AI Detector Text Free

AI Detector Text Free — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

ISO/IEC JTC 1/SC 24

ISO/IEC JTC 1/SC 24 Computer graphics, image processing and environmental data representation is a standardization subcommittee of the joint subcommittee ISO/IEC JTC 1 of the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), which develops and facilitates standards within the field of computer graphics, image processing, and environmental data representation. The international secretariat of ISO/IEC JTC 1/SC 24 is the British Standards Institute (BSI) located in the United Kingdom. == History == ISO/IEC JTC 1/SC 24 was formed in 1987 from ISO/TC 97 as a result of Resolution 21 at the ISO/IEC JTC 1 plenary. The group's origins began in computer graphics, the standardization of which was originally under ISO/IEC JTC 1/SC 21/WG 2. However, when ISO/IEC JTC 1/SC 24 was created, the standardization activity of ISO/IEC JTC 1/SC 21/WG 2 was carried over to the new subcommittee. The initial five working groups of ISO/IEC JTC 1/SC 24 were titled, “Architecture,” “Application programming interfaces,” “Metafiles and interfaces,” “Language bindings,” and “Validation, testing and registration.” The work of ISO/IEC JTC 1/SC 24 began with the Graphical Kernel System (GKS), which was adopted from ISO/IEC JTC 1/SC 21/WG 2. However, since GKS only addressed 2D functionality, attention turned to the standardization of 3D functionality. This resulted in two standards being published: GKS-3D in 1988 and PHIGS in 1989, both of which addressed 3D functionality. Since 1991, ISO/IEC JTC 1/SC 24 has held plenaries in a number of countries, including the Netherlands, Germany, United States, France, Canada, Japan, Sweden, Korea, United Kingdom, Australia, and Czech Republic. == Scope == The scope of ISO/IEC JTC 1/SC 24 is the “Standardization of interfaces for information technology based applications relating to”: Computer graphics Image processing Environmental data representation Support for the Mixed and Augmented Reality (MAR) Interaction with, and visual representation of, information Included are the following related areas: Modeling and simulation and related reference models Virtual reality with accompanying augmented reality/augmented virtuality aspects and related reference models Application program interfaces Functional specifications Representation models Interchange formats, encodings and their specifications, including metafiles Device interfaces Testing methods Registration procedures Presentation and support for creation of multimedia, hypermedia, and mixed reality documents Excluded are the following areas: Character and image coding Coding of multimedia, hypermedia, and mixed reality document interchange formats JTC 1 work in user system interfaces and document presentation ISO/TC 207 work on ISO 14000 environment management, ISO/TC 211 work on geographic information and geomatics Software environments as described by ISO/IEC JTC 1/SC 22 == Structure == ISO/IEC JTC 1/SC 24 is made up of four active working groups, each of which carries out specific tasks in standards development within the field of computer graphics, image processing and environmental data representation, together with ITU-T Study Group 16. As a response to changing standardization needs, working groups of ISO/IEC JTC 1/SC 24 can be disbanded if their area of work is no longer applicable, or established if new working areas arise. The focus of each working group is described in the group's terms of reference. Active working groups of ISO/IEC JTC 1/SC 24 are: == Collaborations == ISO/IEC JTC 1/SC 24 works in close collaboration with a number of other organizations or subcommittees, both internal and external to ISO or IEC, in order to avoid conflicting or duplicative work. Organizations internal to ISO or IEC that collaborate with or are in liaison to ISO/IEC JTC 1/SC 24 include: ISO/IEC JTC 1/WG 7, Sensor Networks ISO/IEC JTC 1/SC 29, Coding of audio, picture, multimedia and hypermedia information ISO/IEC JTC 1/SC 32, Data management and interchange ISO/TAG 14, Imagery and technology ISO/TC 130, Graphic Technology ISO/TC 184/SC 4, Industrial data ISO/TC 211, Geographic information/Geomatics ISO/TC 215, Health informatics IEC TC 100, Audio, video and multimedia system and equipment Some organizations external to ISO or IEC that collaborate with or are in liaison to ISO/IEC JTC 1/SC 24 include: Defence Geospatial Information Working Group (DGIWG) Digital Imaging and Communications in Medicine (DICOM) International Hydrographic Organization (IHO) The Khronos Group NATO - Joint Intelligence Surveillance and Reconnaissance Capability Group (JISRCG) OMG Robotics DTF Open CGM Open Geospatial Consortium (OGC) SEDRIS Organization Simulation Interoperability Standards Organization (SISO) US National Imagery Transmission Format Standard (NITFS) Technical Board (US NTB) Web3D Consortium World Intellectual Property Organization (WIPO) World Wide Web Consortium (W3C) == Member countries == Countries pay a fee to ISO to be members of subcommittees. The 11 "P" (participating) members of ISO/IEC JTC 1/SC 24 are: Australia, China, Egypt, France, India, Japan, Republic of Korea, Portugal, Russian Federation, United Kingdom, and United States. The 22 "O" (observer) members of ISO/IEC JTC 1/SC 24 are: Argentina, Austria, Belgium, Bosnia and Herzegovina, Bulgaria, Canada, Cuba, Czech Republic, Finland, Ghana, Hungary, Iceland, Indonesia, Islamic Republic of Iran, Italy, Kazakhstan, Malaysia, Poland, Romania, Serbia, Slovakia, Switzerland, and Thailand. == Published standards == ISO/IEC JTC 1/SC 24 currently has 80 published standards under their direct responsibility within the field of computer graphics, image processing, and environmental data representation, including:
