Virtual collective consciousness (VCC) is a term rebooted and promoted by two behavioral scientists, Yousri Marzouki and Olivier Oullier in their 2012 Huffington Post article titled: "Revolutionizing Revolutions: Virtual Collective Consciousness and the Arab Spring", after its first appearance in 1999-2000. VCC is now defined as an internal knowledge catalyzed by social media platforms and shared by a plurality of individuals driven by the spontaneity, the homogeneity, and the synchronicity of their online actions. VCC occurs when a large group of persons, brought together by a social media platform think and act with one mind and share collective emotions. Thus, they are able to coordinate their efforts efficiently, and could rapidly spread their word to a worldwide audience. When interviewed about the concept of VCC that appeared in the book - Hyperconnectivity and the Future of Internet Communication - he edited, Professor of Pervasive Computing, Adrian David Cheok mentioned the following: "The idea of a global (collective) virtual consciousness is a bottom-up process and a rather emergent property resulting from a momentum of complex interactions taking place in social networks. This kind of collective behaviour (or intelligence) results from a collision between a physical world and a virtual world and can have a real impact in our life by driving collective action." == Etymology == In 1999-2000, Richard Glen Boire provided a cursory mention and the only occurrence of the term "Virtual collective consciousness" in his text as follows: The trend of technology is to overcome the limitations of the human body. And, the Web has been characterized as a virtual collective consciousness and unconsciousness The recent definition of VCC evolved from the first empirical study that provided a cyberpsychological insight into the contribution of Facebook to the 2011 Tunisian revolution. In this study, the concept was originally called "collective cyberconsciousness". The latter is an extension of the idea of "collective consciousness" coupled with "citizen media" usage. The authors of this study also made a parallel between this original definition of VCC and other comparable concepts such as Durkheim's collective representation, Žižek's "collective mind" or Boguta's "new collective consciousness" that he used to describe the computational history of the Internet shutdown during the Egyptian revolution. Since VCC is the byproduct of the network's successful actions, then these actions must be timely, acute, rapid, domain-specific, and purpose-oriented to successfully achieve their goal. Before reaching a momentum of complexity, each collective behavior starts by a spark that triggers a chain of events leading to a crystallized stance of a tremendous amount of interactions. Thus, VCC is an emergent global pattern from these individual actions. In 2012, the term virtual collective consciousness resurfaced and was brought to light after extending its applications to the Egyptian case and the whole social networking major impact on the success of the so-called Arab Spring. Moreover, the acronym VCC was suggested to identify the theoretical framework covering on-line behaviors leading to a virtual collective consciousness. Hence, online social networks have provided a new and faster way of establishing or modifying "collective consciousness" that was paramount to the 2011 uprisings in the Arab world. == Theoretical underpinnings of VCC == Various theoretical references in fields ranging from sociology to computer science were mentioned in order to account for the key features that render the framework for a virtual collective consciousness. The following list is not exhaustive, but the references it contains are often highlighted: Émile Durkheim's collective representations are at the heart of VCC since collectivity taken decisions according to Durkheim's assumptions will approve or disapprove individuals' actions and help them eventually reach their final goal. Marshall McLuhan's global village: The shrinking of our big world to a small place called cyberspace is made possible by technological extensions of human consciousness. Carl Jung's collective unconscious: When a society witnesses significant changes, the anchoring of archetypal images (e.g., political leaders) seems to be deeply rooted in individuals' collective unconscious that is likely to bias their political choices. Individual memories of public events were also supposed to convey a "collective awareness" that can be subconsciously altered by the instantaneous spread of information through social networking around the world. Daniel Wegner's transactive memory (TM): social-networking platforms such as Facebook during the Tunisian revolution or Twitter during the Egyptian revolution served as placeholders of a VCC where information can be harnessed and steered to the highly specific revolutionary purpose. Although research on TM was originally limited to couples, small groups, and organizations, recent studies strongly suggest that an effective TM can operate on a very large scale too. James Surowiecki's wisdom of crowds Collective influence algorithm: The CI (Collective influence) algorithm is effective in finding influential nodes in a variety of networks, including social networks, communication networks, and biological networks. It has been used to identify influencers on social-media platforms, to identify key nodes in transportation networks, and to identify potential drug-targets in biological networks. == Some illustrations of VCC == Besides the studied effect of social networking on the Tunisian and Egyptian revolutions, the former via Facebook and the latter via Twitter other applications were studied under the prism of VCC framework: The Whitacre's virtual choir: A compelling example of the degree of autonomy and self-identity members of a spontaneously created network through a VCC is Eric Whitacre's unique musical project that involved a collection of singers performing remotely to create a virtual Choir. The effect of all the voices illustrated a genuine virtual collective empathy merging the artist's mind with all the singers through his silent conducting gestures. The Harlem Shake dance: The Bitcoin protocol: It was questioned whether or not the Bitcoin protocol can morph into virtual collective consciousness. The Byzantine generals problem was used as an analogy to understand the behavioral complexity of the community of Bitcoin's users. Artificial Social Networking Intelligence (ASNI): refers to the application of artificial intelligence within social networking services and social media platforms. It encompasses various technologies and techniques used to automate, personalize, enhance, improve, and synchronize users' interactions and experiences within social networks. ASNI is expected to evolve rapidly, influencing how we interact online and shaping our digital experiences. Transparency, ethical considerations, media influence bias, and user control over data will be crucial to ensure responsible development and positive impact.
