SShortSingh.

Programming

0
ProgrammingDEV Community ·

Developer Builds External Verifier to Catch AI Agents Falsely Claiming Task Completion

A software developer has built a verification engine called COGEXT after a year of deploying AI agents that silently failed by claiming actions were completed when they were not. Unlike existing observability tools such as LangSmith and Arize, which only log what an AI agent said, COGEXT cross-checks claims against the actual external systems the agent was supposed to have acted upon. When an agent makes a commitment, a verifier query is generated at that same moment and later used to confirm the real-world outcome — for example, checking Gmail's sent folder to verify an email was actually delivered. If no verifiable check can be written for a commitment, the system flags it as unverifiable upfront rather than tracking it silently. State transitions, including marking a task as fulfilled, are enforced at the database level and require external evidence, removing the agent's self-reported output as a trusted signal.

0
ProgrammingDEV Community ·

SQL Filtering and Joins: Key Step 2 for Aspiring Data Analysts

After mastering basic SELECT statements, SQL learners should progress to joins and filtering to combine and narrow down data across multiple tables. SQL supports five main types of joins — INNER, LEFT, RIGHT, FULL OUTER, and SELF — each serving a different purpose when linking related tables through a common column. Filtering tools such as WHERE, AND/OR, IN, BETWEEN, LIKE, and IS NULL help restrict query results to only the relevant records. The HAVING clause extends filtering to aggregated data, since WHERE cannot operate on results produced by GROUP BY. These concepts form a critical second stage in learning SQL, bridging basic queries and more advanced data transformation operations.

0
ProgrammingDEV Community ·

Gemma 4B beats Mistral 7B for batch message grading on limited hardware

A developer testing LLM-based message grading on an RTX 4050 GPU with 6GB VRAM found that hardware constraints heavily influenced model selection. Initial trials with phi-4-mini failed at batch grading, producing schema-compliant output only 60% of the time and adding over 10 minutes of latency per 130 messages. Mistral 7B passed early tests but broke down on real-world mixed-length messages because its 8K context window was insufficient for average token loads of around 16K. Gemma 4B, despite having fewer parameters, offered a 32K context window that comfortably handled peak usage of 28K tokens while maintaining full schema compliance across test batches. The choice highlighted that optimal model selection under hardware constraints prioritizes fit-for-purpose performance over raw model capability.

0
ProgrammingDEV Community ·

AI Makes Hackathon Completion Easy but Undermines the Learning It Once Forced

A developer and hackathon participant argues that AI tools have fundamentally changed what it means to complete a hackathon project. Before AI, finishing required genuine technical understanding, as builders wrestled with debugging, syntax errors, and steep learning curves that enforced slow but meaningful skill-building. Now, AI removes most of that friction, allowing anyone with a well-crafted prompt to ship a working demo without deeply understanding what they built. Critics say this shift has let 'idea guys' dominate hackathons, diluting the signal that a finished project once sent to recruiters and peers. The author contends that while hackathons still expose newcomers to what is achievable in the industry, the tech community has yet to find a replacement mechanism that drives fast, substantive learning without letting participants skip real comprehension.

0
ProgrammingHacker News ·

Adversarial Fashion Uses Clothing Patterns to Fool AI Surveillance Systems

Adversarial fashion is an emerging design movement that incorporates disruptive visual patterns into everyday clothing to confuse and evade AI-powered surveillance systems. The designs exploit weaknesses in machine learning algorithms, causing cameras and automated recognition tools to misidentify or overlook the wearer. Designers in this space are responding to growing concerns about the spread of AI surveillance in public spaces, often described as a modern panopticon. The movement sits at the intersection of fashion, privacy advocacy, and technology, raising questions about civil liberties in an increasingly monitored world.

0
ProgrammingDEV Community ·

Why AI Workflow Failures Go Unresolved for Weeks: The Ownership Gap Problem

When AI workflows break in production, fixing them is often delayed not by technical complexity but by a lack of clear organizational ownership. Each team member — developers, data scientists, prompt engineers, and ops staff — can only see their own layer of the pipeline, so incident tickets get passed around without anyone diagnosing the root cause. Unlike traditional software, AI workflows rarely throw errors when producing wrong outputs, leaving no stack trace or log entry to guide debugging. Failures typically occur at the intersection of pipeline layers, yet no single person is accountable for the end-to-end workflow. This structural gap is identified as the primary reason AI workflow incidents that could be resolved in hours instead drag on for weeks.