Read more →
Regulation of artificial intelligence in the United States

The United States federal government and state governments have developed some regulation of artificial intelligence, including executive orders, federal laws, and state laws. Federal agencies have also developed some sector-specific regulations related to AI. At the federal level, the Biden administration released an October 2023 executive order about AI safety and security, Executive Order 14110, with directives related to AI development and deployment. President Trump revoked that executive order in January 2025 and issued Executive Order 14179. In December 2025, President Trump signed Executive Order 14365, an executive order directing federal agencies to develop a unified national approach to AI policy, evaluate state AI laws for potential conflicts, challenge them through legal action, and condition certain federal funding on state compliance, while exempting state laws related to child safety, data center infrastructure, and state government procurement. In 2025, Congress passed legislation targeting AI-generated deepfakes, the TAKE IT DOWN Act. Several U.S. states have enacted laws related to artificial intelligence. Some are already in effect, including in California. Other states have AI-related legislation coming into effect in 2026 and 2027. In 2025 and 2026, the Trump administration mentioned the patchwork nature of state legislation as a motivation for its push for unified national legislation regulating AI. The administration has criticized state lawmakers, threatened to sue states, and issued letters to discourage them from regulating AI companies and products; some states have continued to propose and enact related laws. Discussions about regulating AI have included topics such as the timeliness of regulating AI, the nature of the federal regulatory framework to govern and promote AI, including what agency should lead, the regulatory and governing powers of that agency, and how to update regulations in the face of rapidly changing technology, as well as the roles of state governments and courts. == Federal government == === Obama administration (2009–2017) === As early as 2016, the Obama administration had begun to focus on the risks and regulations for artificial intelligence. In an October 2016 report titled Preparing For the Future of Artificial Intelligence, the National Science and Technology Council set a precedent to allow researchers to continue to develop new AI technologies with few restrictions. The report stated that "the approach to regulation of AI-enabled products to protect public safety should be informed by assessment of the aspects of risk". The first National Artificial Intelligence Research And Development Strategic Plan was published in October 2016. === First Trump administration (2017–2021) === On August 13, 2018, Section 1051 of the Fiscal Year 2019 John S. McCain National Defense Authorization Act (P.L. 115-232) established the National Security Commission on Artificial Intelligence "to consider the methods and means necessary to advance the development of artificial intelligence, machine learning, and associated technologies to comprehensively address the national security and defense needs of the United States." Steering on regulating security-related AI is provided by the National Security Commission on Artificial Intelligence. The Artificial Intelligence Initiative Act (S.1558) is a proposed bill that would establish a federal initiative designed to accelerate research and development on AI for, inter alia, the economic and national security of the United States. On January 7, 2019, following an Executive Order on Maintaining American Leadership in Artificial Intelligence, the White House's Office of Science and Technology Policy released a draft Guidance for Regulation of Artificial Intelligence Applications, which includes ten principles for United States agencies when deciding whether and how to regulate AI. In response, the National Institute of Standards and Technology released a position paper, and the Defense Innovation Board issued recommendations on the ethical use of AI. A year later, the administration called for comments on regulation in another draft of its Guidance for Regulation of Artificial Intelligence Applications. Other specific agencies working on the regulation of AI included the Food and Drug Administration, which created pathways to regulate the incorporation of AI in medical imaging. The National Science and Technology Council also published an updated National