New Benchmark PeakBench Reveals AI Agents Can Plan Correctly Yet Crash Systems

A preprint published on August 25 introduces PeakBench, a benchmark designed to expose a production failure mode where AI agents correctly identify parallel tasks but overload the machines executing them. The benchmark tested eight leading AI models — including GPT-5, o3, Claude Sonnet 4.6, and DeepSeek variants — on around 300 executable workflows built from approximately 1,200 MCP-compatible tools. Researchers found that planning accuracy had almost zero correlation with capacity violations, meaning a model could map task dependencies perfectly and still schedule jobs that exceed memory or concurrency limits. PeakBench scores dependency planning and resource scheduling separately, so failures can be traced to reasoning, scheduling, or infrastructure rather than lumped into a single end-to-end result. The authors recommend letting AI models build the logical task graph while deterministic infrastructure enforces resource constraints to prevent crashes.

Guinea's President Dismisses 173 Soldiers, Including Former Coup Ally, for Desertion

Guinea's president has dismissed 173 soldiers from the military, among them a former ally from the coup that brought him to power. The mass dismissal was issued on grounds of desertion. Analysts believe the move is part of a broader effort by the president to tighten his grip on the armed forces. The purge signals an ongoing reshaping of Guinea's military under the current leadership.

Nepal Flash Flood Survivor Loses Everything, Left With Only the Clothes She Wore

A woman named Deepika shared her harrowing experience with NDTV after a devastating flash flood struck her community in Nepal. She described fleeing with nothing but the clothes on her back, leaving all her belongings behind. Even basic necessities like drinking water are now unavailable, with her courtyard reduced to a field of mud. Deepika pointed to where her water pitcher once stood, now buried under flood debris. The incident highlights the severe humanitarian toll the flash floods have inflicted on affected Nepali residents.

IIT Delhi Student's Family Recalls His Suicidal Thoughts Before Tragic Death

Rishikesh Kumar, a student at IIT Delhi, died under tragic circumstances after a period of severe mental distress. His relatives recall that he had been experiencing feelings of failure and suicidal ideation in the days leading up to his death. The family believes that certain administrative interactions contributed to his emotional suffering and sense of shame. They had been planning to seek professional mental health support for him before his death. An external committee has been formed to investigate the circumstances surrounding the incident.

Haryana Boxer Sumit Kundu Battles Hardship and Injury to Reach Asian Games

Haryana boxer Sumit Kundu has secured his spot at the Commonwealth and Asian Games after overcoming a series of personal and professional setbacks. Following his father's death and financial difficulties, Kundu temporarily left boxing as a teenager to work as a security guard and support his family. With support from his village community, he returned to the sport and went on to win medals at national and international levels. A serious knee injury later forced him to miss the Paris Olympics, putting his career in jeopardy. Kundu's recovery and subsequent qualification mark a significant comeback in his athletic journey.

Opinion piece examines AI consciousness, machine politics, and herd acceptance

An opinion article published in Communications of the ACM explores philosophical and political questions surrounding artificial intelligence. The piece, titled 'AI Is a Harsh Mistress,' addresses concepts including anima machina and the notion of herd acceptance in AI systems. It raises questions about the politics of conscious machines and how society collectively adopts or defers to AI judgment. The article appeared on Hacker News, attracting modest engagement with 31 points and 4 comments.

BulkImgConvert lets you compress, resize, and convert images entirely in-browser

A developer has launched BulkImgConvert, a free browser-based tool for batch image processing including compression, resizing, cropping, and format conversion. The tool supports formats such as JPG, PNG, WEBP, AVIF, GIF, SVG, PDF, and HEIC as input. All processing runs client-side using the Canvas API, meaning no files are ever uploaded to an external server. The tool works offline once the page is loaded and requires no account, imposes no watermarks, and has no usage limits. It is available at bulkimgconvert.com and was built by the developer to solve a personal frustration with server-dependent online image tools.

Amit Shah Set to Visit Manipur in October Amid NRC Update Push

Union Home Minister Amit Shah is scheduled to visit Manipur in October, according to reports. The visit comes at a politically sensitive time for the northeastern state. Authorities are pushing to update the National Register of Citizens (NRC) ahead of the upcoming national Census. The NRC exercise is aimed at identifying and documenting legitimate citizens in the region.

Meta agrees to $1.4B settlement with 29 states over child safety on social media

Meta has reached an $1.4 billion settlement with 29 U.S. states over allegations that its social media platforms harmed children. The lawsuit claimed that Meta deliberately engineered Instagram and Facebook to be addictive to young users. Authorities argued the company was fully aware of the potential dangers its platforms posed to minors yet continued these practices. The multi-state legal action represents one of the largest settlements involving social media and child safety to date.

Developer Launches Nitya, an AI Assistant for Coding and General Queries

A developer has introduced Nitya, a personal AI assistant built to support everyday development and problem-solving tasks. The tool is capable of handling general conversations, programming questions, debugging, and context-aware responses. Nitya is publicly accessible via a web interface hosted on Vercel. The project is described as ongoing, with the creator actively working to improve its capabilities over time. The developer is seeking user feedback to help guide future enhancements.

