Science corner, 16-Oct-26 Research papers of the week on AI and the fields around it, in plain English: the ones our sources talked about most, and in each field the ones that matter most. AI RESEARCH * A video model that answers new questions without redoing the work Researchers built a family of models that turn video understanding into structured decisions instead of written-out answers, so they can point to moments in a video or pick among answer options. Queries arriving during a session reuse earlier visual computation. On three moment-finding tests the models improved on their starting versions and ran roughly three times faster; on four video question sets they matched them while often running far faster. – Why it matters: It offers one reusable way for visual agents to make judgments from video as goals change, which the authors say can be combined into complex, open-ended workflows. – Why people are talking about it: A widely followed account flagged the paper, and it drew hundreds of upvotes on Hugging Face. arXiv · AI research * Predictive world models let AI systems improve themselves across tasks Researchers built a framework that uses predictive world models to guide action and pick better training data, keeping updated policy-model pairs that pass compatibility checks. In code repair, fixing the coder, success rose from 26.3% with a static verifier to 31.2% after adapting the world model, with similar test usage. Across five policy versions, lab-planning scores more than doubled. – Why it matters: The results suggest systems can keep improving from their own experience without costly external feedback, which could make AI more useful across coding, lab work and science. – Why people are talking about it: The source highlights that the same self-improvement framework handled very different tasks, from fixing bugs and operating computers to planning lab work and optimizing reactions, without changing the underlying coder model or increasing test budgets. arXiv · AI research * Telling AI models to stop thinking does not stop them Researchers tested six language models with six ways of instructing them to answer without showing reasoning, across true/false, multiple-choice and open-ended questions. Strict instructions did not reliably remove visible reasoning, a pattern they call thinking inertia. It grew stronger as questions became more open, and on open-ended tasks answering without reasoning traded off against accuracy. Giving candidate answers made answer-only replies easier. – Why it matters: The paper argues that no-thinking is a real capability that cannot be assumed from a model's settings or instructions, and should be evaluated alongside reasoning ability. – Why people are talking about it: A widely followed AI account shared the paper, drawing attention to the finding that models keep reasoning even when told not to. arXiv · AI research AI AND MANAGERIAL WORK * Autonomous AI agents cut knowledge work time and cost Researchers studied production data from Perplexity's Search and Computer products to see how autonomous AI agents change knowledge work. On matched tasks, sessions with the agent performed 26 minutes of autonomous work versus 33 seconds for search, cut completion time from 269 to 36 minutes, and lowered estimated time and cost by 87 percent and 94 percent. Dissatisfaction fell 55 percent. – Why it matters: The findings suggest autonomous agents can speed up workflows, improve output quality, reduce costs, and expand the range of work that gets automated, changing how knowledge tasks are delegated. – Why people are talking about it: One source asks how much autonomy an AI agent should have, arguing maximum autonomy is not always the right goal and that teams need ways to test how far to delegate. arXiv · AI and managerial work AI AND ECONOMICS * Higher wealth taxes in Norway led to few departures Norway raised dividend taxes, closed a capital gains loophole for people who move abroad, and nearly doubled the tax on business wealth between 2021 and 2024. The researchers find at most 100 individuals left because of the changes, mostly the very richest, and no drop in new founders or inventors. Firms stayed active and did not invest less. – Why it matters: The results suggest the short-run economic costs of taxing the ultra wealthy, in lost migration and reduced business investment, are smaller than the public debate implies. – Why people are talking about it: It speaks to a live argument about billionaire exits and whether taxing the ultra wealthy drives them away. NBER · AI and economics AI AND JOBS * ChatGPT cut writing time 40% and raised quality Researchers gave occupation-specific, paid writing tasks to 453 college-educated professionals and randomly let half use the ChatGPT chatbot. Those with access finished faster and produced better work, and the gap between stronger and weaker writers narrowed. Fears and enthusiasm about AI both rose briefly. Weeks later, those who had used it were far more likely to use it in their real jobs. – Why it matters: The findings suggest generative AI can make midlevel professional writing