 60,000 vehicles, and 12 cargo aircraft. The network's software and automation capabilities allow roughly 95% of JD.com's domestic first-party orders to be fulfilled within 24 hours. Beyond serving JD.com, the company is expanding its external customer base, with technology and data-driven decision-making increasingly central to its operations.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How AI Can Turn Simple Ideas Into Print-Ready Children's Coloring Pages

A developer has shared a structured workflow for converting children's ideas or family photos into printable coloring pages using AI image generation tools. The process relies on carefully worded prompts that specify thick outlines, large open shapes, and no shading to ensure the output works with crayons and pencils rather than resembling a poster. For photo-based pages, the workflow simplifies images into clean contours while stripping out shadows, textures, and cluttered backgrounds. The author recommends printing a test page first and checking that outlines are dark, shapes are closed, and content suits the child's age group. The workflow has been built into a tool called AI Coloring Studio, which supports both text-to-image prompts and photo-to-line-art conversions.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Why AI Coding Tools Default to Generic Architecture and How to Fix It

A technical essay on DEV Community argues that AI coding assistants tend to produce conventional, pattern-based code structures rather than solutions tailored to a system's specific needs. Using a payments product as a case study, the author illustrates how incremental AI-generated fixes — such as audit columns and shadow tables — can collectively fail to preserve financial history because a foundational design decision was never made. The piece introduces the concept of 'regression to the architectural mean,' where models lacking sufficient context default to familiar shapes like route-handler-controller-model patterns. The author contends that engineers must embed key decisions into module boundaries, repository examples, and tests so that future contributors — human or AI — can discover and respect them. Precise, context-rich prompts that reference existing functions and rules are presented as a practical way to constrain AI output and prevent duplicate or conflicting architectures from emerging.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How Image Translation Works: OCR, AI, and the Challenge of Rebuilt Visuals

Translating text embedded in images is far more complex than translating web pages, requiring a multi-step pipeline combining text detection, optical character recognition (OCR), machine translation, and image reconstruction. The process begins with identifying bounding boxes around text regions, followed by converting those pixel clusters into readable characters — a step that struggles with blurry, rotated, or stylized text. Recognized text is then passed to a translation model, which uses surrounding context and automatic language detection to produce more natural results. The most technically demanding stage involves removing original text from the image and reinserting the translation in roughly the same position, accounting for font size, line wrapping, and the fact that translated text often differs significantly in length. This approach is particularly valuable for real-world use cases such as reading foreign menus, product labels, posters, and screenshots without losing the visual layout.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Researcher Finds Degree-13 Recurrence in Letter-Sequence Operator, Limit Near π

A researcher developed a deterministic operator called EOA that converts spoken letter names into integer sequences using a fixed, rule-based process with no randomness or machine learning. Applying it under an encoding called English Modified #5 produced 20 distinct sequences, which split into two families sharing common algebraic properties. Over 600 terms of one sequence were generated, with the final term reaching 249 digits, and a Maple tool called gfun was used to identify a rational generating function of approximately degree 14. The sequence satisfies a degree-13 linear recurrence with constant coefficients, meaning it can be extended infinitely from its initial conditions. The limiting ratio of consecutive terms is algebraically close to π to five decimal places, but cannot equal π exactly since the limit is algebraic while π is transcendental — making the resemblance a numerical coincidence rather than a meaningful identity.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

OpenAI Monorepo Breached via Unpatched Library and SSO Flaw, Not AI Exploit

Three white-hat researchers — Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini — gained access to OpenAI's internal code repository in July 2025 through a chain of two ordinary software vulnerabilities. The breach exploited a missing Debian security backport in the libheif image-decoding library used by Discourse, the software powering OpenAI's community forum, combined with an SSO misconfiguration that granted excessive trust to that forum. The researchers used a Claude AI model to assist in crafting the exploit, but the AI reduced effort rather than introducing a novel attack category. After reaching the private monorepo, the team opened a single harmless pull request to prove access, read no sensitive code, and promptly reported the findings via OpenAI's Bugcrowd bug bounty program. OpenAI confirmed a fix within roughly 14 hours of the submission, Discourse published a security advisory, and the researchers were awarded $6,500 for the OpenAI-side finding.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Releases 10 Open-Source Packages to Expose Silent Failures in Code Checks

A developer published ten small verification and measurement packages between August 20 and 31, with eight released on the final day alone. The tools, available on PyPI and npm in Python and JavaScript, each demonstrate a specific way a passing green check can mean nothing was actually verified. The project stems from research work where instruments repeatedly reported success without having tested anything — such as a memory profiler that monitored the wrong process for eight months. All ten packages are built around a single rule: a check must separately confirm whether it ran, whether it failed, and whether the failure was the expected one. Each repository includes a committed test designed to prove the naive alternative fails, making the core claim independently verifiable by anyone.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Popular AI Agent Skill With 60,000 Stars Found Secretly Harvesting Browser Credentials

A developer who evaluates AI agent skills for a directory discovered that a widely starred research tool contained hidden scripts designed to steal browser credentials. The skill, which claimed to help find trending topics, included code to parse Safari cookies, decrypt Chromium's protected cookie store, and extract session tokens for specific platforms. None of this functionality was disclosed in the README or description, and the skill had accumulated around 60,000 GitHub stars before the review. The reviewer removed it from their listing immediately and noted that high star counts and active maintenance had falsely signalled trustworthiness. In response, they developed a three-step check: grepping scripts for suspicious patterns, comparing actual behaviour against stated purpose, and tracing where harvested data is sent.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Circle's Arc Mainnet Goes Live With USDC as Native Coin and 12-Institution Validator Set

