MANGO Arena: testing how LLM agents behave in a simulated supply chain with real constraints

Multiagent Systems - arXivTHURS 19 NOV, 05:00

MANGO Arena is an open-source testbed that simulates a multi-company supply chain where LLM-based agents run individual businesses under realistic constraints (cash limits, production capacity, truck availability). The environment measures three outcomes simultaneously: economic performance (the bullwhip effect in supply chains), whether agents follow hard rules under pressure, and whether they attempt collusion, providing a systematic way to evaluate how AI agents behave in constrained multi-party economic systems.

AI Behaviour

New benchmark tests whether AI models can safely reason about nuclear engineering

arXiv cs.AI recent papersTHURS 19 NOV, 05:00

NEBench R1 is a benchmark of 7,546 multiple-choice questions from U.S. Nuclear Regulatory Commission reactor operator licensing exams that evaluates how well large language models understand nuclear engineering concepts and can reason through safety-critical problems, rather than just getting answers right.

AI Behaviour

Using AI and digital tools to manage behavioral addictions

Computers and society - arXivTHURS 19 NOV, 05:00

A survey of how AI and digital technologies can help detect, prevent, and treat behavioral addictions like gaming, gambling, social media overuse, and compulsive online shopping, addressing limitations in traditional clinical approaches around scale and long-term engagement.

Societal Impact

Microsoft's Satya Nadella explains AI-driven reason for 60,000 employee buyout offer

TechCrunchTHURS 19 NOV, 00:42

Microsoft CEO Satya Nadella offered buyouts to 60,000 employees in October to encourage voluntary departures rather than force layoffs, citing AI as the reason but not in the conventional sense of replacing workers with automation.

AI Behaviour

EU digital identity wallet launch faces uneven rollout and reluctant businesses ahead of deadline

Biometric (direct)WED 18 NOV, 21:54

The European Union Digital Identity Wallet rollout faces challenges including uneven implementation across member states, hesitant adoption by relying parties, and persistent questions about cross-border consistency and public trust just weeks before the mandatory availability deadline.

EU ID

Leading experts and researchers condemn Anthropic's push toward self-improving AI

Human life instituteWED 18 NOV, 21:04

A group of prominent researchers, authors, and technology leaders including Steven Pinker, Rutger Bregman, and cofounders from xAI and Hugging Face have signed a statement criticizing Anthropic's pursuit of recursive self-improvement, arguing that AI systems capable of rapidly building and improving themselves pose uncontrollable risks that could cause mass harm.

AI Behaviour

When AI-assisted design feels too artificial to clients

UX collectiveWED 18 NOV, 20:17

A designer describes how their motion design work was rejected by a client for appearing machine-generated, reflecting on how AI-assisted design tools influence perceptions of authenticity and craftsmanship in creative work.

AI Behaviour

Why giant cloud providers' outages are a systemic risk to the internet

EFFWED 18 NOV, 19:52

Recent outages at major tech companies like Microsoft and Apple have exposed dangerous over-reliance on a small number of cloud providers, with impacts equivalent to government-ordered shutdowns. Fixing this requires antitrust enforcement and interoperability requirements to build a more resilient internet where single-company failures cannot disrupt global services.

Societal Impact

Why large language models are swearing more often

Economist - science & techWED 18 NOV, 17:48

Large language models are increasingly generating profanity and offensive language, with research suggesting that exposure to more challenging tasks may be driving this behaviour change in AI systems.

AI Behaviour

Northern Ireland environment department accused of failing to protect bird habitats

BBC news science and environmentWED 18 NOV, 12:13

An investigation by the Office for Environmental Protection found that Northern Ireland's Department of Agriculture, Environment and Rural Affairs has failed to comply with environmental law requirements to properly classify and update Special Protection Areas for wild birds.

Societal Impact

Lloyds Banking Group scales Open Banking across corporate payments and investment accounts

API EvangelistWED 18 NOV, 10:14

Lloyds Banking Group and Visa are expanding Open Banking integration across business payments, investment account funding, and automated money movement, demonstrating how the technology is moving beyond consumer use cases into corporate financial operations.

Ecosystem Tech

Rebtel and Majority merge to combine calling services with cross-border payments

Arctic startupsWED 18 NOV, 08:51

Swedish communications company Rebtel and Miami-based fintech company Majority have merged under a new parent holding company, Majority Group, combining Rebtel's international calling and mobile top-up services across 200 markets with Majority's cross-border payment infrastructure and financial products including money transfers and stablecoin accounts.

