100s of flights cancelled at UK airports due to ATC issue

Article URL: https://www.bbc.com/news/live/c6x2z0yy32ejt Comments URL: https://news.ycombinator.com/item?id=49614557 Points: 15 # Comments: 6

Allostatik Offers a Persistent AI Context System Built on Plain Files You Control

Allostatik is a tool designed to give users a continuous, compounding working relationship with AI by storing context in plain text files the user owns. Unlike default AI sessions that reset with each conversation, Allostatik loads project state at the start of every session and updates it at the close, so each session builds on the last. The system addresses a common problem where AI lacks access to past decisions, reversed preferences, or evolving project context, causing it to contradict or misrepresent the user's intentions. Rather than relying on AI-generated memory summaries the user never authored, Allostatik keeps the user as the sole gatekeeper of every change to their context files. The tool's core value lies not just in capturing context but in a structured maintenance routine that ensures records stay current and accurate across projects.

Bamber's Five-Wicket Haul Skittles Glamorgan for 69, Lifts Warwickshire in Title Race

Warwickshire seamer Ethan Bamber delivered a dominant bowling performance against Glamorgan in a County Championship fixture. He claimed a five-wicket haul, also reaching the milestone of 50 wickets for the season in the process. His efforts helped bundle Glamorgan out for a meager 69 runs. The result has strengthened Warwickshire's position in the County Championship title race.

Raine's resistance keeps Durham afloat after Morgan's four-wicket haul for Middlesex

Middlesex gained the upper hand against Durham after pacer Matt Morgan claimed a four-wicket haul to put the visitors under pressure. Higgins and Sharma also contributed with wickets, further tightening Middlesex's grip on the match. Durham's Ben Raine offered notable resistance with the bat, keeping his side's hopes alive. However, Brydon Carse struck twice to peg back the visitors and shift momentum once more. The match remained finely poised as both sides traded blows across the contest.

Household mfg grows 4x faster than corporates

What Really Happens When You Fail Google Play's 14-Day Testing Requirement

Developers with personal Google Play accounts created on or after November 13, 2023, must complete a closed testing phase with at least 12 opted-in testers for 14 consecutive days before gaining production access. Google reduced this tester threshold from 20 to 12 in December 2024, but the quality of engagement still matters — not just the numbers. Failing the requirement either resets the streak timer or results in a rejected production access application, but does not suspend or penalize the developer account. Google flags inactive, fake, or unengaged testers by tracking downloads, app launches, feedback submissions, and update activity during the testing window. Developers who fail can address the issues and reapply once a new compliant closed test is completed.

Developer Builds Brain-Inspired AI Memory System That Forgets Recent Work by Design

A developer created an AI memory system called 'mycelium' that mimics how biological brains reinforce and forget information, storing memories as interconnected nodes rather than simple searchable documents. The system strengthens links between concepts recalled together frequently and allows unused memories to decay over time, treating forgetting as a deliberate feature. Built on SQLite with FTS5 and a connection graph as its primary retrieval mechanism, it surfaces 'hub' memories — the most frequently accessed and heavily connected concepts — at the start of each session. However, the developer discovered a key limitation: recent memories, lacking connections or access history, consistently ranked below older, well-reinforced ones, making it difficult to resume work from the previous session. The flaw highlights a fundamental tension between salience-based reinforcement and recency, mirroring a trade-off found in human memory itself.

Career Changer Documents Every AWS Serverless Pitfall in Cloud Resume Challenge Build

A former insurance operations professional completed the Cloud Resume Challenge to transition into cloud engineering, building a serverless resume site hosted on AWS. The project uses a private S3 bucket behind CloudFront with Origin Access Control, a Lambda-powered visitor counter backed by DynamoDB, and is fully managed via Terraform and deployed through GitHub Actions. During the build, a domain registration failure on Route 53 was traced to a fraud alert triggered by an account upgrade, leaving a support case unresolved for two days. A separate issue arose when an ACM certificate request failed because DNS validation required a Route 53 hosted zone that did not yet exist, creating a dependency deadlock. The developer worked around both blockers by continuing progress on unaffected components and upgrading to a paid support plan to accelerate AWS's response.

AWS AI Agents Power Real-Time Maritime Vessel Monitoring System

A technical guide published on DEV Community on September 8 details how to build a maritime watch floor using AI agents on AWS. The system is designed to monitor ships that go 'dark,' meaning vessels that stop transmitting location or identification signals. Author Shadab Hussain, an AWS Community Builder, outlines the architecture and system design behind the agentic AI solution. The project leverages AWS cloud infrastructure combined with agentic AI to enable continuous, automated maritime surveillance. The write-up, estimated at a 20-minute read, covers topics including AI, AWS, agentic AI, and system design.

Tata lowest bidder for Rs 20k-cr tank project

India eyes fresh UK investment treaty talks

AI Debugging Tools Cut Bug Resolution Time but Only 6% of Teams Use Them Fully

According to a 2026 JetBrains survey, only 18% of developers trust AI debugging assistants to produce production-ready fixes without human review, while adoption remains shallow across the industry. New data from Sentry shows top AI debugging tools resolve 41% of production bugs within 24 hours, compared to just 13% for human-only teams. A Snyk 2026 report attributes 73% of critical post-release bugs to AI-generated code, underscoring the growing complexity of modern software pipelines. OpenAI reportedly reduced its median bug resolution time by 66%, from 41 hours to 14 hours, after deploying a hybrid Copilot-Sentry AI triage system across its ChatGPT API team. Gartner data from 2026 indicates that 38% of companies adopting advanced AI debugging tools reduced incident response costs by a median of $15,400 per team annually.

