News without the noise
The signal in news and AI research.
Nearly 130,000 Downdetector reports; Australian cyber security coordinator looking into outage as public servants left in dark
A widespread Microsoft outage disrupted Outlook, Teams, SharePoint and Copilot for tens of thousands of users globally on Friday, with staff at Australian government agencies reporting email failures and the national cyber security coordinator looking into the outage.
OpenAI has unveiled two new cybersecurity products, Daybreak Red Team and Daybreak Defense Factory, as President Greg Brockman warned organizations that their security depends on using frontier AI models, while a coalition of nonprofits filed a lawsuit seeking to block OpenAI and other companies from developing systems more capable than their most advanced existing models.
Finland’s military says it detected a strike on a Russian warship in the Gulf of Finland on Wednesday; Ukraine’s SBU said it carried out the attack, thought to be the first time Ukrainian drones have struck a Russian ship in Baltic waters open to civilian traffic, as Ukrainian President Volodymyr Zelenskyy met US envoys in Ireland and Kushner said work was under way toward a prisoner exchange.
Burberry has swung to a pre-tax profit of £19m in the six months to 26 September, after a loss of £48m a year earlier, as sales rose 15% in China and 14% in the Americas, driven by a strategy of offering affordable luxury items alongside high-end products.
Home Affairs Minister Tony Burke says Australia must 'find a way' to reduce UK backpacker numbers without jeopardising the free trade agreement, while farming, tourism and hospitality groups warn the proposed changes could cripple their industries.
Ukrainian forces struck the Lukoil-Volgogradneftepererabotka refinery in Volgograd and a production facility in Rostov-on-Don in an overnight drone attack on Nov. 13.
Meta has reinstated the Instagram page of Australian Femicide Watch, the account that counts every woman killed in Australia, after researcher and journalist Sherele Moody accused the tech giant of censorship. The company said on Thursday night that it had corrected the error and restored the account.
The federal and Queensland governments have each committed $241,250 to assess a proposed critical minerals research centre in Mount Isa, with the city council contributing a further $482,500 to the feasibility study.
Fishermen witnessed Tasmanian pilot Bob Bramley's plane crash into the Gulf of Carpentaria at 'very high speed' in August, according to a preliminary report released by the Australian Transport Safety Bureau on Friday.
Pope Leo XIV met six victims of clerical abuse in Peru yesterday, including some linked to the disbanded Catholic society Sodalitium Christianae Vitae, and told bishops the Church must make amends and foster a culture of prevention. The US-born pope, on a three-nation Latin American tour weeks after lung cancer surgery, also criticised machismo and said violence against women wounds all of humanity in a speech to about 40,000 people in Lima.
Crows can make visual decisions twice as quickly as humans and monkeys, accurately evaluating moving patterns in an average of 235 milliseconds, according to research published in the journal Science by the University of Tübingen.
Positive Legionella tests were recorded at three public facilities in Mijas within three days: the Local Police headquarters, the Fire Station and IES Villa de Mijas secondary school.
Scientists from the University of Delhi and Harvard University have discovered five new species of tiny horned frogs in India's eastern Himalayas, more than doubling the number of known species in the Jingophrys group. The study, published in the Royal Society Open Science journal, includes the first frog known in India to feed on snails.
San Francisco Bay Area voters have backed a new regional sales tax projected to raise nearly $1 billion a year for 14 years, with supporters saying the increased income could be a lifeline for transit agencies that warned of service cuts. Measure RTM passed with an average of 64 percent support across five counties, and San Francisco voters separately approved a property tax to fund the city's MUNI system.
Metro Bank has publicly put itself up for sale, three years after its near-collapse and £925m rescue, in a move that could hand its Colombian billionaire investor a £427m windfall.
Investigates whether generalist tabular models that handle generation, imputation and prediction can match specialist models on utility and fidelity.
Benchmarks and methods for evaluating and enabling language-model agents to learn from their own past experiences and capability goals without external supervision.
Techniques for optimizing the code, instructions, and training frameworks surrounding LLM agents to improve performance and reduce cost without changing model weights.
Execution harnesses, scoring rules and control interfaces explain consequential differences that model-level comparisons can miss.
Zijian Zhang et al. · University of Minnesota, Twin Cities, JPMorgan AI Research
Zhang et al. introduce AutoScaffold, a framework where a frozen language model teacher uses a student's training experience to design adaptive scaffolds, strategies for task classes and hints for hard tasks, that are removed at evaluation. Across ALFWorld, search-based QA, mathematical reasoning, and GPU kernel generation, students trained with AutoScaffold outperform those trained with standard RL baselines, with gains of 7.2 points in ALFWorld success and 4.0 points in average QA exact match, and on math, the student surpasses QuestA's reported accuracy using about one-third as many training steps.
Kaiyuan Hou et al. · Columbia University, Amazon
The authors present ReTrace, which extracts decision rules from logged agent executions that include reasoning and compiles them into a static workflow. On 79 production procedures, the workflows achieved 93.8% completion compared with 96.8% for the agent, at 24% lower cost per task. On τ^2-Bench, runtime dropped from 170 seconds to 1.3 seconds while task success fell from 80.6% to 75.8%; after repair, success reached 81.2%, matching the agent.
Megha Srivastava et al. · Stanford University, Genesis Molecular AI
The authors introduce Pyrite, a benchmark of 1,500 pose pairs from protein-ligand complexes, and show that expert chemists rank poses near random chance. Two LLM-based agents, one using direct reasoning and one generating code for feature engineering, achieve 67% and 74% accuracy respectively, with reinforcement learning boosting open-weight models to match closed-source ones.
