Three dead, one injured in Ramgarh steel plant fire in Jharkhand

A major fire broke out at a steel plant in Ramgarh, Jharkhand, late Friday night. The blaze originated in the main control room before quickly engulfing the rest of the facility. Three workers lost their lives and one person was injured in the incident. A number of employees were present at the plant when the fire broke out. Authorities are working to verify the identities of the victims and confirm the full casualty count.

Three Engineers Dead After Turbine Blast and Fire at Ramgarh Steel Plant

A turbine explosion at a steel plant in Ramgarh, Jharkhand, killed three engineers early Saturday morning. The blast triggered a large fire that quickly engulfed the facility's control room. A fourth employee was injured in the incident and has been taken to hospital for treatment. Residents of nearby villages staged protests, accusing the factory management of serious safety negligence. Authorities have launched a formal investigation into the cause of the fatal accident.

India News 22 August Live Updates: Interpol issues Red Corner Notice against alleged mastermind in Sambhal violence case

Developer's Cost Monitor Silently Failed for a Week Due to a Common Shell Scripting Trap

A developer building an autonomous Claude Code environment discovered their cost-monitoring script had two critical flaws: it initially overcounted API spending by 4–6x due to double-summing cumulative log lines, then went completely silent for a week after being 'hardened' with set -euo pipefail. The strict error-handling flag caused the entire script to exit whenever a single command failed, leaving the status bar blank while significant API costs accumulated unnoticed. The developer argues that monitoring scripts should be designed to 'fail open' — displaying a visible 'no data' indicator rather than terminating silently — so users can distinguish between healthy operation and missing data. This principle emerged from real financial stakes: the developer notes that Claude Code's metered API costs can exceed $3,000 in a single week if left unmonitored. The resulting 212-line bash script now exits with code 0 on errors while displaying a visible placeholder, ensuring the dashboard never goes quiet.

VideoChat3 and RoboTTT Challenge Single-Backbone AI with Specialized Architectures

Two new AI systems, VideoChat3 and RoboTTT, push back against the dominant trend of using one shared transformer backbone for vision, language, and action tasks. VideoChat3 employs a dedicated video-centric backbone with just 4 billion parameters, halving visual token counts while outperforming larger open-source multimodal models on video benchmarks. RoboTTT takes a different approach, integrating test-time training into existing vision-language-action policies to extend context length for robotic tasks, boosting manipulation performance by 87 percent. It also successfully completed a five-minute, ten-stage assembly task that no prior baseline model could finish. Together, the two systems raise the question of whether modality-specific architectures could broadly replace monolithic designs, though questions around training costs and untested deployment scenarios remain open.

ReBA scheduler fixes expert load imbalance in vision-language AI models

Researchers have proposed ReBA, a geometry-guided routing scheduler designed to address load imbalance in vision-language mixture-of-experts (MoE) models. The core problem is that existing routing methods treat all tokens equally, allowing high-resolution image patches to overwhelm certain experts while text tokens remain underutilized. ReBA solves this by enforcing separate routing pathways for visual and linguistic tokens and applying equal weighting per image instance. In tests across four model backbones, ReBA reduced layer-level load variation more consistently than both the standard auxiliary loss method and a coupled image-instance baseline. The study notes limitations, including the absence of wall-clock speed or memory benchmarks and untested performance on very large-scale deployments.

FlashMorph automates hybrid attention design using just 20M tokens and 2 GPU-hours

Researchers have introduced FlashMorph, a method that automates the conversion of standard attention models into hybrid architectures by framing layer selection as a global subset-optimization problem. Unlike previous approaches that relied on costly grid searches or hand-crafted layer patterns, FlashMorph requires only 20 million tokens and approximately 2.1 GPU-hours to determine the optimal mix of full and linear attention layers. The system freezes pretrained weights and jointly learns binary gates on synthetic retrieval data, followed by standard distillation to finalize the architecture. In evaluations, FlashMorph matched dense baseline performance on short and medium context retrieval benchmarks with negligible accuracy loss. However, the study has not yet tested generative tasks, and the authors note that synthetic training data may not fully reflect real-world distribution shifts.

