Yi Zeng (Chinese: 曾毅) is a Chinese artificial intelligence researcher and professor at the Chinese Academy of Sciences, who also serves as the founding director of Center for Long-term AI, and as a member of the United Nations Advisory Body on AI. == Career == On May 25, 2019, Zeng led the team that published the Beijing Artificial Intelligence Principles, proposed as an initiative for the long-term research, governance and planning of AI, and the "realization of beneficial AI for mankind and nature". He was named on the Time 100 AI list, a list featuring the hundred most influential figures in artificial intelligence of the year, in 2023. In July 2023, Zeng addressed the United Nations Security Council in a meeting on the risks posed by recent strides in artificial intelligence. He said that AI models “cannot be trusted as responsible agents that can help humans to make decisions,” and warned of the risk of extinction posed by both near-term and long-term AI, arguing that “in the long term, we haven’t given superintelligence any practical reasons why they should protect humans”. Zeng stated that humans should always be responsible for final decision-making on the use of nuclear weapons, and that the United Nations must produce an international framework on AI development and governance, to ensure global peace and security. In October 2023, UN Secretary-General António Guterres announced the creation of an advisory body on issues surrounding the international governance of AI, of which Zeng would be a member. He leads teams of researchers at the Institute of Philosophy and the Institute of Automation of the Chinese Academy of Sciences, including doctoral candidates, postdoctoral fellows, research fellows, assistant professors, and associate professors. Among them is his first international PhD student, Ammar Younas, a lawyer and arbitrator whose research focuses on cross-cultural dimensions of AI ethics and governance.
World model (artificial intelligence)
A world model in artificial intelligence is a machine learning system that builds an internal representation of an environment. The model predicts how that environment changes over time in response to actions. Researchers design world models to help agents plan, reason, and act without constant real-world trial and error. World models differ from systems that merely classify or generate outputs. They simulate dynamics such as physics, object interactions, and causality. Early ideas date to the 1990s. Modern versions power robots, autonomous driving, and interactive video generation. == History == Jürgen Schmidhuber introduced the term world model in machine learning in 1990. He proposed recurrent neural networks that predict future states from observations and use those predictions to train agents. David Ha and Schmidhuber revived the concept in a 2018 paper. Their agents learned to drive virtual cars and play video games inside self-generated simulations. Yann LeCun advanced the idea in a 2022 position paper titled "A Path Towards Autonomous Machine Intelligence". He argued that intelligence requires predictive models of the world rather than pure pattern matching. LeCun proposed the joint embedding predictive architecture (JEPA) as a practical foundation. LeCun and collaborators developed several JEPA variants. V-JEPA 2 reached state-of-the-art performance on video understanding and physical reasoning at the time. It supports zero-shot robot control in unfamiliar environments. Introduced in March 2026, LeWorldModel trains stably end-to-end from raw pixels and uses two loss terms and avoids hand-crafted heuristics. LeCun founded Advanced Machine Intelligence Labs in 2026 to further develop world models. Google DeepMind introduced Genie in 2024. The model learned interactive environments from unlabeled internet videos. Genie 2 followed in late 2024 and added three-dimensional generation. The Genie series set benchmarks for general-purpose simulation. Genie 3 was introduced in August 2025. It produces photorealistic, real-time interactive worlds from text prompts which are displayed at 24 frames per second and explored in real time with text or image prompts. The model supports persistent three-dimensional worlds and real-time interaction. Waymo adopted Genie 3 in February 2026 and used it to create a specialized world model for autonomous driving simulation, called the Waymo World Model. It produces synchronized camera and lidar outputs and creates edge cases that real robotaxis rarely encounter. The edge cases were reported to be unusual by PCMag. General Intuition announced a $133.7 million seed round. World Labs raised $1 billion. AMI raised $1.03 billion. In April 2026, Alibaba announced Happy Oyster, its world model designed for real-time and “flowy” world model. It includes a directing mode for world building based on text and image prompts and a wandering mode for exploring the resulting world. It can generate 3-minute in-world video clips. Also in April, World Labs, co-founded by Li Fei Fei, unveiled Spark 2.0, an open-source 3D Gaussian splatting rendering engine that targets smartphone-class devices. In June 2026, Nvidia released Cosmos 3, a family of open-weight models. It combines previously independent physical reasoning, world simulation, and action generation. Cosmos 3 integrates can process and generate text, image, video, audio, and action sequences. The model employs a