Large language model
A large language model (LLM) is a neural network trained on a vast amount of text for natural language processing tasks, especially language generation. LLMs can typically generate, summarize, translate and analyze text in many contexts, and are a foundational technology behind modern chatbots. Biased or inaccurate training data can make an LLM's output less reliable. As of 2026, the most capable LLMs are based on transformer architectures, which, according to the 2017 paper "Attention Is All You Need", can be more efficient and parallelizable than earlier statistical and recurrent neural network models. Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety. == History == Before the emergence of transformer-based models in 2017, some language models were considered large relative to the computational and data constraints of their time. In the early 1990s, IBM's statistical models pioneered word alignment techniques for machine translation, laying the groundwork for corpus-based language modeling. In 2001, a smoothed n-gram model, such as those employing Kneser–Ney smoothing, trained on 300 million words, achieved state-of-the-art perplexity on benchmark tests. During the 2000s, with the rise of widespread internet access, researchers began compiling massive text datasets from the web ("web as corpus") to train statistical language models. Moving beyond n-gram models, researchers started in 2000 to use neural networks as language models. Following the breakthrough of deep neural networks in image classification around 2012, similar architectures were adapted for language tasks. This shift was marked by the development of word embeddings (e.g., Word2Vec by Mikolov in 2013) and sequence-to-sequence (seq2seq) models using LSTM. In 2016, Google transitioned its translation service to neural machine translation (NMT), replacing statistical phrase-based models with deep recurrent neural networks. These early NMT systems used LSTM-based encoder-decoder architectures, as they preceded the invention of transformers. At the 2017 NeurIPS conference, Google researchers introduced the transformer architecture in their landmark paper "Attention Is All You Need". This paper's goal was to improve upon 2014 seq2seq technology, and was based mainly on the attention mechanism developed by Bahdanau et al. in 2014. The following year in 2018, BERT was introduced and quickly became "ubiquitous". Though the original transformer has both encoder and decoder blocks, BERT is an encoder-only model. Academic and research usage of BERT began to decline in 2023, following rapid improvements in the abilities of decoder-only models (such as GPT) to solve tasks via prompting. Although decoder-only GPT-1 was introduced in 2018, it was GPT-2 in 2019 that caught widespread attention because OpenAI claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use. GPT-3 in 2020 went a step further and as of 2025 is available only via API with no offering of downloading the model to execute locally. But it was the consumer-facing chatbot ChatGPT in late 2022 that received extensive media coverage and public attention by 2023. The 2023 GPT-4 was praised for its increased accuracy and as a "holy grail" for its multimodal capabilities. OpenAI did not reveal the high-level architecture and the number of parameters of GPT-4. The release of ChatGPT led to an uptick in LLM usage across several research subfields of computer science, including robotics, software engineering, and societal impact work. In 2024, OpenAI released the reasoning model OpenAI o1, which generates long chains of thought before returning a final answer. Many LLMs with parameter counts comparable to those of OpenAI's GPT series have been developed. Since 2022, weights-available models have been gaining popularity, especially at first with BLOOM and LLaMA, though both have restrictions on usage and deployment. Mistral AI's open-weight models Mistral 7B and Mixtral 8x7B have a more permissive Apache License. In January 2025, DeepSeek released DeepSeek R1, a 671-billion-parameter open-weight model that performs comparably to OpenAI o1 but at a much lower price per token for users. Since 2023, many LLMs have been trained to be multimodal, having the ability to also process or generate other types of data, such as images, audio, or 3D meshes. Open-weight LLMs have become more influential since 2023. Per Vake et al. (2025), community-driven contributions to open-weight models improve their efficiency and performance via collaborative platforms such as Hugging Face. == Dataset preprocessing == === Tokenization === As machine learning algorithms process numbers rather than text, the text must be converted to numbers. In the first step, a vocabulary is decided upon, then integer indices are arbitrarily but uniquely assigned to each vocabulary entry, and finally, an embedding is associated with the integer index. Algorithms include byte-pair encoding (BPE) and WordPiece. There are also special tokens serving as control characters, such as [MASK] for masked-out token (as used in BERT), and [UNK] ("unknown") for characters not appearing in the vocabulary. Also, some special symbols are used to denote special text formatting. For example, "Ġ" denotes a preceding whitespace in RoBERTa and GPT and "##" denotes continuation of a preceding word in BERT. For example, the BPE tokenizer used by the legacy version of GPT-3 would split tokenizer: texts -> series of numerical "tokens" as Tokenization also compresses the datasets. Because LLMs generally require input to be an array that is not jagged, the shorter texts must be "padded" until they match the length of the longest one. ==== Byte-pair