17-year-old Naga boy killed in Manipur ambush

EU's Kallas Urges Keeping Sanctions on Russia to Cut War Funding

EU foreign policy chief Kaja Kallas has called for maintaining sanctions against Russia, describing them as a central element of Europe's response to the ongoing war. She made the remarks to reporters in New York following a meeting of EU foreign affairs ministers. Kallas emphasized that the sanctions serve to deprive Moscow of the financial resources needed to sustain its military campaign. The statement signals continued European resolve to keep economic pressure on Russia.

Dev builds Chrome extension for live translated captions on Meet, Zoom and Teams

An independent developer has built a Chrome extension that displays real-time translated captions for browser-based video calls on Google Meet, Zoom, and Teams, without requiring a bot to join the call. The extension captures tab audio using Chrome's MV3 service worker architecture and an offscreen document, routing the stream back to speakers so users can still hear the call. Early versions used separate Deepgram and DeepL services for speech recognition and translation, but interim-result retranslation inflated costs to roughly $2.55 per hour of call. The developer switched to Soniox's stt-rt-v5 model, which handles both recognition and translation in a single stream, reducing complexity and cost. Key technical hurdles included handling stale tab capture locks, managing activeTab permission resets after page navigation, and correctly parsing server-side end-of-line signals undocumented by the provider.

AI-Built 'Museum of Almost Right' Shows How Assumptions Skew Data Results

A developer named Jianqiang, working with AI assistant Codex, has built an interactive web exhibit called 'Museum of Almost Right' for the Sanity Challenge. The project presents four synthetic datasets where visitors can toggle between competing assumptions — such as numeric versus default sorting or different treatments of missing values — to see how the same data yields different results. Built using Astro and Sanity, the museum requires no visitor account and is publicly accessible without any paid API dependency. The goal is not to score users but to help them identify the specific condition that makes a given data interpretation defensible. All datasets are invented for educational purposes, and the project's code and content are released under the MIT license.

Independent Researcher Trains 1.18B LLM From Scratch on a Single Consumer GPU

Independent researcher Mert Çetin of Me Force Technology has trained CetinLM Base-v1, a 1.18-billion-parameter language model, entirely from scratch using a single RTX 4070 Ti SUPER desktop GPU with 16GB of VRAM. The model has surpassed 3.80 billion training tokens, with validation loss continuing to fall steadily from 2.592976 at 3.60B tokens to 2.577079 at 3.80B tokens, showing no signs of plateauing. Rather than relying on large-scale web data dumps, Çetin built a curated proprietary dataset over a week-long pipeline, aiming for higher logical density per token. The base model, which has not yet undergone instruction fine-tuning or alignment, reportedly runs at around 48 tokens per second on a local interface and has shown coherent responses in both English and Turkish. Çetin argues the project demonstrates that foundational language model research does not require massive compute clusters or venture capital funding to produce stable, meaningful results.

Developer Proposes PHP-Based AI Agent Framework to Leverage Existing Web Hosting

A developer has proposed building AI agents using PHP, arguing that the language already powers at least half of the internet and runs on widely available, low-cost hosting infrastructure. The idea challenges the dominance of Python and Node.js in AI development, which the author says locks builders into commercial providers. The author draws a parallel to WordPress, which succeeded by being open, accessible, and easy to deploy on mainstream shared hosting. The proposal questions whether the millions of servers already configured for WordPress-style setups will realistically migrate to Django or Node.js environments within the next five years. The core argument is that an agent-first PHP framework could replicate WordPress's broad adoption by meeting developers where existing infrastructure already lives.

Developer Documents QA Automation Learning Journey From Basics to CI/CD Pipelines

A developer has shared their ongoing experience building a professional QA automation framework from scratch, currently working on Project 152. A key lesson learned involved ensuring that CI/CD pipelines fail when tests fail, while still preserving failure evidence and data for later review. The project, though small in scope, provided practical insight into how real-world CI pipelines should behave. The developer plans to continue documenting their progress, including mistakes and problem-solving approaches, as they advance their automation skills.

How Graph-Based Audit Trails Can Track Why AI Agents Make Decisions

A developer has built a system to automatically record the full reasoning chain of AI agents, addressing the problem of models fabricating justifications when their decision traces are lost. The solution uses a hook-based recorder built on the Strands Agents framework, which captures tool calls and inputs without requiring any changes to existing agent tools. Decision traces are stored in Neo4j using the graph database vendor's own agent-memory SDK, rather than flat log files or key-value stores. The graph structure enables a reverse audit capability, allowing engineers to run a single traversal query to identify every past decision that relied on a data source later found to be incorrect. The approach is framework-agnostic in principle and is presented as part of a broader series on agent memory management.

Courts must not interfere in cop probe: Supreme Court

The Real AI Risk Lies With the Companies That Control It

A recently published opinion piece argues that artificial intelligence itself is not the primary threat facing society. Instead, the author contends that the companies developing and deploying AI systems pose the greater danger. The concern centers on how corporate interests, power concentration, and lack of accountability shape how AI is built and used. The piece draws a distinction between the technology as a tool and the institutional decisions made around it. It has sparked discussion online, appearing on Hacker News where it received early attention from readers.