Mixture-of-Transformers" (MoT) approach. An autoregressive (AR) transformer handles reasoning and next-token prediction, while a diffusion transformer (DT) does multimodal generation. Encoders (ViT for vision, VAE for visual/audio, and domain-specific for actions) and generate a shared representation space using 3D multi-dimensional rotary position embedding (mRoPE) for spatial and temporal information. The family includes Cosmos3-Nano (16B parameters) for workstations; Cosmos3-Super (64B parameters) for research. == Architecture == World models process raw sensory data such as video frames or lidar scans. They compress this input into compact latent representations. The system then predicts future representations rather than pixel-by-pixel reconstructions. Many modern world models use joint embedding predictive architecture (JEPA). An encoder turns observations into embeddings. A predictor estimates one or a suite of embeddings from the current one and an action. In some cases a critic chooses one embedding as the best result. A regularizer keeps embeddings well-behaved. The model trains by minimizing prediction error in embedding space. This approach avoids the high cost of generating every detail. Some architectures add explicit components. A fast reactive path handles immediate responses. A slower deliberative path performs longer-horizon planning. Video prediction accuracy or robot success rates are key metrics, but do not always predict real-world performance. Generative world models such as Genie 3 combine these with a simulator. They accept text prompts or layouts and output consistent video, lidar, or three-dimensional scenes. World models often train with self-supervised learning. They use large unlabeled datasets of video or robot interactions. Self-supervised learning can speed learning. Reinforcement learning can fine-tune a model for specific tasks. == Applications == World models support robot learning. Agents train inside simulations and transfer skills to the physical world. This reduces the need for dangerous or expensive real-world trials. Autonomous vehicles use world models to test rare events. Waymo's system simulates tornadoes or unusual pedestrian behavior. Companies train planners without putting vehicles on public roads. Interactive entertainment benefits from world models. Genie 3 lets users generate playable environments from simple descriptions. Game studios prototype levels faster. Scientific simulation gains from these models. Researchers model physical systems or biological processes at scale. Planners in logistics or urban design test strategies inside accurate digital twins. == Comparison with large language models == Both world models and large language models (LLMs) use inferencing on their inputs to make predictions. LLMs operate on textual inputs. They predict the next token in text sequences. They excel at language-oriented tasks such as translation or summarization. However, they lack understanding of physics. World models operate on sensor inputs such as pixels. They predict state changes in that data in latent space. This design supports planning and causal reasoning. LLMs generate fluent text but often fail at consistent physical predictions. Their architecture employs transformers with refinements such as mixture of experts. World models divide an inferencing task into work performed by encoders, predictors, simulators, and other pieces. They typically handle multimodal inputs such as video, lidar, radar, and audio, guided by textual prompting. LLMs power chatbots and code assistants. World models drive embodied agents that act in dynamic environments, such as autonomous driving. The two may be combined in hybrid systems. For example, a LLM handles instructions, while a world model manages low-level control. World model proponents such as LeCun claim that because LLMs are trained only on text, they have no ability to predict anything beyond text, such as real-world events. == Benchmarks == World model benchmarks test physical understanding, long-term consistency, planning, and generalization from sensor data. Meta introduced three benchmarks for V-JEPA 2. IntPhys 2 measures a model's ability to detect physics violations. It presents pairs of videos that diverge when one breaks physical rules. Humans score near 100% accuracy. V-JEPA 2 achieves little better than random chance on many conditions. Minimal Video Pairs (MVPBench) tests physical understanding through multiple-choice questions based on short video clips. It probes object interactions and causality. Something-Something tests action recognition. Epic-Kitchens-100 tests human action anticipation. DeepMind benchmark: Interactive evaluation measures consistency over minutes of interaction, memory of off-screen objects, and response to user actions or text prompts. Waymo benchmark: Output generation quality: Metrics include realism, controllability (via text prompts), and usefulness for training planners in simulated worlds. However, pixel reconstruction error rate with episodic rewards often fails. Other: Epic-Kitchens-100 (often measured with Recall@5) Ego4D 50 Salads, Breakfast, etc. Potential benchmarks: Zero-shot transfer to robots Long-horizon planning Implausible prediction rate
Group concept mapping