Artificial Intelligence Research and Development Strategic Plan in 2019, which received public scrutiny and recommendations to further improve it towards enabling Trustworthy AI. === Biden administration (2021–2025) === In March 2021, the National Security Commission on Artificial Intelligence released their final report. In the report, they stated, "Advances in AI, including the mastery of more general AI capabilities along one or more dimensions, will likely provide new capabilities and applications. Some of these advances could lead to inflection points or leaps in capabilities. Such advances may also introduce new concerns and risks and the need for new policies, recommendations, and technical advances to assure that systems are aligned with goals and values, including safety, robustness and trustworthiness." In June 2022, Senators Rob Portman and Gary Peters introduced the Global Catastrophic Risk Management Act. The bipartisan bill "would also help counter the risk of artificial intelligence... from being abused in ways that may pose a catastrophic risk". On October 4, 2022, President Joe Biden unveiled a new AI Bill of Rights, which outlines five protections Americans should have in the AI age: 1. Safe and Effective Systems, 2. Algorithmic Discrimination Protection, 3.Data Privacy, 4. Notice and Explanation, and 5. Human Alternatives, Consideration, and Fallback. The bill was formally published in October 2022 by the Office of Science and Technology Policy (OSTP), a U.S. government office that advises the President on science and technology policy matters. In July 2023, the Biden administration secured voluntary commitments from seven companies – Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI – to manage the risks associated with AI. The companies committed to ensure AI products undergo both internal and external security testing before public release; to share information on the management of AI risks with the industry, governments, civil society, and academia; to prioritize cybersecurity and protect proprietary AI system components; to develop mechanisms to inform users when content is AI-generated, such as watermarking; to publicly report on their AI systems' capabilities, limitations, and areas of use; to prioritize research on societal risks posed by AI, including bias, discrimination, and privacy concerns; and to develop AI systems to address societal challenges, ranging from cancer prevention to climate change mitigation. In September 2023, eight additional companies – Adobe, Cohere, IBM, Nvidia, Palantir, Salesforce, Scale AI, and Stability AI – subscribed to these voluntary commitments. In January 2023, the National Institute of Standards and Technology (NIST) released the Artificial Intelligence Risk Management Framework (AI RMF 1.0), providing voluntary guidance for organizations to identify, assess, and manage risks associated with AI systems. The Biden administration, in October 2023 signaled that they would release an executive order leveraging the federal government's purchasing power to shape AI regulations, hinting at a proactive governmental stance in regulating AI technologies. On October 30, 2023, President Biden released Executive Order 14110 on Safe, Secure, and Trustworthy Artificial Intelligence. The Executive Order includes directives on standards for critical infrastructure, AI-enhanced cybersecurity, and federally funded biological synthesis projects. The Executive Order provides the authority to various agencies and departments of the US government, including the Energy and Defense departments, to apply existing consumer protection laws to AI development. The Executive Order builds on the Administration's earlier agreements with AI companies to instate new initiatives to "red-team" or stress-test AI dual-use foundation models, especially those that have the potential to pose security risks, with data and results shared with the federal government. The Executive Order also recognizes AI's social challenges, and calls for companies building AI dual-use foundation models to be wary of these societal problems. For example, the Executive Order states that AI should not "worsen job quality", and should not "cause labor-force disruptions". Additionally, Biden's Executive Order mandates that AI must "advance equity and civil rights", and cannot disadvantage marginalized groups. It also called for foundation models to include "watermarks" to help the publi
Read more →
Xinhua–Sogou AI news anchor

Xinhua News Agency and Sogou of China developed an artificial intelligence (AI) for news reporting purposes. The AI was unveiled in 2018. It is touted to be the "world's first AI news anchor". == History == The AI was unveiled at the 2018 World Internet Conference in Wuzhen, Zhejiang, China. The AI devises avatars patterned after real life Xinhua anchors. The AI patterned after Qiu Hao spoke in Chinese, while the one derived from the likeness of Zhang Zhao speaks in English. The unveiling of the AI raised concerns of its impact on employment. Xinhua and Sogou unveiled Xin Xiaomeng, an AI with a female avatar in 2019. People's Daily followed suit by unveiling its own AI newscaster in 2023.