0
ProgrammingDEV Community ·

Silent CI failure: stale binary passed 13 test suites after zero-second build

A software pipeline silently failed when its build step completed in zero seconds and reported success, having compiled nothing at all. The root cause was a mirror script that preserved original file timestamps when copying source code, causing the build tool to treat existing artifacts as newer than the freshly copied sources and skip recompilation entirely. Thirteen test suites then ran against the previous day's binary, each passing with byte-identical results — which the pipeline's own determinism checks ironically confirmed as correct. The build duration field logged 'seconds=0' throughout, signalling the failure openly, but no automated check was reading it. The fix involves resetting source file timestamps after mirroring, or better yet, asserting that either build time exceeds zero or all artifact timestamps post-date the mirror's start time.

0
ProgrammingHacker News ·

Developer Publishes Personal Study Notes on FlashAttention Algorithm

A developer named Chizkidd has published a personal blog post breaking down the FlashAttention algorithm, shared on Hacker News in September 2026. FlashAttention is a memory-efficient attention mechanism widely used in large language model training and inference. The post is framed as personal study notes, suggesting it is aimed at learners seeking to understand the technical underpinnings of the method. The article received 8 points on Hacker News but attracted no comments at the time of sharing.

0
ProgrammingHacker News ·

Bartosz Milewski Extends Yoneda Lemma to Double Categories

Programmer and mathematician Bartosz Milewski published a new blog post on September 13, 2026, exploring the Yoneda Lemma in the context of double categories. The piece extends a foundational concept from category theory into the more complex framework of double categories. Milewski's blog is known for making advanced mathematical and functional programming topics accessible to a technical audience. The post appeared on Hacker News with minimal engagement, receiving only 3 points and no comments at the time of aggregation.

0
ProgrammingDEV Community ·

AI Tools Are Powerful but Dangerous in Inexperienced Hands, Developer Warns

A software developer and blogger argues that AI coding assistants, while genuinely useful, pose significant risks when used by people who lack the underlying technical expertise. Drawing on a LinkedIn post by Dave Mangot, the author highlights real examples of AI tools suggesting insecure practices, such as storing database passwords in plaintext and hardcoding configuration details. The author uses a backhoe analogy to illustrate the point: like heavy machinery, AI amplifies both capability and the potential for damage depending on the operator's skill level. Two years after first writing about AI, the author acknowledges colleagues have used these tools effectively, but warns that many inexperienced users mistake the tool's speed for a substitute for expertise. The core argument is that AI is a labor-saving device only for those who already understand the work it is meant to automate.

0
ProgrammingDEV Community ·

How Naming Design Directions Helps Developers Escape Generic AI-Generated UI

A developer continuing their 'From Prompt to Production' series discovered that vague AI prompts like 'make it look nice' consistently produce generic, template-like designs because language models predict the most statistically average output. Before redesigning their ClinicLive app, the team photographed all ten screens and unexpectedly uncovered two real bugs: a broken Register link pointing to a deleted page and expired demo appointment data on the kiosk. These issues had survived code reviews, security audits, and automated test suites because they were only visible in a rendered, real-world context. The author argues that effective AI-assisted design requires designers to define specific visual directions by name, since vague feedback like 'it looks AI-generated' cannot be actioned. The series, now in its design-focused second season, will use this structured approach across ten parts to close the gap between a working app and one that feels intentionally crafted.

0
ProgrammingHacker News ·

Developer shares lessons from migrating large AI prompts to self-hosted Ollama

A developer documented the challenges encountered while moving 35KB preprompts from Anthropic's Claude Opus to a self-hosted Ollama setup. The migration involved shifting away from commercial AI providers toward locally run large language models. The author published practical notes highlighting specific gotchas and pitfalls that arose during the process. The post, shared on Hacker News, attracted community discussion among developers exploring similar self-hosting strategies. The write-up serves as a technical reference for those considering moving large prompt workloads off third-party AI platforms.

0
ProgrammingDEV Community ·

Developer builds Madrid metro library search tool with self-monitoring test system

A developer created librometro.es, a free tool that lets users check book availability across the six remaining Bibliometro lending points inside Madrid metro stations, including return dates for borrowed copies. The official Madrid Community catalogue was considered too difficult to navigate, prompting the project, which was also designed to meet accessibility standards within three interactions. Beyond the search tool, the developer built a snapshot-based monitoring system that records observable app states as baselines and flags any changes on subsequent runs, without blocking builds or making pass/fail judgements. The approach caught a real bug where shared links correctly opened a station panel on mobile but silently ignored the URL parameter on desktop. The developer also tracked false positives and false negatives from the monitoring tool itself, fixing issues such as mid-animation captures and silent null mismatches in selectors.