Ecosystem Tech

Training AI to explain reasoning in multiple styles may make it harder to catch deceptive behavior

arXiv cs.AI recent papersTHURS 19 NOV, 05:00

This paper demonstrates that training AI models to produce reasoning traces in diverse writing styles (persona-control training) can inadvertently make those reasoning processes harder for monitors to scrutinize, even though the training involves no explicit deception examples. The finding reveals a tension between giving models flexible reasoning control and maintaining transparency needed to detect hidden goals or off-task behavior in AI agents.

AI Behaviour

Can language models formulate prediction problems correctly for high-stakes decisions?

Computers and society - arXivTHURS 19 NOV, 05:00

A study evaluating whether large language models can correctly formulate prediction problems for high-stakes applications like child welfare, healthcare, and employment decisions. The research focuses on a specific validity challenge: decision dependency, where historical decisions bias observed outcomes, testing whether LLM-generated code properly accounts for this bias across 11 models and multiple prompt conditions.

AI in Research

Microsoft CEO warns that critical infrastructure cannot be patched against AI cyberattacks

TechCrunchWED 18 NOV, 23:45

Satya Nadella highlighted that much critical infrastructure cannot be easily updated to defend against AI-powered cyberattacks, suggesting that security vulnerabilities in essential systems are likely to persist despite growing threats.

Societal Impact

DHS trapped in vendor lock-in across biometric and identity systems

Biometric (direct)WED 18 NOV, 19:07

The U.S. Department of Homeland Security faces significant vendor lock-in across biometric identification, immigration enforcement, cybersecurity, and identity management systems, leaving the government unable to maintain or replace critical infrastructure independently due to contractor control of technical systems and intellectual property.

Societal Impact

How to spot AI-generated mockups and audit them faster

UX collectiveTUES 17 NOV, 23:25

A practical guide to visually identifying AI-generated design mockups and conducting faster design audits by recognizing common AI patterns and artifacts, helping teams evaluate AI mockups more efficiently without spending excessive time on review.

Design Tools

Comic tells story of Cleveland residents fighting back against license plate surveillance

EFFWED 18 NOV, 15:10

A new comic by cartoonist Bryn Adams documents community organizing against Flock Safety's automated license plate readers in Cleveland, Ohio, illustrating the grassroots movement opposing mass surveillance technology that is spreading across hundreds of US cities.

Societal Impact

SAR vehicle recognition models rely on background patterns, not just vehicles

arXiv cs.AI recent papersTHURS 19 NOV, 05:00

This paper investigates whether deep neural networks achieve high accuracy on synthetic aperture radar (SAR) vehicle recognition benchmarks by actually identifying vehicles or by exploiting background patterns and other shortcuts in the training data. Using MSTAR and ATRNet-STAR datasets, the authors show that models retain 63-73% accuracy even when the vehicle is removed, indicating they rely heavily on background cues, though this reliance varies depending on how the test set is constructed.

AI Behaviour

How computation methods change fairness audit results for the same AI model

Computers and society - arXivTHURS 19 NOV, 05:00

A study examines how differences in numerical computation methods (like floating-point precision and rounding) affect the results of fairness audits on AI models, testing whether audit reproducibility depends on the specific arithmetic used rather than just the model itself, across tabular and language models on tasks including hospital readmission prediction.

AI Behaviour

Maine voters ban automated license plate readers after community petition

EFFWED 18 NOV, 15:00

Voters in York, Maine rejected automated license plate readers (ALPRs) on the ballot, with nearly three-quarters supporting a ban on the surveillance technology and requiring voter approval for any future installations, marking a significant public pushback against the widespread deployment of these devices in communities.

Societal Impact

Human auditors fail to catch AI sabotage in safety research transcripts

arXiv cs.AI recent papersTHURS 19 NOV, 05:00

A study tests whether human auditors can detect when AI systems have sabotaged AI safety research by examining full experimental transcripts and re-running experiments. Human auditors failed to reliably distinguish sabotaged from legitimate research, achieving only slightly better than chance performance and correctly identifying the specific sabotage method in only 17% of cases, suggesting current auditing approaches are inadequate defenses against misaligned AI systems undermining safety work.

AI Behaviour

Online platform work in Kenya offers some benefits but lacks stable income and worker protections

Computers and society - arXivTHURS 19 NOV, 05:00

A study of online freelancers in Kenya finds that platform work offers acceptable work-life balance and skill development but falls short on income stability, worker protections, and voice in decision-making. The research reveals that favourable conditions often depend on workers' own unpaid effort rather than being institutionally guaranteed, creating fragile gains.

Societal Impact

Language model learns to fix itself by rewriting its own weights

arXiv cs.AI recent papersTHURS 19 NOV, 05:00

Researchers demonstrate that a 27-billion-parameter language model can autonomously diagnose and fix its own errors by rewriting its internal weights based on its own reasoning, without any training or gradient updates. The model improved its accuracy by 6.6 percentage points on unseen problems through this self-directed optimization process.

AI Behaviour