Group concept mapping is a structured methodology for organizing the ideas of a group on any topic of interest and representing those ideas visually in a series of interrelated maps. It is a type of integrative mixed method, combining qualitative and quantitative approaches to data collection and analysis. Group concept mapping allows for a collaborative group process with groups of any size, including a broad and diverse array of participants. Since its development in the late 1980s by William M.K. Trochim at Cornell University, it has been applied to various fields and contexts, including community and public health, social work, health care, human services,, instructional interventions, and biomedical research and evaluation. == Overview == Group concept mapping integrates qualitative group processes with multivariate analysis to help a group organize and visually represent its ideas on any topic of interest through a series of related maps. It combines the ideas of diverse participants to show what the group thinks and values in relation to the specific topic of interest. It is a type of structured conceptualization used by groups to develop a conceptual framework, often to help guide evaluation and planning efforts. Group concept mapping is participatory in nature, allowing participants to have an equal voice and to contribute through various methods. A group concept map visually represents all the ideas of a group and how they relate to each other, and depending on the scale, which ideas are more relevant, important, or feasible. == Process == Group concept mapping involves a structured multi-step process, including brainstorming, sorting and rating, multidimensional scaling and cluster analysis, and the generation and interpretation of multiple maps. The first step requires participants to brainstorm a large set of statements relevant to the topic of interest, usually in response to a focus prompt. Participants are then asked to individually sort those statements into categories based on their perceived similarity and rate each statement on one or more scales, such as importance or feasibility. The data is then analyzed using The Concept System software, which creates a series of interrelated maps using multidimensional scaling (MDS) of the sort data, hierarchical clustering of the MDS coordinates applying Ward's method, and the computation of average ratings for each statement and cluster of statements. The resulting maps display the individual statements in two-dimensional space with more similar statements located closer to each other, and grouped into clusters that partition the space on the map. The Concept System software also creates other maps that show the statements in each cluster rated on one or more scales, and absolute or relative cluster ratings between two cluster sets. As a last step in the process, participants are led through a structured interpretation session to better understand and label all the maps. == History == Group concept mapping was developed as a methodology in the late 1980s by William M.K. Trochim at Cornell University. Trochim is considered to be a leading evaluation expert, and he has taught evaluation and research methods at Cornell since 1980. Originally called "concept mapping", the methodology has evolved since its inception with the maturation of the field and the continued advancement of the software, which is now a Web application. == Uses == Group concept mapping can be used with any group for any topic of interest. It is often used by government agencies, academic institutions, national associations, not-for-profit and community-based organizations, and private businesses to help turn the ideas of the group into measurable actions. This includes in the areas of organizational development, strategic planning, needs assessment, curriculum development, research, and evaluation. Group concept mapping is well-documented, well-established methodology, and it has been used in hundreds of published papers. == Versus concept mapping and mind mapping == More generally, concept mapping is any process used for visually representing relationships between ideas in pictures or diagrams. A concept map is typically a diagram of multiple ideas, often represented as boxes or circles, linked in a graph (network) structure through arrows and words where each idea is connected to another. The technique was originally developed in the 1970s by Joseph D. Novak at Cornell University. Concept mapping may be done by an individual or a group. A mind map is a diagram used to visually represent information, centering on one word or idea with categories and sub-categories radiating off of it in a tree structure. Popularized by Tony Buzan in the 1970s, mind mapping is often a spontaneous exercise done by an individual or group to gather information about what they think around a single topic. Unlike Novak's concept maps and Buzan's mind maps, group concept mapping has a structured mathematical process (sorting and rating, multidimensional scaling and cluster analysis) for organizing and visually representing multiple ideas of a group through a series of specific steps. In other words, in group concept mapping, the resulting visual representations are mathematically generated from mixed (qualitative and quantitative) data collected from a group of research subjects, whereas in Novak's concept maps and Buzan's mind maps the visual representations are drawn directly by the subjects resulting in diagrams that are qualitative data and final product at the same time.