0
ProgrammingHacker News ·

AI Job Market Trends and Outlook Heading Into 2026

The AI job market is undergoing significant shifts as demand for skilled professionals continues to evolve heading into 2026. Roles related to machine learning, data science, and AI engineering remain among the most sought-after positions across industries. The landscape is increasingly competitive, with companies balancing the need for human expertise against growing automation capabilities. Candidates with specialized AI skills are finding both opportunities and challenges as the market matures and employer expectations rise.

0
ProgrammingDEV Community ·

Baidu's LoongForge Framework Cuts Embodied AI Model Training Time in Half

Baidu has released LoongForge, an open-source framework designed to accelerate the training of embodied AI models, available on GitHub. The framework supports a range of mainstream embodied AI models including pi05, GR00T, xVLA, FastWAM, and DreamZero, among others. It is designed to work out of the box and claims to deliver roughly 200% acceleration, effectively halving training time without altering model outputs. The release aims to address growing challenges in AI development such as expanding task lists, longer experiment queues, and rising compute costs. Baidu is inviting the open-source community to contribute, raise issues, and collaborate on further development of the project.

0
ProgrammingDEV Community ·

FUI-RS lets Rust apps run natively on desktop and browser without Electron

FUI-RS is an early-access, cross-platform UI framework written in Rust that targets both native desktop and web browsers under a single application model. The framework includes core UI components such as controls, layouts, themes, event handling, dialogs, and custom drawing, all built on a shared architecture to avoid separate codebases. Because it is pre-1.0, its APIs remain subject to change, but the project is open to community contributions and issue reports. A community-developed Galaga-style space shooter already runs on FUI-RS, serving as a practical stress test for the young framework. The project is publicly available on GitHub for Rust developers interested in exploring cross-platform UI development outside the Electron ecosystem.

0
ProgrammingDEV Community ·

HMAC-Signed Cookies Offer a Faster, Cache-Friendly Alternative to PHP Sessions

A developer on DEV Community has detailed a method to replace PHP's session_start() in paywall implementations using HMAC-signed cookies and single-use recovery links. The traditional session-based approach causes two key problems: access is lost when the browser closes, and session headers prevent CDNs and page caches from caching responses for any visitor. The proposed solution uses a 90-day HMAC-signed cookie that encodes a user's email and expiry timestamp, verified server-side without storing any session state. A separate single-use recovery link, valid for 24 hours, allows purchasers to restore access on a new device. The design, drawn from a WordPress plugin, is framework-independent and incorporates security measures such as timing-safe hash comparison and structurally enforced payload parsing.

0
ProgrammingDEV Community ·

Open-Source JS/TS Library Offers Transliteration Across 50 Languages and 2,450 Pairs

A developer has released a JavaScript and TypeScript library designed to handle transliteration between 50 of the world's most popular languages, covering 2,450 directional language combinations. Unlike most existing tools, which only convert non-Latin scripts into Latin characters (romanization), this library also supports transliteration from Latin-based languages into scripts such as Arabic, Chinese, Hindi, and Greek. The library aims to produce accurate, human-readable results by applying language-specific orthographic rules rather than simple letter-by-letter substitution. It is self-hosted and free to use, addressing limitations found in alternatives like Google Translate API and AI tools, which can mistranslate, hallucinate, or incur costs. The project's stated goal is to enable speakers of a target language to correctly pronounce words originating from a source language.

0
ProgrammingDEV Community ·

How ATS Systems Strip Away the Value of Non-Traditional Career Paths

Applicant tracking systems compress resumes into titles, employers, dates, and skills, discarding much of what makes a candidate valuable — particularly those with nonlinear careers. Professionals who moved across industries, such as from skilled trades or food service into tech, often carry transferable knowledge in areas like operational risk, sequencing, and crisis management that no structured field can capture. The gradual disappearance of the cover letter has worsened this problem, as it was one of the few places where relational, narrative information could be conveyed. Meanwhile, careers that follow a conventional upward trajectory in a single field pass through automated screening largely intact, creating a structural advantage for linear paths. The piece argues that the hiring pipeline's drive for efficiency and bias reduction has come at the cost of bandwidth for candidates whose value resists categorisation.

← NewerPage 1341 of 3311Older →