Why Measuring Internet-Exposed RDP Services Matters for Cybersecurity Defenders

Remote Desktop Protocol (RDP), typically running on port 3389, remains one of the most widely exposed and high-value services on the public internet, making it a persistent target for attackers. Security researchers use large-scale passive and active scanning tools to map this exposure, capturing service banners and handshake data to identify open ports and software versions. However, these scans produce only point-in-time snapshots, as firewalls, NAT configurations, and dynamic IP allocation can cause assets to appear or disappear between scans. Experts caution that a publicly visible RDP service does not automatically indicate a breach or vulnerability, since many organizations expose RDP intentionally with strong authentication and MFA controls in place. Defenders are advised to treat such measurement data as an indicator of potential risk rather than a definitive census of compromised or insecure systems.

From $2 Bets to $500K: Horseplayer Roger Cettina Helps Rebuild Belmont Park

Roger Cettina, a 59-year-old Hall of Fame horseplayer, started placing $2 bets as a teenager before going on to win nearly $500,000 through National Horseplayers Championship contests. Over decades, he developed a deep personal connection with Belmont Park, one of horse racing's most iconic venues. Cettina was recently given the opportunity to contribute to the legendary racetrack's rebuilding efforts. He has described the experience as an unbelievable, once-in-a-lifetime opportunity that brings his lifelong passion full circle.

Ireland's DPC fines Google €403M for mishandling user location data

Ireland's Data Protection Commission (DPC) has imposed a €403 million fine on Google following an inquiry into how the company processed users' location data. The investigation examined whether Google's handling of location information complied with the EU's General Data Protection Regulation (GDPR). The DPC, which serves as Google's lead data regulator in the EU due to the company's European headquarters being in Ireland, found violations significant enough to warrant the substantial penalty. This ruling adds to a growing list of major GDPR enforcement actions taken against large technology companies operating in Europe.

Angular Concepts Explained Through a Simulated Junior Developer Interview

A DEV Community article presents Angular fundamentals through a fictional technical interview between a candidate named Pedro and an interviewer named João. The dialogue-based format covers core Angular concepts including project structure, Services, and separation of responsibilities. Key topics addressed include why HTTP calls belong in Services rather than components, how Route Guards protect authenticated routes, and the role of HTTP Interceptors in automatically attaching authorization tokens. The format is designed to help beginners learn by observing how a real interview scenario might unfold, including moments where the candidate answers incorrectly and receives corrections.

Supreme Court flags grosslypoor infra at IBC tribunals

I don't want to read what you didn't write

Article URL: https://blog.colinbreck.com/i-dont-want-to-read-what-you-didnt-write/ Comments URL: https://news.ycombinator.com/item?id=49794330 Points: 46 # Comments: 13

Google Gemini Expands Into AI-Powered Shopping With Cart and Visual Search Tools

Google is integrating shopping capabilities across Gemini, AI Mode, and its Shopping Graph, aiming to make AI-assisted product discovery a more connected part of the buying journey. A promotional demo suggests users may soon research purchases, compare products, and browse using camera images directly within Gemini. Google's I/O 2026 materials also highlight a Universal Cart concept that can aggregate items from Search, Gemini, YouTube, and Gmail into a single intelligent cart. The initiative includes Buy buttons on eligible product listings within AI Mode, pointing toward purchase paths that extend beyond traditional search results. Exact availability, supported markets, and merchant requirements for these features have not yet been officially confirmed.

Analysis of 22,901 Resume PDFs Finds 80% Have Major Parsing Flaws

A free resume-checking tool analyzed 22,901 PDF resumes submitted between May and September 2026, finding that only 5.8% had zero layout issues, while 79.8% contained at least one major or critical problem. Contrary to popular advice, hidden layout tables were the most common critical flaw, flagged in 32.7% of files, nearly four times more frequent than the two-column layouts that most guides warn against. These invisible tables, often inherited from word processor templates, cause text extractors to scramble job titles, dates, and bullet points into the wrong order. Other widespread issues included decorative icons or emoji disrupting text parsing (48.9%), non-standard fonts (46%), and non-standard section headings like 'Professional Journey' that can cause entire resume sections to go undetected (39.2%). Contact details placed in header or footer bands were missed by many parsers in 22.1% of files, since some applicant tracking systems only read the main body layer of a PDF.

Meta's Muse AI Assistant Found Vulnerable to Critical Zero-Day Exploit

Meta's AI assistant Muse has been found to contain a serious zero-day vulnerability, raising significant security concerns. The flaw can be exploited through a relatively simple ClickFix-style attack, which is just one of multiple methods capable of fully hijacking the agent. The vulnerability is considered especially serious given the elevated privileges Muse operates with. Security researchers have flagged the issue as it poses a substantial risk to users of the AI assistant. The discovery highlights growing concerns around the security of AI agents with broad system access.

Markdown Bold Text in AI Agent Instruction Files Does Not Affect Model Behavior

Developers commonly bold rules in AI agent instruction files like CLAUDE.md or AGENTS.md expecting the model to prioritize them, but no documented mechanism translates text emphasis into increased model attention. Language models process asterisks as plain characters in a sequence, with no pathway from visual formatting to weighted instruction following. Research from Georgia Tech, UC Berkeley, and Microsoft (PASTA) shows attention reweighting is possible at inference, but this is not how current agent file parsing works. Anthropic's own Claude Code guidance recommends using the word 'IMPORTANT' on a single line instead of bold markup, warning that over-emphasizing many lines makes none stand out. Across Claude, Codex, and Gemini ecosystems, none document Markdown emphasis as a functional steering mechanism in instruction files.