41% of AI Integration Projects Fail or Overspend, Gartner 2026 Data Shows

A 2026 industry analysis highlights that 41% of AI software integration projects exceed budget or fail entirely, even as 88% of US enterprises rank AI-powered integration as their top IT priority. The median annual spend on AI integration stands at $624,000, well below the $1.4 million average cost of a failed project, creating significant financial risk for organizations. Experts identify poor model selection, dirty data, and lack of real-time monitoring as the leading causes of failure, with 72% of integration errors linked to inconsistent or incomplete data. Automated validation tools, continuous model monitoring platforms, and early-stage compliance checks are recommended as essential safeguards rather than optional add-ons. Non-compliant AI integrations averaged $2.2 million in regulatory fines in 2026, underscoring that security and data governance must be built in from the start.

UK to Fine Apple and Google Over Failure to Block Child Nudity on Devices

Apple and Google have missed a UK government deadline to implement measures blocking child nudity on their platforms. The failure to comply has prompted British authorities to announce plans for new legislation targeting the tech giants. Under the proposed laws, companies could face significant financial penalties as well as potential criminal liability. The UK government has also signaled that the regulatory crackdown may be extended to other platforms, including Snapchat and Instagram.

Only One AWS Region Supports Llama 3.3 70B Fine-Tuning on Bedrock

A developer querying all 33 Amazon Bedrock regions via the AWS SDK found that only us-west-2 supports fine-tuning of Meta's Llama 3.3 70B model, as of September 2026. Two other regions returned valid responses but lacked Llama 3.3 support, 17 regions reported zero tunable models, and 13 opt-in regions were inaccessible on the test account. Amazon Bedrock's public pricing page lists training costs for Llama 2 and several other models but omits any published rate for Llama 3.3 70B, despite the capability existing and generating real AWS charges. Third-party guides have been found incorrectly applying Llama 2 70B pricing to Llama 3.3 70B, which the developer warns could lead to inaccurate budget estimates. The author recommends deriving cost estimates from an actual training job rather than relying on figures inferred from documentation or community sources.

Jeff Bezos’ $75M yacht fix faces Miami ban

Romanian websites score average 17/100 on AI agent discoverability, study finds

A study using the Exista.io ADO Score framework tested 130 Romanian domains across 12 machine-verifiable signals to assess how well AI agents — not human users — can find and evaluate companies autonomously. The mean score was just 17 out of 100, with no domain exceeding 35, and zero out of 130 sites serving an agent card file, which carries the highest weight in the scoring system. Most points came from legacy SEO elements such as JSON-LD structured data and Wikidata presence, not from agent-specific infrastructure. The limited interoperability detected was largely incidental, driven by Shopify and WordPress platforms rather than deliberate strategy, with only one domain — IT firm snsys.ro — having manually built its own agent manifest. The findings suggest that even companies already visible to AI-powered search tools are largely unprepared for the emerging wave of autonomous AI agents that verify rather than browse.

Thomas Century Puts Somerset in Command Against Leicestershire on Day One

Somerset's Thomas scored a crucial century on the opening day of their County Championship clash against Leicestershire. The innings came on a green-tinged pitch that made batting challenging, giving Somerset a strong platform early in the match. Somerset currently sit at the top of the County Championship table, and the performance reinforced their title-challenging credentials. Leicestershire, who have struggled for form this season, found themselves under pressure as Somerset set the pace. The century proved pivotal in allowing the table-toppers to dominate proceedings on a difficult surface.

llama.cpp Drops Key Multi-GPU Flag, But New Features Restore Lost Speed

A developer running a dual Tesla P40 GPU setup built their entire inference pipeline around the llama.cpp flag '-sm row', which delivered 12–14 tokens/sec compared to roughly 7 for layer split mode. In July 2026, upstream llama.cpp removed the flag entirely — not as a deprecation but a full deletion — forcing a shift to layer split as the only multi-GPU option. Earlier testing had also revealed that an alternative 'graph' split mode, claimed to be 40% faster, caused hard CUDA crashes on Pascal-generation GPUs, underscoring the risk of relying on unverified community benchmarks. A separate architecture-level bug in Gemma 4's shared KV layers had already made row split unusable for that model family, revealing that split mode affects correctness, not just performance. However, parallel inference slots and speculative decoding on the new stack pushed aggregate throughput beyond what the removed flag had previously achieved, ultimately leaving the setup faster than before.

Engineer Builds $2,000 Local AI Inference Server Using 2016-Era Hardware

A software engineer built a local AI inference server for roughly $2,000 using enterprise-surplus components, including two NVIDIA Tesla P40 GPUs from 2016 and an AMD EPYC processor, sourced mostly from eBay in January 2026. The build was driven by cost constraints rather than ideology, and the author estimates it paid for itself within two months compared to equivalent frontier API expenses. Running on Pascal-generation GPUs with no Tensor Cores and a compute capability of 6.1, the setup required workarounds including INT8 quantization and ruling out popular tools like vLLM entirely. Despite the hardware limitations, the server processed over 8,100 requests across a 40-hour production window with a failure rate of just 0.17% and no manual interventions. The author notes that most publicly available LLM performance advice targets newer Ampere-generation hardware, making Pascal-specific tuning a significant and often misleading challenge.