Researchers at Columbia University tested how well language models — including GPT-4o and GPT-4 Turbo — are able to reproduce the values of different societies. They compared the models' responses with the results of the World Values Survey, an international survey that records the values of citizens in 21 country-language pairs. A total of 94 questions were used, translated and verified by native speakers. The models answered these questions as if they were residents of the respective countries, and the answers were compared with those given by humans. A deviation of more than 50% was considered an error.

The researchers then compared the accuracy of the responses with the level of digital representation of the language — the proportion of websites and texts on the internet. The result was clear: 44% of the variation in GPT-4o's ability to imitate societal values can be explained by the amount of digital data available in that language. For GPT-4 Turbo, this correlation reached 72%, indicating its greater dependence on web-scraped content. Errors in low-resource languages (i.e., languages with an unsaturated digital footprint — Hausa, Amharic, Burmese, Shona) occur more than five times more often than in English, German, or Japanese.

The discrepancies are particularly noticeable in the topics of security, ethical norms, and political culture — precisely where the digital presence of local communities is limited or censored. The authors view this not simply as a technical error, but as a new form of digital inequality, where AI inherits the cultural bias of hegemonic languages.

Even though GPT-4o has reduced the error rate in low-resource languages by 10 percentage points, the overall trend remains: models better understand cultures with a rich digital footprint, while others are represented only fragmentarily.

In response, researchers are calling for the creation of multilingual LLMs with the participation of local communities, open datasets, and ethical standards of diversity and fairness — so that AI not only speaks different languages, but also understands their worlds.

The agent AI industry is growing rapidly: by 2025, investments exceeded $2 billion, and the forecast for the next 10 years is estimated to be 100 times higher. Behind this growth is the confidence that large language models (LLMs) will become the core of intelligent agents capable of reasoning, writing code, and interacting with tools.

But researchers at NVIDIA offer a different perspective: in most scenarios, such models are redundant. Their work can be taken over by small language models (SLMs) — compact systems with up to 10 billion parameters, capable of working locally and serving a single user.

The methodology replicates an experiment of a gradual transition from LLM to SLM. Researchers collected data on the use of agent systems and analyzed which tasks are repeated most often and do not require complex reasoning. These patterns were then grouped into clusters, after which small models capable of solving them autonomously were selected for each group. SLMs were retrained and tested in real chains of tool calls, with multiple retraining cycles, until they achieved performance comparable to LLM.

It turned out that up to 70% of typical tasks — document generation, coding, intent recognition, data formatting — can already be performed by small models without any loss of quality. In three case studies — MetaGPT, Open Operator, and Cradle — SLMs replace LLMs in 40 to 70% of calls. Models with 1.5–7 billion parameters are already comparable to previous generations of systems with 30–70 billion parameters, especially where stability and predictability are important rather than open reasoning. At the same time, the cost is reduced tenfold.

This transition enables the decentralization of agent architectures, where many small models operate autonomously without resorting to centralized cloud APIs. In the long term, researchers see this as not just a technical shift, but a structural shift from monolithic AI hubs to a distributed ecosystem of “small AI specialists.”

According to scientists, this will lower barriers to entry for developers, reduce privacy risks, and pave the way for more sustainable AI. Large models will remain a tool for complex reasoning, but everyday intelligence will live on a small scale — on devices closer to the user and the task at hand.

In recent years, AI has often been championed as a tool for sustainable development — a technology that should help humanity reduce emissions, improve energy efficiency, and accelerate the transition to a green economy. But there is little empirical evidence to support these expectations.

Researchers decided to examine the extent to which artificial intelligence is actually bringing the US closer to its carbon neutrality goal and whether it is creating new risks in the process. For their analysis, they collected quarterly data for 1996–2020 and built models that took into account not only the impact of AI itself, but also four related factors: energy policy uncertainty, green growth, and regulatory quality. This context allowed them to see not a linear relationship, but a complex system of mutual effects.

Methodologically, the study relies on NARDL (Nonlinear Autoregressive Distributed Lag), Wavelet Time Coherence, and Quantile-on-Quantile Regularized Least Squares — three approaches that allow us to capture asymmetric relationships and time lags. This is important because the impact of AI on emissions is heterogeneous: positive and negative shocks act at different speeds and at different times.