BM25 Outperforms AI Agents in Large-Scale Document Retrieval, Study Finds

A new scaling study has found that the decades-old BM25 lexical retrieval algorithm outperforms sophisticated AI-driven agentic search systems as document collections grow larger. At the largest corpus size tested, BM25 achieved an accuracy score of 50.5, compared to 30.7 for the File-System Agent and 29.9 for DenseRAG. The performance gap widens significantly beyond approximately 10 million corpus tokens, where BM25 consistently dominates all competing methods. Agentic approaches also proved far more expensive, with the File-System Agent consuming 39 times more query tokens than a single BM25 pass. Researchers suggest a hybrid approach — using BM25 first to narrow candidates, then applying lightweight agents for post-ranking refinement — could offer the best balance of cost and accuracy.

Training-Free Head Pruning Cuts 20% of Attention FLOPs in Diffusion Transformers

Researchers have found that structural template tokens in diffusion transformers act as implicit semantic registers, absorbing text semantics while allowing a significant portion of attention heads to be safely removed. A training-free pruning rule that targets heads attending most strongly to prompt tokens eliminates roughly 20% of joint-attention FLOPs with only a 1.4-point drop in GenEval scores. Around 20–30% of attention heads can be pruned without retraining, additional data, or gradient-based saliency analysis. The finding marks a departure from prior pruning strategies, which treated all heads as equally essential and relied on weight magnitude or costly fine-tuning loops. The study is currently limited to text-to-image diffusion transformers evaluated on GenEval, and its applicability to other architectures or quality metrics remains to be tested.

Self-Geometry method boosts vision model accuracy using epipolar constraints at test time

Researchers have introduced Self-Geometry, a test-time adaptation pipeline that enforces multi-view geometric consistency in vision foundation models without requiring full retraining. Unlike prior methods that rely on a model's own outputs for self-consistency, Self-Geometry uses pseudo ground-truth 2D correspondences optimized against epipolar and multi-view losses. The approach was evaluated across six vision foundation models and four benchmark datasets, consistently improving both pose accuracy and depth scores. On the challenging ETH3D benchmark, the method improved pose accuracy of the VGGT model by up to 37.3% on a key metric. Despite running in under two minutes per scene, the pipeline is not suited for real-time use and may struggle in texture-poor or dynamic environments where reliable correspondences are hard to obtain.

Why Correlation Heatmaps Miss Multicollinearity in Regression Models

A common practice in regression analysis involves using pairwise correlation heatmaps to detect multicollinearity, but this method has a fundamental blind spot. Multicollinearity can exist across three or more features simultaneously even when no individual pair exceeds the correlation threshold, such as when one variable is the sum of two others. The Variance Inflation Factor (VIF) is a more reliable diagnostic because it regresses each feature against all others, capturing multivariate dependencies that pairwise statistics cannot detect. A high VIF signals that a coefficient's variance is inflated, making it statistically unstable and potentially misleading for causal or policy-driven interpretation. Crucially, multicollinearity only poses a real problem when the goal is to interpret coefficients rather than generate predictions, so the appropriate response depends on whether the model is being used for explanation or forecasting.

Developer shares structured process for shipping startup web apps on time and budget

A freelance web developer has outlined a detailed workflow for building and launching startup web apps, emphasizing that most projects fail during planning rather than development. The process begins with a paid discovery session producing a concise two-page brief and milestone plan, deliberately avoiding lengthy requirement documents. Projects are divided into three to five milestones, each ending in a browser-accessible, deployed slice of the product, with client sign-off required before proceeding to the next stage. Staging environments are set up from day one, and tools like Next.js, Vercel, and Zod-based environment validation are used to maintain consistency across deployment stages. The developer argues that strict scope management and incremental delivery are the primary factors in preventing budget overruns and project delays.