encoding ==== As an example, consider a tokenizer based on byte-pair encoding. In the first step, all unique characters (including blanks and punctuation marks) are treated as an initial set of n-grams (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) n-grams that most frequently occur together are then again merged into even lengthier n-gram, until a vocabulary of prescribed size is obtained. After a tokenizer is trained, any text can be tokenized by it, as long as it does not contain characters not appearing in the initial-set of uni-grams. === Dataset cleaning === In the context of training LLMs, datasets are typically cleaned by removing low-quality, duplicated, or toxic data. Cleaned datasets can increase training efficiency and lead to improved downstream performance. A trained LLM can be used to clean datasets for training a further LLM. With the increasing proportion of LLM-generated content on the web, data cleaning in the future may include filtering out such content. LLM-generated content can pose a problem if the content is similar to human text (making filtering difficult) but of lower quality (degrading performance of models trained on it). === Synthetic data === Training of largest language models might need more linguistic data than naturally available, or that the naturally occurring data is of insufficient quality. In these cases, synthetic data might be used. == Training == An LLM is a type of foundation model (large X model) trained on language. LLMs can be trained in different ways. In particular, GPT models are first pretrained to predict the next word on a large amount of data, before being fine-tuned. === Cost === Substantial infrastructure is necessary for training the largest models. The tendency towards larger models is visible in the list of large language models. For example, the training of GPT-2 (i.e. a 1.5-billion-parameter model) in 2019 cost $50,000, while training of the PaLM (i.e. a 540-billion-parameter model) in 2022 cost $8 million, and Megatron-Turing NLG 530B (in 2021) cost around $11 million. The qualifier "large" in "large language model" is inherently vague, as there is no definitive threshold for the number of parameters required to qualify as "large". === Fine-tuning === Before being fine-tuned, most LLMs are next-token predictors. The fine-tuning shapes the LLM's behavior via techniques like reinforcement learning from human feedback (RLHF) or constitutional AI. Instruction fine-tuning is a form of supervised learning used to teach LLMs to follow user instructions. In 2022, OpenAI demonstrated InstructGPT, a version of GPT-3 similarly fine-tuned to follow instructions. Reinforcement learning from human feedback (RLHF) involves training a reward model to predict which text humans prefer. Then, the LLM can be fine-tuned through reinforcement learning to better satisfy this reward model. Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers. == Architecture == LLMs are generally based on the tra
ICAART
The International Conference on Agents and Artificial Intelligence (ICAART) is a meeting point for researchers (among others) with interest in the areas of Agents and Artificial Intelligence. There are 2 tracks in ICAART, one related to Agents and Distributed AI in general and the other one focused in topics related to Intelligent Systems and Computational Intelligence. The conference program is composed of several different kind of sessions like technical sessions, poster sessions, keynote lectures, tutorials, special sessions, doctoral consortiums, panels and industrial tracks. The papers presented in the conference are made available at the SCITEPRESS digital library, published in the conference proceedings and some of the best papers are invited to a post-publication with Springer. ICAART's first edition was in 2009 counting with several keynote speakers like Marco Dorigo, Edward H. Shortliffe and Eduard Hovy. Since then, the conference had several other invited speakers like Katia Sycara, Nick Jennings, Robert Kowalski, Boi Faltings and Tim Finin. Bart Selman is one of the names confirmed for the next edition of this conference. Since 2012 the conference is held in conjunction with 2 other conferences: the International Conference on Operations Research and Enterprise Systems (ICORES) and the International Conference on Pattern Recognition Applications and Methods (ICPRAM). == Areas == === Agents === Agent communication languages Cooperation and Coordination Distributed Problem Solving Economic Agent Models Emotional Intelligence Group Decision Making Intelligent Auctions and Markets Mobile Agents Multi-agent systems Negotiation and Interaction Protocols Nep News Detection Agent Models and Architectures Physical Agents at Work Privacy, Safety and Security Programming Environments and Languages Robot and Multi-Robot Systems Self Organizing Systems Semantic Web Simulation Swarm Intelligence Task Planning and Execution Transparency and Ethical Issues Agent-Oriented Software Engineering Web Intelligence Agent Platforms and Interoperability Autonomous systems Cloud Computing and Its Impact Cognitive robotics Collective Intelligence Conversational Agents === Artificial intelligence === AI and Creativity Deep Learning Evolutionary Computing Fuzzy Systems Hybrid Intelligent Systems Industrial Applications of AI Intelligence and Cybersecurity Intelligent User Interfaces Knowledge Representation and Reasoning Knowledge-Based Systems Ambient Intelligence Machine learning Model-Based Reasoning Natural Language Processing Neural Networks Ontologies Planning and Scheduling Social Network Analysis Soft Computing State Space Search Bayesian Networks Uncertainty in AI Vision and