Read more →
GuideGeek

GuideGeek is an AI-powered travel assistant that was launched by travel publisher Matador Network in April 2023 and is accessed by users through Instagram, WhatsApp and Facebook Messenger to plan itineraries or provide travel tips and recommendations. It uses generative artificial intelligence technology from OpenAI. Matador Network is a San Francisco-based digital media company and online travel publication with millions of monthly visitors and social media followers. == Features == Users message GuideGeek questions about travel and receive customized answers and itineraries that are pulled from ChatGPT in addition to over 1,000 additional travel-specific integrations such as live flight, hotel and vacation rental data. Travelers can specify their budget and needs to generate custom itineraries. GuideGeek is not an app and does not require the user to download anything, instead relying on messaging apps such as Instagram to connect users with the AI. GuideGeek is free to use, doesn't include ads, and doesn't sell user data. Matador Network has a team of staff members monitoring conversations to correct them if the AI makes a false statement; for example, one user incorrectly inputted “Crete Freeze” instead of “Crete, Greece”, and the AI made up a fictional soft serve company. Using a technique known as reinforcement learning from human feedback (RLHF), the accuracy of GuideGeek increased to 98%, according to Matador Network CEO, Ross Borden. == Destination partnerships == Matador Network is monetizing GuideGeek via white-label partnerships with tourism bureaus and destination marketing organizations (DMOs). As of March 2024, it had over a dozen such clients. Estes Park, Colorado, was one of the first DMOs to partner with Matador for a custom version of GuideGeek called “Rocky Mountain Roamer.” For Discover Greece, Matador created Pythia, a custom AI named after the high priestess of the Temple of Apollo at Delphi. As Borden explained to Travel + Leisure, “Visitors to the Discover Greece website will find Pythia in the bottom right corner, and they can converse with the AI like a friend who knows everything about Greece.” Other DMOs who have partnerships with GuideGeek include the Aruba Tourism Authority, Visit Reno Tahoe, Illinois Office of Tourism, and Tourism Richmond. == Awards == In recognition of GuideGeek, Fast Company named Matador Network to its 2024 list of Most Innovative Companies. Following growth driven by the launch of GuideGeek, Matador Network was ranked on the 2024 Inc. 5000 list of fastest-growing private companies in America. The 2024 Skift IDEA Awards recognized Matador Network as a finalist in the category of Best Use of AI for GuideGeek's customized AI for the travel industry. == Michael Motamedi experiment == Travel influencer and chef Michael Motamedi traveled the world with his wife Vanessa Salas and their 2-year-old daughter on a six-month trip (which was later extended to a full year) led by GuideGeek. The family started off in Morocco before heading to Spain and continuing east. The experiment became the basis of a web series called “No Fixed Address.” Motamedi used GuideGeek's AI to select countries the family visited, where they ate, and what sites they saw. Motamedi and Salas first tested out the technology in April 2023 while using the chatbot to plan a date night in Mexico City. GuideGeek provided speakeasy and drink recommendations as well as local history facts.