faster and better while narrowing differences between workers, and that a single hands-on experience shapes whether people keep using it at work. – Why people are talking about it: One source cites the study while arguing that agent execution time is only one term in a broader calculation, urging engineering leaders to check other factors before scaling pilots. AI and jobs * AI assistant raises support agent productivity, helping less-skilled workers most Researchers studied the staggered rollout of a generative AI chat assistant to 5,172 customer-support agents. Access to the assistant raised productivity by 15 percent on average, measured as issues resolved per hour. Less experienced and lower-skilled agents sped up and improved quality; the most experienced saw small speed gains but slightly lower quality. Gains were largest for moderately rare problems. – Why it matters: AI assistance can lift output and help newer workers learn faster and write better English, while top performers gain little. It also changes the job: customers are more polite and ask for a manager less often. – Why people are talking about it: One source uses the paper to ask whether AI agents actually save work, arguing that agent execution time is only one term in a seven-term equation and that leaders should check the other six before scaling a pilot. AI and jobs AI AND SOFTWARE DEVELOPMENT * Training setup helps coding agents build whole programs from scratch The researchers built a framework that turns existing software repositories into practice environments where coding agents can learn to build programs from scratch. It produced 1,648 training environments across four programming languages and paired them with a workflow that separates analysis, implementation and review. Agents that trained this way performed better at reproducing programs, and the workflow lifted results further. – Why it matters: Coding agents are moving from fixing bugs to writing whole programs, but reliable practice grounds for that work are scarce. The researchers say they will release the framework to support more work on agents that build software from scratch. arXiv · AI and software development * Coding agents fail to build agents that meet the brief Researchers set up a test where coding agents are commissioned to build, edit and tune working software agents, much like real clients would ask. Across 25 tasks, every agent tested fell short: what a builder shows on a few example cases rarely matches what it delivers when checked on fresh cases. Common failures include tuning too closely to the visible examples and code that looks finished but does not actually run. – Why it matters: Software is increasingly built by software itself, and the results show a gap between an agent appearing to meet a requirement and actually meeting it. That gap matters when teams delegate work to coding agents and assume the result is faithful. arXiv · AI and software development AI AND BUSINESS EDUCATION * Structured AI use in business classes lifts students' confidence and job readiness Researchers ran a field experiment with 612 students in 34 seminar groups at three business schools in China. One set followed a structured routine for using generative AI in case work; the other had the same curriculum with unstructured AI access. The structured group reported better self-perceived employability, through stronger professional self-efficacy and motivation. – Why it matters: The paper argues the question for business schools is not whether to use generative AI, but how to use it while students keep ownership of their own reflection and judgment. SSRN · AI and business education AI AND CONSULTING * Consulting firms redesign themselves for human-AI delivery The paper looks at consulting and professional services firms that moved AI-assisted delivery from trials into everyday work between 2024 and 2026, and asks what such a firm sells when clients can license the same models, who answers for an error passing through several agents, and whether reviewing AI work builds judgment. It treats business model, operating design and incentives as one problem. – Why it matters: It argues a firm's lasting advantages are independence from the client's frame, accountability a regulator can reach, and a data feedback loop competitors cannot buy. SSRN · AI and consulting AI AND EDUCATION * Disadvantaged schools offer fewer chances to learn to judge AI output Researchers compared students' reported chances to evaluate AI-generated information with their chances to evaluate ordinary online information, using survey data from 78 education systems. The gap between richer and poorer students was larger for AI material. Most of that extra gap fell between schools rather than inside them, and it showed up mainly in whether students met such tasks at all, not how often. – Why it matters: It suggests students in disadvantaged schools may be less likely to learn how to judge AI-generated information, so schooling could widen rather than narrow this new skills gap. SSRN · AI and education --- The Observatory · https://observatory.mgorsky.net/archive/135/