Perception Visualization Big Data Case-Based Reasoning Cognitive Systems Constraint Satisfaction Data Mining Data Science == Editions == === ICAART 2023 – Lisbon, Portugal === === ICAART 2020 – Valletta, Malta === === ICAART 2019 – Prague, Czech Republic === Proceedings - Proceedings of the 11th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-350-6 Proceedings - Proceedings of the 11th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-350-6 === ICAART 2018 – Funchal, Madeira, Portugal === Proceedings - Proceedings of the 10th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-275-2 Proceedings - Proceedings of the 10th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-275-2 === ICAART 2017 – Porto, Portugal === Proceedings - Proceedings of the 9th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-219-6 Proceedings - Proceedings of the 9th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-220-2 === ICAART 2016 – Rome, Italy === Proceedings - Proceedings of the 8th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-172-4 Proceedings - Proceedings of the 8th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-172-4 === ICAART 2015 – Lisbon, Portugal === Proceedings - Proceedings of the 7th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-073-4 Proceedings - Proceedings of the 7th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-074-1 === ICAART 2014 – ESEO, Angers, Loire Valley, France === Proceedings - Proceedings of the 6th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-758-015-4 Proceedings - Proceedings of the 6th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-758-016-1 === ICAART 2013 – Barcelona, Spain === Proceedings - Proceedings of the 5th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-8565-38-9 Proceedings - Proceedings of the 5th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-8565-39-6 === ICAART 2012 – Vilamoura, Algarve, Portugal === Proceedings - Proceedings of the 4th International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-8425-95-9 Proceedings - Proceedings of the 4th International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-8425-96-6 === ICAART 2011 – Rome, Italy === Proceedings - Proceedings of the 3rd International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-8425-40-9 Proceedings - Proceedings of the 3rd International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-8425-41-6 === ICAART 2010 – Valencia, Spain === Proceedings - Proceedings of the 2nd International Conference on Web Information Systems and Technologies - Volume 1. ISBN 978-989-674-021-4 Proceedings - Proceedings of the 2nd International Conference on Web Information Systems and Technologies - Volume 2. ISBN 978-989-674-022-1 === ICAART 2009 – Porto, Portugal === Proceedings - Proceedings of the 1st International Conference on Web Information Systems and Technologies. ISBN 978-989-8111-66-1
Model collapse
Model collapse, also known by other names such as "AI inbreeding", "AI cannibalism", "Habsburg AI", and "model autophagy disorder" or "MAD" is a phenomenon noted in artificial intelligence studies, where machine learning models gradually degrade due to errors coming from uncurated synthetic data, or due to training on the outputs of another model such as prior versions of itself. It is unclear to what extent the phenomenon threatens the long-term development of such models, and some techniques have been proposed to mitigate the effect. == Characteristics == Shumailov et al. coined the term to describe two specific stages to the degradation of machine learning models: early model collapse and late model collapse: In early model collapse, the model begins losing information about the tails of the distribution – mostly affecting minority data. Later work highlighted that early model collapse is hard to notice, since overall performance may appear to improve, while the model loses performance on minority data. In late model collapse, the model loses a significant proportion of its performance, confusing concepts and losing most of its variance. == Mechanism == Using synthetic data as training data can lead to issues with the quality and reliability of the trained model. Model collapse occurs for three main reasons: functional approximation errors sampling errors learning errors Importantly, it happens in even the simplest of models, where not all of the error sources are present. In more complex models the errors often compound, leading to faster collapse. == Disagreement over real-world impact == Some researchers and commentators on model collapse warn that the phenomenon could fundamentally threaten future generative AI development: As AI-generated data is shared on the Internet, it will inevitably end up in future training datasets, which are often crawled from the Internet. If training on "slop" (large quantities of unlabeled synthetic data) inevitably leads to model collapse, this could therefore pose a difficult problem. However, recently, other researchers have disagreed with this argument, showing that if synthetic data accumulates alongside human-generated data, model collapse is avoided. The researchers argue that data accumulating over time is a more realistic description of reality than deleting all existing data every year, and that the real-world impact of model collapse may not be as catastrophic as feared. An alternative branch of the literature investigates the use of machine learning detectors and watermarking to identify model generated data and filter it out. == Mathematical models of the phenomenon == === 1D Gaussian model === In 2024, a first attempt has been made at illustrating collapse for the simplest possible model — a single dimensional