SIMURG Tool Claims to Reduce LLM Hallucinations by Intercepting Bad Outputs

A developer has shared early results using a tool called SIMURG to reduce hallucinations in locally run large language models. SIMURG sits in front of an OpenAI-compatible endpoint, monitors text generation in real time, and automatically aborts and retries requests when it detects the model going off track. The developer tested it with a heavily compressed 2-bit quantized version of Qwen3.8 27B running on a consumer-grade RTX 3060 GPU, a setup prone to generating hallucinations and garbled outputs. According to the developer, responses became noticeably more consistent after deploying SIMURG, with end users never seeing the problematic outputs. The tool is described as a 15-parameter machine learning model and is available via pip install, with early testing still ongoing across varied prompts.

Pure Python library whoosh3 can power faceted search without Elasticsearch or APIs

Priya Sundaram, maintainer of whoosh3 — an actively maintained fork of the pure-Python Whoosh full-text search library — has demonstrated how to build faceted search entirely in Python without any external server or API. Faceted search is the sidebar feature common on shopping sites that displays category, brand, and price filters with live counts that update as users drill down. Using whoosh3 version 3.48.0, the implementation requires only a schema definition, an embedded index, and a single groupedby argument on a search call to generate facet counts. The library also supports numeric range bucketing for price filters and can recompute facet counts dynamically against any text search query. When a user selects a filter, a lightweight filter parameter is added to the existing query rather than rerunning the full search, keeping the process efficient.

How Browsers Resolve Conflicting ARIA Roles and Properties in the Accessibility Tree

ARIA roles tell assistive technologies what an element is, but conflicts arise when multiple roles or mismatched ARIA properties appear on the same element. Browsers resolve these conflicts through a priority system: an explicit role attribute takes precedence, followed by implicit roles inferred from HTML tags, with certain roles like alert overriding more generic ones. ARIA properties such as aria-haspopup or aria-expanded can also alter how screen readers announce an element, and misaligning them with the assigned role can produce unpredictable results. To handle this, browsers build an accessibility tree from the DOM, CSS, and ARIA attributes, validating properties against roles and discarding conflicts. Understanding this resolution system is essential for developers debugging accessibility issues in complex or custom UI widgets.

LLM Cost Router Ranks 16th of 18 on Benchmark — Developer Explains Why That Matters

A developer tested OmnisRouter, an AI cost-routing tool built to reduce LLM API bills, against RouterArena, an independent benchmark from an ICLR 2026 paper covering 809 queries across 39 datasets. The router scored 72.7% accuracy at $3.71 per thousand queries, placing 16th out of 18 entrants. The developer noted the low ranking stems from using a premium model pool — GPT-5, Claude Opus, Haiku, and GPT-5-nano — while top-ranked routers rely almost exclusively on cheaper open-source models like DeepSeek and Gemini Flash variants costing as little as four cents per thousand queries. The benchmark's cost metric effectively rewards routing to the cheapest models that pass academic tests, which differs from real-world enterprise use cases where output quality on complex tasks matters more. The developer argues OmnisRouter's value is demonstrated on production traffic, where routing decisions cut a projected $30,000 monthly API bill by 50–58% on actual coding-agent workloads.

How Combining Agent Loops With Graph Memory Creates Self-Improving AI

Most AI agents today operate as stateless loops — they complete tasks but retain no memory of what they did or learned, forcing them to restart from scratch each time a similar problem arises. A proposed architecture combines agent loops with graph-based memory, allowing agents to record relationships between files, actions, outcomes, and people involved in each completed task. Unlike storing experience as flat text documents, a graph structure captures the connections between facts, making retrieval more precise and contextually relevant at scale. When an agent encounters a new problem, it can query its accumulated graph of past experiences to prioritize relevant files and strategies rather than beginning with no prior knowledge. This approach differs from standard retrieval-augmented generation by emphasizing structured, experience-driven memory that grows and improves through the agent's own work.

MIT Press Profiles a Thinker Who Envisioned Humanity on a Cosmic Timescale

MIT Press Reader published a profile exploring the life and ideas of a thinker known for contemplating humanity's place across vast cosmic timescales. The piece examines how this individual conceptualized human existence from a perspective spanning two billion years. The article appeared on the MIT Press Reader platform and generated discussion on Hacker News. It reflects a broader intellectual tradition of placing human civilization within deep-time and astronomical frameworks.

France reaches 94.9% fiber coverage in 2026

Article URL: https://cartefibre.arcep.fr Comments URL: https://news.ycombinator.com/item?id=49448872 Points: 7 # Comments: 0

Army Veteran Jeret Howard Joins NASA Stennis Space Center as Civil Servant

Jeret Howard, a former U.S. Army combat medic who served in Iraq, has been hired as a civil servant at NASA's Stennis Space Center near Bay St. Louis, Mississippi. He was brought on board as part of NASA's workforce directive aimed at restoring core competencies within the agency. Howard had previously worked at Stennis as a contractor before making the transition to a federal civil servant role this past summer. His journey from military service at age 17 to a NASA career reflects the agency's broader effort to build an experienced in-house workforce.

Meta settles social media harms case for $16.68 bn

Meta settles with 29 states, agrees to strict safety limits for teen users

Meta has reached a settlement with a coalition of 29 state attorneys general over child online safety concerns, avoiding a trial that could have resulted in hundreds of billions of dollars in penalties. As part of the deal, Meta must implement an age verification system subject to independent testing, with strict limits on false identification rates for users aged 13 to 17. The company will also give teenagers the ability to opt out of algorithmically personalized content feeds. Additionally, Meta agreed to restrict push notifications for younger users, reflecting broader efforts to limit the platform's influence on minors.