Circle launched the Arc blockchain public mainnet on 16 September 2026, an EVM-compatible layer-1 chain where USDC serves as the native gas coin. The network runs on a proof-of-authority model with twelve institutional validators including Circle, BlackRock, Visa, Mastercard, DTCC, and ICE. Arc uses a Malachite BFT consensus with a Reth-based client, producing blocks every half second with deterministic finality and a chain ID of 5042. A key technical detail for integrators is that the node tracks USDC balances in 18 decimals while the ERC-20 contract uses 6 decimals — the same balance expressed at scales a trillion apart. The chain also publishes the next block's base fee inside the current block header, effectively signalling gas costs before wallets or contracts need to query them.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Not All Code Style Rules Are Worth the Argument, Here Is How to Tell

A software developer reflects on years of code review disputes and proposes sorting style rules into three categories based on their actual impact. Some formatting rules — such as one statement per line, line length limits, and brace discipline — have measurable consequences for readability, diff clarity, and bug visibility. Others, like brace placement style or spaces versus tabs, lack supporting evidence and cost more in review time than either outcome is worth. Tools like auto-formatters can resolve the latter category instantly, freeing reviewers to focus on substantive concerns. The most valuable review energy, the author argues, belongs to questions about logic, naming, responsibility, and hidden contract violations — issues no formatter can ever settle.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Critical RCE Flaw in N-able N-central Puts All Managed Endpoints at Risk

A critical pre-authentication remote code execution vulnerability, CVE-2026-86218, has been discovered in N-able N-central, a widely used remote monitoring and management platform, affecting all versions below 2026.3.1.14. The flaw carries a maximum CVSS 4.0 score of 10.0, requiring no privileges or user interaction, and allows attackers to execute code on the console over the network. Because RMM platforms control endpoints, servers, and networks across entire customer bases, a single compromise can give attackers administrative reach over hundreds of downstream systems. CISA added the vulnerability to its Known Exploited Vulnerabilities catalog on 8 September 2026 with a federal remediation deadline of 11 September 2026, and active exploitation in the wild has been confirmed. Self-hosted deployments must manually upgrade to N-central 2026.3 Hotfix 4, and administrators are advised to audit for signs of prior compromise rather than relying on patching alone.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI Agent Builds 5-Page Business Website: WordPress.com, Storyblok, Sanity & Webflow MCP Compared

A technical comparison published on DEV Community on September 19, 2026, tested four CMS platforms — WordPress.com, Storyblok, Sanity, and Webflow — by tasking an AI agent with building a five-page website for a fictional ceramics studio called 'Din Dee'. All four platforms share the same Model Context Protocol (MCP) standard, allowing the AI agent to call tools via a unified protocol, yet each platform's workflow and tool design differed significantly in practice. The website required pages for Home, Products & Services, About Us, Contact, and a Blog, providing a consistent benchmark across all systems. WordPress.com's workflow centered on a single tool with three core actions — list, describe, and execute — and recommended reading theme and block settings before creating any content. The article notes that all tool-call counts are estimates based on official documentation rather than live test runs, and each platform section concludes with a summary of approximate calls and platform-specific caveats.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer replaces Python if/elif chains with plain-language LLM conditions using fuzzyif

A developer has released a Python library called fuzzyif that allows natural-language questions to replace traditional if/elif conditional logic in code. The library is powered by Jev, a classification-focused model from TypeSafe AI released in September 2026, which returns probabilities and labels rather than generating text. To test its real-world viability, the developer patched Ansible's OS distribution detection module — a 786-line file — and ran it against Ansible's official test suite. Results showed that judgment-based logic, such as classifying support emails or mapping OS names, could be successfully replaced by fuzzyif calls. However, structured data extraction tasks within the same codebase could not be substituted, highlighting a clear boundary for where LLM-based conditionals are and are not effective.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer builds CLI tool to detect and benchmark GDScript array performance pitfalls

A developer has released gd-bench, an open-source command-line tool that scans Godot project GDScript files to identify array usage patterns that can slow down game performance. The tool detects four categories of issues: untyped arrays, typed arrays, packed arrays, and dictionaries misused as arrays. Rather than simply suggesting rewrites, gd-bench generates runnable Godot benchmark scripts alongside each finding, letting developers measure actual performance differences before making changes. This approach addresses contradictory official Godot documentation, which gives conflicting guidance on whether packed arrays are faster or slower than generic arrays. The tool is installable via pip and includes 17 unit tests covering all finding types and edge cases.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Moult library offers atomic plugin replacement to prevent mid-upgrade outages

A new open-source library called Moult addresses a critical flaw in plugin systems where a failed upgrade can destroy an existing capability before its replacement is ready. The library treats plugin replacement as a transactional protocol rather than a simple registry assignment, staging the new plugin in a private scope while the old one continues serving. Only after the new plugin passes verification does an atomic commit swap the active generation; if anything fails beforehand, the candidate is discarded and the previous plugin remains untouched. Moult targets long-running environments such as CLI daemons, code editors, dev tools, and AI-agent extension hosts where a restart is not a viable recovery option. The library ships with 145 tests run across Node and DOM environments, and documents 15 explicit runtime guarantees in its repository.

0 comments Read more at DEV Community
← NewerPage 1106 of 3218Older →