normal distribution fit using unbiased estimators of mean and variance, computed on samples from the previous generation. To make this more precise, we say that original data follows a normal distribution X 0 ∼ N ( μ , σ 2 ) {\displaystyle X^{0}\sim {\mathcal {N}}(\mu ,\sigma ^{2})} , and we possess M 0 {\displaystyle M_{0}} samples X j 0 {\displaystyle X_{j}^{0}} for j ∈ { 1 , … , M 0 } {\displaystyle j\in {\{\,1,\dots ,M_{0}\,{}\}}} . Denoting a general sample X j i {\displaystyle X_{j}^{i}} as sample j ∈ { 1 , … , M i } {\displaystyle j\in {\{\,1,\dots ,M_{i}\,{}\}}} at generation i {\displaystyle i} , then the next generation model is estimated using the sample mean and variance: μ i + 1 = 1 M i ∑ j X j i ; σ i + 1 2 = 1 M i − 1 ∑ j ( X j i − μ i + 1 ) 2 . {\displaystyle \mu _{i+1}={\frac {1}{M_{i}}}\sum _{j}X_{j}^{i};\quad \sigma _{i+1}^{2}={\frac {1}{M_{i}-1}}\sum _{j}(X_{j}^{i}-\mu _{i+1})^{2}.} Leading to a conditionally normal next generation model X j i + 1 | μ i + 1 , σ i + 1 ∼ N ( μ i + 1 , σ i + 1 2 ) {\displaystyle X_{j}^{i+1}|\mu _{i+1},\;\sigma _{i+1}\sim {\mathcal {N}}(\mu _{i+1},\sigma _{i+1}^{2})} . In theory, this is enough to calculate the full distribution of X j i {\displaystyle X_{j}^{i}} . However, even after the first generation, the full distribution is no longer normal: It follows a variance-gamma distribution. To continue the analysis, instead of writing the probability density function at each generation, it is possible to explicitly construct them in terms of independent random variables using Cochran's theorem. To be precise, μ 1 {\displaystyle \mu _{1}} and σ 1 {\displaystyle \sigma _{1}} are independent, with μ 1 ∼ N ( μ , σ 2 M 0 ) {\displaystyle \mu _{1}\sim {\mathcal {N}}\left(\mu ,{\frac {\sigma ^{2}}{M_{0}}}\right)} and ( M 0 − 1 ) σ 1 2 ∼ σ 2 Γ ( M 0 − 1 2 , 1 2 ) {\displaystyle (M_{0}-1)\,\sigma _{1}^{2}\sim \sigma ^{2}\,\Gamma \left({\frac {M_{0}-1}{2}},{\frac {1}{2}}\right)} , following a Gamma distribution. Denoting with Z {\displaystyle Z} Gaussian random variables distributed according to N ( 0 , 1 ) {\displaystyle {\mathcal {N}}(0,1)} and with S i {\displaystyle S^{i}} random variables distributed with 1 M i − 1 − 1 Γ ( M i − 1 − 1 2 , 1 2 ) {\displaystyle {\frac {1}{M_{i-1}-1}}\Gamma \left({\frac {M_{i-1}-1}{2}},{\frac {1}{2}}\right)} , it turns out to be possible to write samples at each generation as X j 0 = μ + σ Z j 0 , {\textstyle X_{j}^{0}=\mu +\sigma Z_{j}^{0},} X j 1 = μ + σ M 0 Z 1 + σ S 1 Z j 1 , {\textstyle X_{j}^{1}=\mu +{\frac {\sigma }{\sqrt {M_{0}}}}Z^{1}+\sigma {\sqrt {S^{1}}}Z_{j}^{1},} and more generally X j n = μ + σ M 0 Z 1 + σ M 1 S 1 Z 2 + ⋯ + σ M n − 1 S 1 × ⋯ × S n − 1 Z n + σ S 1 × ⋯ × S n Z j n . {\displaystyle X_{j}^{n}=\mu +{\frac {\sigma }{\sqrt {M_{0}}}}Z^{1}+{\frac {\sigma }{\sqrt {M_{1}}}}{\sqrt {S^{1}}}Z^{2}+\dots +{\frac {\sigma }{\sqrt {M_{n-1}}}}{\sqrt {S^{1}\times \dots \times S^{n-1}}}Z^{n}+\sigma {\sqrt {S^{1}\times \dots \times S^{n}}}Z_{j}^{n}.} Note, that these are not joint distributions, as Z n {\displaystyle Z^{n}} and S n {\displaystyle S^{n}} depend directly on Z j n − 1 {\displaystyle Z_{j}^{n-1}} , but when considering X j n {\displaystyle X_{j}^{n}} on its own the formula above provides all the information about the full distribution. To analyse the model collapse, we can first calculate variance and mean of samples at generation n {\displaystyle n} . This would tell us what kind of distributions we expect to arrive at after n {\displaystyle n} generations. It is possible to find its exact value in closed form, but the mean and variance of the square root of gamma distribution are expressed in terms of gamma functions, making the result quite clunky. Following, it is possible to expand all results to second order in each of 1 / M i {\displaystyle 1/M_{i}} , assuming each sample size to be large. It is then possible to show that 1 σ 2 Var ( X j n ) = 1 M 0 + 1 M 1 + ⋯ + 1 M n − 1 + 1 + O ( M i − 2 ) . {\displaystyle {\frac {1}{\sigma ^{2}}}\operatorname {Var} (X_{j}^{n})={\frac {1}{M_{0}}}+{\frac {1}{M_{1}}}+\dots +{\frac {1}{M_{n-1}}}+1+{\mathcal {O}}\left(M_{i}^{-2}\right).} And if all sample sizes M i = M {\displaystyle M_{i}=M} are constant, this diverges linearly as n → ∞ {\displaystyle n\to \infty } : Var ( X j n ) = σ 2 ( 1 + n M ) ; E ( X j n ) = μ . {\displaystyle \operatorname {Var} (X_{j}^{n})=\sigma ^{2}\left(1+{\frac {n}{M}}\right);\quad \mathbb {E} (X_{j}^{n})=\mu .} This is the same scaling as for a single dimensional Gaussian random walk. However, divergence of the variance of X j n {\displaystyle X_{j}^{n}} does not directly provide any information about the corresponding estimates of μ n + 1 {\displaystyle \mu _{n+1}} and σ n + 1 {\displaystyle \sigma _{n+1}} , particularly how different they are from the original μ {\displaystyle \mu } and σ {\displaystyle \sigma } . It turns out to be possible to calculate the distance between the true distribution and the approximated distribution at step n + 1 {\displaystyle n+1} , using the Wasserstein-2 distance (which is also sometimes referred to as risk): E [ W 2 2 ( N ( μ , σ 2 ) , N ( μ n + 1 , σ n + 1 2 ) ) ] = 3 2 σ 2 ( 1 M 0 + 1 M 1 + ⋯ + 1 M n ) + O ( M i − 2 ) , {\displaystyle \mathbb {E} \left[\mathbb {W} _{2}^{2}\left({\mathcal {N}}(\mu ,\sigma ^{2}),{\mathcal {N}}(\mu _{n+1},\sigma _{n+1}^{2})\right)\right]={\frac {3}{2}}\sigma ^{2}\left({\frac {1}{M_{0}}}+{\frac {1}{M_{1}}}+\dots +{\frac {1}{M_{n}}}\right)+{\mathcal {O}}\left(M_{i}^{-2}\right),} Var [ W 2 2 ( N ( μ , σ 2 ) , N ( μ n + 1 , σ n + 1 2 ) ) ] = 1 2 σ 4 ( 3 M 0 2 + 3 M 1 2 + ⋯ + 3 M n 2 + ∑ i ≠ j 4 M i M j ) + O ( M i − 3 ) . {\displaystyle \operatorname {Var} \left[\mathbb {W} _{2}^{2}\left({\mathcal {N}}(\mu ,\sigma ^{2}),{\mathcal {N}}(\mu _{n+1},\sigma _{n+1}^{2})\right)\right]={\frac {1}{2}}\sigma ^{4}\left({\frac {3}{M_{0}^{2}}}+{\frac {3}{M_{1}^{2}}}+\dots +{\frac {3}{M_{n}^{2}}}+\sum _{i\neq j}{\frac {4}{M_{i}M_{j}}}\right)+{\mathcal {O}}\left(M_{i}^{-3}\right).} This directly shows why model collapse occurs in this simple model. Due to errors from re-sampling the approximated distribution, each generation ends up corresponding to a