MuZero
MuZero is a computer program developed by artificial intelligence research company DeepMind, a subsidiary of Google, to master games without knowing their rules and underlying dynamics. Its release in 2019 included benchmarks of its performance in Go, chess, shogi, and a suite of 57 different Atari games. The algorithm uses an approach similar to AlphaZero, where a combination of a tree-based search and a learned model is deployed. It matched AlphaZero's performance in chess and shogi, improved on its performance in Go, and improved on the state of the art in mastering a suite of 57 Atari games (the Arcade Learning Environment), a visually-complex domain. MuZero was trained via self-play, with no access to rules, opening books, or endgame tablebases. The trained algorithm used the same convolutional and residual architecture as AlphaZero, but with 20 percent fewer computation steps per node in the search tree. == History == MuZero really is discovering for itself how to build a model and understand it just from first principles. On November 19, 2019, the DeepMind team released a preprint introducing MuZero. === Derivation from AlphaZero === MuZero (MZ) is a combination of the high-performance planning of the AlphaZero (AZ) algorithm with approaches to model-free reinforcement learning. The combination allows for more efficient training in classical planning regimes, such as Go, while also handling domains with much more complex inputs at each stage, such as visual video games. MuZero was derived directly from AZ code, sharing its rules for setting hyperparameters. Differences between the approaches include: AZ's planning process uses a simulator. The simulator knows the rules of the game. It has to be explicitly programmed. A neural network then predicts the policy and value of a future position. Perfect knowledge of game rules is used in modeling state transitions in the search tree, actions available at each node, and termination of a branch of the tree. MZ does not have access to the rules, and instead learns one with neural networks. AZ has a single model for the game (from board state to predictions); MZ has separate models for representation of the current state (from board state into its internal embedding), dynamics of states (how actions change representations of board states), and prediction of policy and value of a future position (given a state's representation). MZ's hidden model may be complex, and it may turn out it can host computation; exploring the details of the hidden model in a trained instance of MZ is a topic for future exploration. MZ does not expect a two-player game where winners take all. It works with standard reinforcement-learning scenarios, including single-agent environments with continuous intermediate rewards, possibly of arbitrary magnitude and with time discounting. AZ was designed for two-player games that could be won, drawn, or lost. === Comparison with R2D2 === The previous state of the art technique for learning to play the suite of Atari games was R2D2, the Recurrent Replay Distributed DQN. MuZero surpassed both R2D2's mean and median performance across the suite of games, though it did not do better in every game. == Training and results == MuZero used 16 third-generation tensor processing units (TPUs) for training, and 1000 TPUs for selfplay for board games, with 800 simulations per step and 8 TPUs for training and 32 TPUs for selfplay for Atari games, with 50 simulations per step. AlphaZero used 64 second-generation TPUs for training, and 5000 first-generation TPUs for selfplay. As TPU design has improved (third-generation chips are 2x as powerful individually as second-generation chips, with further advances in bandwidth and networking across chips in a pod), these are comparable training setups. R2D2 was trained for 5 days through 2M training steps. === Initial results === MuZero matched AlphaZero's performance in chess and shogi after roughly 1 million training steps. It matched AZ's performance in Go after 500,000 training steps and surpassed it by 1 million steps. It matched R2D2's mean and median performance across the Atari game suite after 500 thousand training steps and surpassed it by 1 million steps, though it never performed well on 6 games in the suite. == Reactions and related work == MuZero was viewed as a significant advancement over AlphaZero, and a generalizable step forward in unsupervised learning techniques. The work was seen as advancing understanding of how to compose systems from smaller components, a systems-level development more than a pure machine-learning development. While only pseudocode was released by the development team, Werner Duvaud produced an open source implementation based on that. MuZero has been used as a reference implementation in other work, for instance as a way to generate model-based behavior. In late 2021, a more efficient variant of MuZero was proposed, named EfficientZero. It "achieves 194.3 percent mean human performance and 109.0 percent median performance on the Atari 100k benchmark with only two hours of real-time game experience". In early 2022, a variant of MuZero was proposed to play stochastic games (for example 2048, backgammon), called Stochastic MuZero, which uses afterstate dynamics and chance codes to account for the stochastic nature of the environment when training the dynamics network.