The results confirm that AI has a dual nature. Technological progress and algorithmic optimization reduce emissions by approximately 0.11% for every percentage point increase in AI activity. However, the computing infrastructure, data centers, and energy chains that power AI systems have the opposite effect, increasing emissions by 1.96%. Overall, the negative consequences outweigh the positive ones, especially given the low quality of regulation and political volatility in the energy sector.

At the same time, the quality of regulation is a decisive factor: a one-unit increase in this indicator reduces emissions by almost 47%, offsetting the harm caused by technological expansion. The green growth indicator also shows a sustained negative impact on emissions, but its effect is manifested with a significant time lag associated with the long cycle of development, implementation, and scaling of environmental innovations.

The authors conclude that artificial intelligence alone does not contribute to reducing emissions or bringing us closer to carbon neutrality goals. Its “environmental potential” is only realized in systems with stable institutions, transparent energy policies, and long-term investments in clean technologies. Without this, AI becomes yet another accelerator of carbon dependence — an integral part of the industry it was supposed to transform.

The authors track how symptoms of problematic Internet use (PIU) change in adolescents as they grow older. PIU is not a clinical diagnosis, but rather a behavioral pattern in which online activity begins to supplant other areas of life: study, communication, sleep, and emotional recovery. For adolescents, this condition often becomes a way to cope with stress, anxiety, or pressure from school and family.

The study lasted three years and included 302 school students aged 11–14. At each stage — from sixth to ninth grade — the authors measured PIU manifestations across 11 symptoms using the Problematic Internet Use Scale. The data was analyzed using network analysis (Gaussian Graphical Model), which allows determining which symptoms occupy central positions and how their interrelationships change over time. This approach reveals not only the level of addiction, but also its internal architecture — which states lead to others.

The study lasted three years and included 302 school students aged 11–14. At each stage — from sixth to ninth grade — the authors measured PIU manifestations across 11 symptoms using the Problematic Internet Use Scale. The data was analyzed using network analysis (Gaussian Graphical Model), which allows determining which symptoms occupy central positions and how their interrelationships change over time. This approach reveals not only the level of addiction, but also its internal architecture — which states lead to others.

The results show that PIU evolves with age. At an early stage (school years 6-7), the central symptom is “being comfortable” — the desire for the internet as a source of emotional comfort and relaxation. By years 8–9 at school, “being conflicted” comes to the fore — an internal contradiction between awareness of harm and the inability to control time spent online.
This shift in focus marks the transition from compensatory to dependent behavior. Over time, the network of symptoms becomes denser: the connections between elements strengthen, and PIU ceases to be a set of separate habits and becomes an interconnected system of symptoms, where emotional and behavioral responses support each other. There are no gender differences — the pattern of development is the same for boys and girls.

The authors emphasize that PIU is a dynamic and structurally stable form of behavior, not a temporary addiction. This requires differentiated prevention strategies. During the early school years, the key is to develop offline alternatives to comfort, teach digital hygiene and time management. During final years at school, the focus is on self-regulation, anxiety and the development of emotional stability.

In the long term, researchers suggest introducing systematic monitoring of PIU at the school level—regular surveys, inclusion of data in psychological and pedagogical records, and support for adolescents in risk groups.

Researcher Claudia Goldin analyzes why fertility rates in industrialized countries remain low despite economic growth, rising living standards, and the development of social institutions. Her work focuses not on economic incentives, but on the structural mismatch between the expansion of women's autonomy and the slow adaptation of male and institutional roles.

The basis of the analysis is the “matching problem” hypothesis. Goldin shows that the decline in fertility is not related to a rejection of family, but to a mismatch of expectations: the fewer men who are willing to share childcare and household responsibilities, the more rational it is for women to postpone having children or to forego motherhood altogether.

To test this idea, Goldin combines historical and macroeconomic data series for 12 countries—from the US and France to Japan and Korea—and applies cohort and comparative analysis using Total Fertility Rate, General Fertility Rate, and Cohort Fertility Rate indicators. She compares fertility trends with data on the distribution of domestic labor (OECD, Time Use Database) and models the impact of “traditional” and “modern” types of partnerships. In the model, the probability of having a child depends not only on income, but also on the chance that a woman will be able to combine motherhood with employment — a variable determined by the proportion of “modern men” in society.