19 Chrome DevTools Features Most Developers Overlook Daily

Chrome DevTools offers far more functionality than most developers regularly use, according to a guide published on DEV Community. The article highlights 19 lesser-known tricks, including a built-in Command Palette (Ctrl+Shift+P) for quickly accessing features without navigating menus. Developers can take full-page screenshots, force CSS hover and focus states, block specific network requests, and enable live page editing directly within DevTools. Other tips cover reading minified code with a one-click formatter, simulating slow network conditions, and using Console shortcuts like $0 to inspect recently selected elements. These features are designed to speed up debugging and reduce reliance on external tools or browser extensions.

Y Combinator CEO urges founders to spend big on AI tokens as strategic investment

Y Combinator CEO Garry Tan is encouraging startup founders to spend heavily on AI tokens rather than worry about the cost. He argues that doing so effectively gives businesses access to future AI capabilities in the present. Tan recommends deploying AI agents for various tasks and carefully documenting the workflows that prove successful. These documented processes can then be converted into repeatable, automated systems for businesses. While the so-called 'tokenmaxxing' approach has sparked debate, Tan frames it as a worthwhile long-term investment.

Shillong Sees Panic Buying After Union Threatens Fuel Supply Suspension

The North-East Petroleum Mazdoor Union (NEPMZ) on Friday threatened to halt the transportation of petrol, diesel, and LPG from Assam to Meghalaya. The union cited safety concerns for Assam-registered vehicles operating within Meghalaya as the primary reason for the threatened action. The announcement triggered panic buying at fuel pumps across Shillong, as residents rushed to stock up on fuel. The threatened suspension raises concerns about potential fuel shortages in Meghalaya, which depends on supplies transported from neighbouring Assam.

Opinion: Modern Life Mirrors AI 'Slop' — Plausible but Ungrounded, Argues Developer

A software developer writing on DEV Community argues that everyday systems — job markets, institutions, and social advice — exhibit the same structural flaw as AI language models: producing outputs optimized for the appearance of correctness rather than accuracy itself. The author defines 'slop' not as garbage but as fluent, confident output generated without any reliable mechanism to verify claims against reality. Drawing on personal debugging experience, they contend that this pattern emerges not from deliberate deception but from optimization pressure applied in the wrong direction. The piece extends a previous argument the author made about large language models hallucinating, applying the same critique to real-world systems people cannot simply opt out of. The post is framed explicitly as opinion rooted in personal feeling rather than technical evidence.

Why Canada suspended US trade talks despite Trump's ‘good relationship’ with Carney

EPF vs stock market: EPFO reveals 6 differences

As AI Writes More Code, Developer Judgment Becomes the Real Skill

AI coding tools can now generate components, fix bugs, write tests, and build entire features with minimal human input, prompting a debate about the evolving role of software developers. Experts argue that while code generation is becoming cheaper and faster, the ability to evaluate, architect, and understand that code remains a distinctly human responsibility. A key concern is that easy access to AI-generated code may allow developers — especially beginners — to skip the foundational problem-solving experience that traditionally builds expertise. The shift does not eliminate software development but moves the core work from writing code to making technical decisions, reviewing output, and ensuring systems are secure, accessible, and scalable. The industry now faces the challenge of redefining what developer skill means in an era where producing code is no longer the hardest part of the job.

Telangana Revenue Officer Found with Rs 6.87 Crore in Assets During Raids

A Telangana government revenue officer named Desani Srinivas Reddy has been found in possession of disproportionate assets worth Rs 6.87 crore following official raids. Reddy currently serves as a divisional administrative officer in Ibrahimpatnam, located in Ranga Reddy district. The assets uncovered include a penthouse, gold, and cash. The raids were conducted as part of an investigation into the officer's alleged accumulation of wealth beyond his known sources of income.