Linagora
Linagora is a French open source software editor, founded in June 2000 by Alexandre Zapolsky and Michel-Marie Maudet. Located in France, as well as in Belgium, Canada, Vietnam, the United States and Tunisia, the company employs around 200 people. In 2023, Linagora created the OpenLLM France community, alongside other French Artificial Intelligence companies and organizations. In 2025, the company launched Lucie, an opensource Large Language Model. == History == Linagora was founded on June 28, 2000. Its name is a contraction of the words "Linux" and "Agora". The company was founded by Alexandre Zapolsky and Michel-Marie Maudet. Soon after, the two entrepreneurs were joined by Alexandre Zapolsky's wife and brother, who took on the roles of commercial director and administrative and financial director of the SME. In 2007, the company was selected by the French National Assembly to provide the software for Linux computers, replacing Microsoft Windows. Linagora then claimed the position of the leading French open source software company by revenue. In 2015, French Prime Minister Manuel Valls allocated €10.7 million from the "Investments for the Future" fund for a research program aimed at developing a new generation of open source software platforms based on Linagora's offerings. In September 2016, Linagora launched the social network "La Cerise" for the newspaper L'Humanité. This app offered a service and tool for readers and citizens mobilizing for causes. It aimed to share engagement through petitions, discussions, agendas, and contacts. In October 2016, the company won two public contracts for supporting open source software in forty-two French ministries and other administrative entities. In May 2019, Linagora organized a fundraising event in the presence of the French Secretary of State for Digital Affairs, Cédric O, to celebrate its 19th anniversary. The funds were intended for: Supporting parents of hospitalized Polynesian children in France. Equipping primary school students with digital devices (tablets or PCs). Establishing a digital academy "OpenHackademy" in French Polynesia to train unemployed youth in digital skills and help them find jobs. In December 2022, Linagora acquired a property known as "Maison Rocher" and later "Maison Chocolat," located on the Île Saint-Germain in Issy-les-Moulineaux. Renamed "Villa Good Tech" by Linagora, this award-winning architectural work by Éric Daniel-Lacombe became the company's new headquarters, aiming to provide a space for associative actors and companies to develop technologies that contribute to a better world. In July 2023, Linagora launched OpenLLM France, a community initially comprising around twenty actors focused on generative AI. The goal was to develop a sovereign and open source large language model. This initiative, led by co-founder and CEO Michel-Marie Maudet, had more than four hundred French members by early 2024. and announced its expansion to the European sphere during Fosdem 2024. In February 2024, the CNRS and Linagora signed a framework agreement to strengthen their research collaboration. In January 2025, Linagora released Lucie, an open source and sovereign AI that faced ridicule due to tests on an unfinished, uncensored version designed for scientific and experimental use. The platform divided opinions between those who saw it as a technological achievement and those who criticized it as "French bashing" compared to American and Chinese AIs. == Acquisitions == The company acquired: In July 2007, the SME AliaSource, based in Ramonville-Saint-Agne and led by its founder, Pierre Baudracco. In 2008, the open source web hosting company Netaktiv, a member of the GIE Gitoyen, announced during the 2008 Solutions Linux trade show. In 2012, the Toulouse-based company EBM Websourcing, the publisher of the open-source software Petals Link, and took over its development. In 2016, the digital agency Neoma Interactive, specializing in UX design and digital communication strategy. == Locations == In 2017, the company's headquarters was located in Issy-les-Moulineaux, with branches in Lyon, Toulouse, Marseille, and internationally in Brussels, San Francisco, Montreal, Vietnam, and Tunisia. In 2005, the company attempted to establish a presence in Nantes. In 2024, the headquarters was moved to Issy-les-Moulineaux. == Activity == === Software === Twake Workplace One of Linagora's flagship products is Twake Workplace, which stands out as a 100% open-source solution compared with those of the GAFAMs. Twake Workplace is available as a complete platform or module by module. It includes : Twake Mail, a powerful modern messaging solution based on the JMAP protocol and the James email server from the Apache Foundation, for which Linagora provides technical management; Twake Chat, an instant communications solution for businesses developed using the Matrix protocol and compatible with