International Olympiad in Artificial Intelligence
The International Olympiad in Artificial Intelligence (IOAI) is an annual International Science Olympiad in the field of artificial intelligence (AI) for secondary education students under the age of 20. The first IOAI was held in Burgas, Bulgaria, in 2024. Each country or territory may send up to two teams, each consisting of up to four students supported by one leader. Participants are selected through a multi-stage National Olympiad in Artificial Intelligence (NOAI) and/or a Regional Olympiad such as the NAOAI or APOAI. Participants at the IOAI compete on an individual basis. As of 2025, there were 61 countries and territories participating in the IOAI. Three hundred students participated in IOAI 2025. As of 2026, 130 countries and territories are accredited for participation in the IOAI. == Competition Structure == The IOAI consists of three contests: the Individual Contest, the Team Challenge, and the GAITE contest. Medals are awarded based solely on the Individual Contest. === Individual Contest === The Individual Contest is the main competition of the IOAI in which contestants compete individually on separate computers and are not permitted to communicate during the contest. Medals are awarded solely on the basis of the total score from the two-day Individual Contest. The Individual Contest consists of two on-site contest days (six hours per day), preceded by an at-home practice round and an on-site practice session. In IOAI 2025, three at-home problems were released for preparation approximately one month before the on-site contest. Results from this at-home round do not affect final results. The first on-site contest day (Individual Contest 1) comprises three tasks as extensions and continuations of the at-home tasks, while the second day (Individual Contest 2) comprises two or three tasks which are novel and different from the at-home tasks. The Individual Contest tasks span various AI domains such as machine learning, natural language processing, and computer vision. The IOAI 2025 contest rules describe tasks as requiring typical machine-learning workflows, including writing code, fitting models on training data, and running inference on test data, using identical local machines and GPU resources (minimum 24 GB RAM). Tasks, datasets, and submissions are handled through a contest platform (Bohrium), including a web-based Jupyter notebook environment for GPU access. Internet access is restricted to a whitelist of documentation sites and an integrated compact large language model accessible within the platform. The use of external APIs are prohibited unless a task explicitly allows them. In IOAI 2025, each contest task was scored up to 100 points and could include multiple subtasks. Scores are normalized using a baseline solution and a maximum score derived from either a Scientific Committee solution or the best contestant submission. Contestants can view only their own scores during the contest; a live scoreboard may be available publicly outside the contest hall but is not permitted to be viewed by contestants during the contest. For non-English-speaking teams, the IOAI hold a translation session beginning three hours before each contest day in which team leaders review and may amend machine-translated task statements; translations must match the English original and are published after the contest. The IOAI committee also enforces quarantine restrictions during these translation sessions, where neither contestants or team leaders may not use cell phones, laptops, and other communication devices. === Team Challenge === The Team Challenge is a team-based component of the IOAI. The results of this part do not affect the distribution of medals. The IOAI 2025 rules describe it as a “creative and AI-oriented challenge” in which a team's contestants sit together and cooperate, with the format varying by year. In IOAI 2024, teams worked with existing AI image and video generation tools to produce a visual result. In IOAI 2025, teams were assigned to program a robot to complete various tasks. === GAITE Contest === The GAITE (Global AI Talent Empowerment) contest is a simplified version of the individual contest with a separate scoreboard, where participants may ask for hints. It is designed for countries and territories with limited International Science Olympiads history, and it awards alternative prizes instead of medals. == Awards Distribution == The top 50% of the participants in the individual contest receive gold, silver and bronze medals in ratio of 1:2:3, respectively. The top three individuals receive honorary trophies. As in other International Science Olympiads, if an individual is in the top 50% on one of the days, but does not receive a medal, they receive an honorary mention during the awards ceremony. The GAITE contest has similar cutoff logic, but receives a reward instead of a medal. The top three teams in the Team Challenge receive trophies. == National selection and regional competitions == National delegations are selected through country-level qualification processes referred to as National Olympiads in Artificial Intelligence (NOAI) or equivalent, which are widely known for their low success