The results show that fertility rates are directly linked to the gender gap in domestic workload. In countries with low TFRs (total fertility rates defined as average number of children born to one woman during her lifetime, where a rate below 2.1 indicates demographic decline), for example, in Japan (1.36) and Italy (1.27), women spend about three hours a day on household chores, while in Sweden and Denmark (where the TFR is closer to 1.7), they spend less than an hour. The smaller the gap, the higher the fertility rate. As women become more autonomous, this gap becomes a major factor in demographic decline: women adapt to new social conditions faster than institutions and men.

A comparison of two groups of countries confirms this trend. In the first group — the US, France, and the Nordic countries — the birth rate is declining moderately: the TFR remains within the range of 1.6–1.9, and the childlessness rate remains stable. In the second group—Italy, Spain, Korea, and Japan—the indicators are falling to the “lowest low” level (less than 1.5, in Korea—0.78), while the proportion of women without children is growing. According to the researcher, the gap is not so much related to the economy as to the pace of cultural change: in the latter group, women are adopting and reproducing new models of education and career faster than social institutions and perceptions of male roles are changing.

Goldin concludes that the “reverse side of fertility” is not a crisis of desire to have children, but a consequence of uncoordinated modernization. Women's autonomy is developing faster than the readiness of systems — labor, family, cultural — to support new forms of everyday equality. As a result, the decline in fertility becomes not an economic but a social problem, requiring a rethinking of the role of institutions in post-industrial society.

This study transfers Milgram's classic “six degrees of separation” experiment to a digital environment in order to understand how people navigate not social networks, but information networks. The paper focuses on comparing two search strategies: hub-driven, based on movement through large “centers” of knowledge (countries, cities, professions), and proximity-driven, where the search is based on semantic proximity, bringing the user closer to the goal step by step. The researchers ask what types of people tend to choose one strategy over another and what this choice says about the mechanisms of human thinking in the digital world.

The experiment was conducted in two stages — in 2020 and 2023 — with the participation of 802 people from the United States. Participants completed nine rounds of navigating Wikipedia: from one famous person to another (for example, from Barack Obama to Vincent van Gogh). Two modes were used: Speed-race (150-second limit) and Least-clicks (no more than seven transitions). For analysis, the researchers constructed a Wikipedia graph of 5.9 million pages and 133.6 million links, trained a 64-dimensional vector representation, and calculated two metrics: hierarchical index and semantic proximity to the target. This made it possible to quantitatively determine which strategy prevailed for each participant and how it correlated with individual characteristics.

The results showed that the strategies are in mutual tension (correlation –0.55): strengthening one reduces the intensity of the other. In speed tasks, hub-driven proved to be more effective —participants saved an average of 3.25 seconds and 0.41 clicks, while in minimum number of steps tasks, both strategies resulted in increased efficiency.

FIxed effects model explained up to 49% of the variance in results. Older participants were more likely to use proximity-driven (+0.013 per unit of age), while women (+0.040) and white respondents (+0.038) also showed a greater tendency toward thematic search, whereas younger participants were more likely to choose the hub-based approach (-0.008 per unit of age). Language background and political views did not play a significant role.

In addition, researchers identified a difference between geographical and professional navigation: in tasks such as Trump → Tchaikovsky, up to 92% of participants went through geographical pages, while in Jobs → Chaplin, only 21% did so. Geographical routes more often corresponded to the hub-driven model, while professional routes corresponded to the proximity-driven model.

The researchers demostrate that searching for information is not only a technical process, but also a cognitive one that reflects individual thinking patterns. Some users think “through hubs” — globally and structurally — while others think “through proximity” — locally and contextually.

These patterns are similar to social navigation strategies, where people look for the connections they need either through influence centers or through thematic similarities. The authors conclude that navigation in the knowledge space is a form of intellectual adaptation to digital excess: effective search engines and educational platforms must take into account that the path to knowledge is always individual and is determined by how a person builds and reduces connections.