the French government's chat solution, Tchap; Twake Drive, an easy-to-use collaborative platform for group work using OnlyOffice. ==== OpenPaaS ==== In 2018, the search engine Qwant announced that its email service Qwantmail would be based on the OpenPaaS product. In 2022, Qwant announced the abandonment of its Qwantmail project due to Linagora's collection of personal email addresses and serious security breaches. The site Next (formerly PC INpact) published an article in January 2020 criticizing the "failures and delays" of the Qwantmail project led by Linagora, which led to the CNIL's intervention regarding Qwant and Linagora. ==== LinTO ==== In 2017, Linagora launched its open source voice assistant project named LinTO. This enterprise voice assistant, described as "GAFAM Free," was presented at CES 2018 in Las Vegas. The LinTO voice framework was developed as part of the eponymous research project funded by Bpifrance (Grands Défis du Numérique instrument). === Services === ==== OSSA (Open Source Software Assurance) ==== One of the company's main activities is OSSA. Through OSSA, Linagora provided support for open source software for 42 ministries and other administrative entities in 2012. == Legal issues == === Dispute with BlueMind === In 2012, a legal dispute arose between BlueMind and Linagora. Linagora accused BlueMind of copyright infringement, unfair competition, and breach of a non-compete clause, leading to several legal actions. Linagora sued BlueMind for copyright infringement and unfair competition in the Bordeaux court, which ruled in Linagora's favor for unfair competition and parasitism but rejected the copyright claim. BlueMind was ordered to pay nearly €170,000 to Linagora. Linagora sued former associates Pierre Baudracco and Pierre Carlier in the Paris Commercial Court for breach of a non-compete clause and violation of a warranty of eviction. The court dismissed Linagora's claims and ordered it to pay €20,000 each to Baudracco and Carlier. Linagora appealed, and the Paris Court of Appeal partially overturned the decision, awarding Linagora €480,000. BlueMind sued Linagora for defamation and public insult in the Toulouse Criminal Court. The court ruled against Linagora, but the decision was overturned by the Court of Cassation in January 2024, and the case was remanded for retrial. === Conviction for wrongful termination and harassment === On June 14, 2017, France 3 reported on a decision by the Versailles Court of Appeal, which ruled that Linagora had wrongfully terminated an employee and subjected them to moral harassment. The court ordered Linagora to pay the employee €22,000 for wrongful termination, €11,000 for notice pay, €6,600 for legal severance pay, €3,200 for conservative suspension, and €3,000 for moral harassment.
Morphological antialiasing
Morphological antialiasing (MLAA) is a spatial anti-aliasing technique used in real-time computer graphics. It reduces artifacts, such as jaggies, when representing a high-resolution image at a lower resolution. MLAA is a post-process filtering which detects borders in the resulting image and then finds specific patterns in these. Anti-aliasing is achieved by blending pixels in these borders, according to the pattern they belong to and their position within the pattern. Introduced in 2009, MLAA was an early and influential example of anti-aliasing techniques done in post-processing, which makes them suitable for deferred shading. A similar method in this class is fast approximate anti-aliasing (FXAA). Temporal anti-aliasing, also a post-process, has become the most common anti-aliasing method for real-time rendering and video games. Enhanced subpixel morphological antialiasing, or SMAA, is an image-based GPU-based implementation of MLAA developed by Universidad de Zaragoza and Crytek.
Comet (browser)
Comet is an AI-powered web browser based on Chromium. It was released by Perplexity AI for Microsoft Windows and macOS on July 9, 2025, for Android on November 20, 2025, and for iOS on March 18, 2026. Initial access to the browser was limited to users subscribed to Perplexity's most expensive tier, with broader availability expected over time. The browser was released for free download in October 2025. == Features == Comet is integrated with Perplexity's AI-assisted search engine. The browser features an assistant which enables users to perform a variety of tasks such as generating article summaries, sending emails, or buying products. == Security concerns == Researchers at LayerX Security identified a malicious attack vector which they call CometJacking. The exploit could possibly exfiltrate a user's personal sensitive data to a remote server controlled by the attacker. LayerX attempted to responsibly disclose their findings to Comet's developer Perplexity AI in August 2025. Perplexity responded that they saw no security impact and marked the disclosure as not applicable.