rates. Although the total number of participants worldwide is not published, available data indicate exceptionally competitive national pools; for example, Brazil reports over 716,000 competitors, while Russia reports more than 72,000. In addition, Regional Olympiads (for example, APOAI or NAOAI) provide continent-level competition and preparation platforms in most regions. === National Selection (National Olympiads in Artificial Intelligence) === Participating countries and territories select their students for the IOAI through a National Olympiad in Artificial Intelligence (NOAI) or an equivalent process. The names of these selection processes differ by country, but almost all of them (excluding newer countries participating in the GAITE contest) have in common that the process comprises multiple and/or extremely rigorous selection stages. United States / Canada – The USA–North America AI Olympiad (USAAIO) is a three-round process including an invitational in-person round and a subsequent selection camp, after which a national delegation is selected for IOAI. Russia – The Russian Olympiad in Artificial Intelligence is organized as a multi-stage process (training, qualification, main round, final). Organizers reported 72,316 registrations for the training round and 52,260 registrations for the qualifying round in one season, with tasks spanning mathematics, algorithms/programming, and machine learning; 977 students were disqualified following plagiarism checks. Japan – Japan's national selection consists of multiple stages, beginning with the Japan Olympiad in Artificial Intelligence (JOAI), a large-scale Kaggle-style competition. High-performing participants advance through additional assessment stages, including written solution reports and technical interviews. From this process, eight students are selected for the APOAI team, with four ultimately chosen to represent Japan at the IOAI. Brazil – Brazil's National Olympiad in Artificial Intelligence (ONIA) is conducted as a large competition which consists of progressive rounds of evaluation. It identifies 28 top students from over 716,000 competitors, four of which are selected for the IOAI. The competition is held in four phases across two cycles, including a two-step third phase and a final training-and-evaluation phase that selects a four-student national team. Singapore – Singapore's national Olympiad consists of two rounds: an online preliminary round (300 MCQs in 3 hours) selects the top 150 performers to advance to the final assessment, which includes both theory questions and Python programming tasks. Additional training and selection may follow the finals for top performers. Poland – The Polish AI Olympiad adopts a two-stage structure: an open online first stage (at-home tasks) and a second-stage competitive camp with 30 selected participants competing for a four-person IOAI team. France – The Olympiades Françaises d'Intelligence Artificielle (OFIA), organized by France-IOI, follow a three-stage structure consisting of an open online qualification round, a second selection round, and a multi-day national training camp and final in Paris. Bangladesh – The Bangladesh AI Olympiad (BdAIO) selects competitors in three rounds: the online preliminary round, the national finals, and the team selection camp. In 2025, 406 participants competed in the national finals. Norway – The Norwrgian AI Olympiad (NOKI) is a three-stage selection system; however, unlike other countries, its first two rounds are shared with the Norwegian Informatics Olympiad. The national Olympiad reports 1,180 participants in the first round. Hong Kong – The national Olympiad reported more than 800 preliminary-round entrants, narrowing through multiple rounds to 25 finalists, with a subsequent
Smart speaker industry in South Korea
Smart speakers, or AI speakers, have been developed by multiple domestic electronics and telecommunications firms in South Korea. Since their introduction to the local market in 2016, they have been used by millions of people in the country. == Brands == === Google === In September 2018, Google Home (including the Google Home Mini) launched in South Korea. Running Google Assistant, it featured simultaneous recognition of two languages among a total of seven, including Korean. At launch, it could play music from Bugs!, in addition to YouTube. === Kakao === In November 2017, Kakao launched the Kakao Mini, featuring integrated KakaoTalk functionality. === KT === KT launched the GiGA Genie smart speaker in January 2017, using a Harman Kardon speaker. In November 2017, KT announced GiGA Genie LTE, a portable AI speaker with LTE support. They also released a mini speaker called GiGA Genie Buddy. In 2018, KT created a special version of GiGa Genie with a screen for use in hotels. On 29 April 2019, KT announced the GiGA Genie Table TV, a consumer-oriented smart speaker with a display. It featured paid TV access through Wi-Fi. Based on usage data from the hotel model, KT decided not to add a touchscreen. The Table TV also featured a limited-access "personalized-text-to-speech technology" which could use parents' voice recording inputs to read children books. In February 2022, KT began rolling out Amazon Alexa integration into its speakers for English support. === Naver === In August 2017, Naver announced the Wave smart speaker, operating on Clova. In October 2017, Naver launched the Friends smart speaker, which were designed based on Line characters. ==== LG Uplus ==== In December 2017, LG Uplus launched the Friends+ speaker with Naver, operating on U+ Home AI. === Samsung === In August 2018, Samsung announced the Samsung Galaxy Home in partnership with Spotify. The original size was delayed, while the Galaxy Home Mini appeared briefly as a bonus for Samsung Galaxy S20 preorders in South Korea in February 2020. === SK Telecom === SK Telecom launched the Nugu smart speaker in September 2016, using an Astell & Kern audio system. In August 2017, SKT released a portable speaker named Nugu mini. In July 2018, SKT launched the Nugu Candle, featuring expanded mood lighting. The first-generation Nugu was subsequently discontinued. On 18 April 2019, SKT released the NUGU Nemo AI, which featured a display and JBL stereo speaker. In August 2019, SKT collaborated with SM Entertainment, incorporating functions related to the agency's artists into Nugu. In January 2022, SKT showcased the NUGU Candle SE, introducing Alexa support. == Usage == In 2018, approximately 3 million people in South Korea used smart speakers. According to data from KT in 2018, the most common commands to its speakers were for controlling televisions. Based on a broader survey in 2017, music was selected as the most frequent use case. By 2018, smart speaker companies were partnering with reading and other education services, adding potential use-cases for children. By 2022, smart speakers were being utilized by the South Korean government. SKT, in partnership with 70 regional governments, distributed smart speakers to 12,000 senior citizens living alone. The government paid for monthly subscriptions to help seniors stay mentally engaged. Naver made an agreement with the Seoul Metropolitan Government to provide Clova CareCall, an automated health checkup program to hundreds of senior citizens living alone. KT's AI care service included an emergency dispatch call function and medication notifications. == Criticism == === Communication === In a survey of 300 users in 2017, approximately half reported having some type of communication issue with their smart speakers. === Privacy === South Korean smart speakers sparked privacy concerns when they were found to be collecting and documenting user audio data in 2019. The speaker companies responded that only a minority of data was collected and that it was anonymized. They stated that such recordings were collected for performance improvements.
AAAI Conference on Artificial Intelligence
The AAAI Conference on Artificial Intelligence is a leading international academic conference in artificial intelligence held annually. It ranks 4th in terms of H5 Index in Google Scholar's list of top AI publications, after ICLR, NeurIPS, and ICML. It is supported by the Association for the Advancement of Artificial Intelligence (AAAI), after which it is named. Precise dates vary from year to year, but paper submissions are generally due at the end of August to beginning of September, and the conference is generally held during the following February. The first AAAI was held in 1980 at Stanford University, Stanford California. During AAAI-20 conference, AI pioneers and 2018 Turing Award winners (often referred to as the Nobel Prize of Computing) Yann LeCun and Yoshua Bengio, among eight other researchers, were honored as the AAAI 2020 Fellows. Along with other conferences such as NeurIPS and ICML, AAAI uses an artificial-intelligence algorithm to assign papers to reviewers. == Sponsors == Many leading technology companies, including Google, Microsoft, Amazon (company), IBM, Baidu, Bytedance, and Huawei, generously sponsor and participate in AAAI to publish and showcase their latest theoretical and applied research. Sponsoring companies also actively recruit AI talents at the conference. == Locations == AAAI-2026 Singapore Expo, Singapore AAAI-2025 Pennsylvania Convention Center, Philadelphia, Pennsylvania, United States AAAI-2024 Vancouver Convention Centre, Vancouver, British Columbia, Canada AAAI-2023 Washington Convention Center, Washington, D.C., United States AAAI-2022 Virtual Conference AAAI-2021 Virtual Conference AAAI-2020 Hilton New York Midtown, New York, New York, United States AAAI-2019 Hilton Hawaiian Village, Honolulu, Hawaii, United States AAAI-2018 Hilton New Orleans Riverside, New Orleans, Louisiana, United States AAAI-2017 San Francisco, California, United States AAAI-2016 Phoenix, Arizona, United States AAAI-2015 Austin, Texas, United States AAAI-2014 Québec Convention Center, Québec City, Québec, Canada AAAI-2013 Bellevue, Washington, United States AAAI-2012 Toronto, Ontario, Canada AAAI-2011 San Francisco, California, United States AAAI-2010 Westin Peachtree Plaza, Atlanta, Georgia, United States AAAI-2008 Chicago, Illinois, United States AAAI-2007 Toronto, Ontario, Canada AAAI-2006 Boston, Massachusetts, United States AAAI-2005 Pittsburgh, Pennsylvania, United States AAAI-2004 San Jose, California, United States AAAI-2002 Shaw conference center in Edmonton, Alberta, Canada AAAI-2000 Austin, Texas, United States AAAI-1999 Orlando, Florida, United States AAAI-1998 Madison, Wisconsin, United States AAAI-1997 Providence, Rhode Island, United States AAAI-1996 Portland, Oregon, United States AAAI-1994 Seattle, Washington, United States AAAI-1993 Washington Convention Center, Washington, D.C., United States AAAI-1992 San Jose Convention Center, San Jose, California, United States AAAI-1991 Anaheim Convention Center, Anaheim, California, United States AAAI-1990 Boston, Massachusetts, United States AAAI-1988 Saint Paul, Minnesota, United States AAAI-1987 Seattle, Washington, United States AAAI-1986 Philadelphia, Pennsylvania, United States AAAI-1984 University of Texas, Austin, Texas, United States AAAI-1983 Washington, D.C., United States AAAI-1982 Carnegie Mellon University and the University of Pittsburgh, Pittsburgh, Pennsylvania, United States AAAI-1980 Stanford, California, United States