Read more →
Neural operators

Neural operators are a class of deep learning architectures designed to learn maps between infinite-dimensional function spaces. Neural operators represent an extension of traditional artificial neural networks, marking a departure from the typical focus on learning mappings between finite-dimensional Euclidean spaces or finite sets. Neural operators directly learn operators between function spaces; they can receive input functions, and the output function can be evaluated at any discretization. The primary application of neural operators is in learning surrogate maps for the solution operators of partial differential equations (PDEs), which are critical tools in modeling the natural environment. Standard PDE solvers can be time-consuming and computationally intensive, especially for complex systems. Neural operators have demonstrated improved performance in solving PDEs compared to existing machine learning methodologies while being significantly faster than numerical solvers. Neural operators have also been applied to various scientific and engineering disciplines such as turbulent flow modeling, computational mechanics, graph-structured data, and the geosciences. In particular, they have been applied to learning stress-strain fields in materials, classifying complex data like spatial transcriptomics, predicting multiphase flow in porous media, and carbon dioxide migration simulations. Finally, the operator learning paradigm allows learning maps between function spaces, and is different from parallel ideas of learning maps from finite-dimensional spaces to function spaces, and subsumes these settings as special cases when limited to a fixed input resolution. == Operator learning == Understanding and mapping relationships between function spaces has many applications in engineering and the sciences. In particular, one can cast the problem of solving partial differential equations as identifying a map between function spaces, such as from an initial condition to a time-evolved state. In other PDEs this map takes an input coefficient function and outputs a solution function. Operator learning is a machine learning paradigm to learn solution operators mapping the input function to the output function . Using traditional machine learning methods, addressing this problem would involve discretizing the infinite-dimensional input and output function spaces into finite-dimensional grids and applying standard learning models, such as neural networks. This approach reduces the operator learning to finite-dimensional function learning and has some limitations, such as generalizing to discretizations beyond the grid used in training. The primary properties of neural operators that differentiate them from traditional neural networks is discretization invariance and discretization convergence. Unlike conventional neural networks, which are fixed on the discretization of training data, neural operators can adapt to various discretizations without re-training. This property improves the robustness and applicability of neural operators in different scenarios, providing consistent performance across different resolutions and grids. == Definition and formulation == Architecturally, neural operators are similar to feed-forward neural networks in the sense that they are composed of alternating linear maps and non-linearities. Since neural operators act on and output functions, neural operators have been instead formulated as a sequence of alternating linear integral operators on function spaces and point-wise non-linearities. Using an analogous architecture to finite-dimensional neural networks, similar universal approximation theorems have been proven for neural operators. In particular, it has been shown that neural operators can approximate any continuous operator on a compact set. Neural operators seek to approximate some operator G : A → U {\displaystyle {\mathcal {G}}:{\mathcal {A}}\to {\mathcal {U}}} between function spaces A {\displaystyle {\mathcal {A}}} and U {\displaystyle {\mathcal {U}}} by building a parametric map G ϕ : A → U {\displaystyle {\mathcal {G}}_{\phi }:{\mathcal {A}}\to {\mathcal {U}}} . Such parametric maps G ϕ {\displaystyle {\mathcal {G}}_{\phi }} can generally be defined in the form G ϕ := Q ∘ σ ( W T + K T + b T ) ∘ ⋯ ∘ σ ( W 1 + K 1 + b 1 ) ∘ P , {\displaystyle {\mathcal {G}}_{\phi }:={\mathcal {Q}}\circ \sigma (W_{T}+{\mathcal {K}}_{T}+b_{T})\circ \cdots \circ \sigma (W_{1}+{\mathcal {K}}_{1}+b_{1})\circ {\mathcal {P}},} where P , Q {\displaystyle {\mathcal {P}},{\mathcal {Q}}} are the lifting (lifting the codomain of the input function to a higher dimensional space) and projection (projecting the codomain of the intermediate function to the output dimension) operators, respectively. These operators act pointwise on functions and are typically parametrized as multilayer perceptrons. σ {\displaystyle \sigma } is a pointwise nonlinearity, such as a rectified linear unit (ReLU), or a Gaussian error linear unit (GeLU). Each layer t = 1 , … , T {\displaystyle t=1,\dots ,T} has a respective local operator W t {\displaystyle W_{t}} (usually parameterized by a pointwise neural network), a kernel integral operator K t {\displaystyle {\mathcal {K}}_{t}} , and a bias function b t {\displaystyle b_{t}} . Given some intermediate functional representation v t {\displaystyle v_{t}} with domain D {\displaystyle D} in the t {\displaystyle t} -th hidden layer, a kernel integral operator K ϕ {\displaystyle {\mathcal {K}}_{\phi }} is defined as ( K ϕ v t ) ( x ) := ∫ D κ ϕ ( x , y , v t ( x ) , v t ( y ) ) v t ( y ) d y , {\displaystyle ({\mathcal {K}}_{\phi }v_{t})(x):=\int _{D}\kappa _{\phi }(x,y,v_{t}(x),v_{t}(y))v_{t}(y)dy,} where the kernel κ ϕ {\displaystyle \kappa _{\phi }} is a learnable implicit neural network, parametrized by ϕ {\displaystyle \phi } . In practice, one is often given the input function to the neural operator at a specific resolution. For instance, consider the setting where one is given the evaluation of v t {\displaystyle v_{t}} at n {\displaystyle n} points { y j } j n {\displaystyle \{y_{j}\}_{j}^{n}} . Borrowing from Nyström integral approximation methods such as Riemann sum integration and Gaussian quadrature, the above integral operation can be computed as follows: ∫ D κ ϕ ( x , y , v t ( x ) , v t ( y ) ) v t ( y ) d y ≈ ∑ j n κ ϕ ( x , y j , v t ( x ) , v t ( y j ) ) v t ( y j ) Δ y j , {\displaystyle \int _{D}\kappa _{\phi }(x,y,v_{t}(x),v_{t}(y))v_{t}(y)dy\approx \sum _{j}^{n}\kappa _{\phi }(x,y_{j},v_{t}(x),v_{t}(y_{j}))v_{t}(y_{j})\Delta _{y_{j}},} where Δ y j {\displaystyle \Delta _{y_{j}}} is the sub-area volume or quadrature weight associated to the point y j {\displaystyle y_{j}} . Thus, a simplified layer can be computed as v t + 1 ( x ) ≈ σ ( ∑ j n κ ϕ ( x , y j , v t ( x ) , v t ( y j ) ) v t ( y j ) Δ y j + W t ( v t ( y j ) ) + b t ( x ) ) . {\displaystyle v_{t+1}(x)\approx \sigma \left(\sum _{j}^{n}\kappa _{\phi }(x,y_{j},v_{t}(x),v_{t}(y_{j}))v_{t}(y_{j})\Delta _{y_{j}}+W_{t}(v_{t}(y_{j}))+b_{t}(x)\right).} The above approximation, along with parametrizing κ ϕ {\displaystyle \kappa _{\phi }} as an implicit neural network, results in the graph neural operator (GNO). There have been various parameterizations of neural operators for different applications. These typically differ in their parameterization of κ {\displaystyle \kappa } . The most popular instantiation is the Fourier neural operator (FNO). FNO takes κ ϕ ( x , y , v t ( x ) , v t ( y ) ) := κ ϕ ( x − y ) {\displaystyle \kappa _{\phi }(x,y,v_{t}(x),v_{t}(y)):=\kappa _{\phi }(x-y)} and by applying the convolution theorem, arrives at the following parameterization of the kernel integral operator: ( K ϕ v t ) ( x ) = F − 1 ( R ϕ ⋅ ( F v t ) ) ( x ) , {\displaystyle ({\mathcal {K}}_{\phi }v_{t})(x)={\mathcal {F}}^{-1}(R_{\phi }\cdot ({\mathcal {F}}v_{t}))(x),} where F {\displaystyle {\mathcal {F}}} represents the Fourier transform and R ϕ {\displaystyle R_{\phi }} represents the Fourier transform of some periodic function κ ϕ {\displaystyle \kappa _{\phi }} . That is, FNO parameterizes the kernel integration directly in Fourier space, using a prescribed number of Fourier modes. When the grid at which the input function is presented is uniform, the Fourier transform can be approximated using the discrete Fourier transform (DFT) with frequencies below some specified threshold. The discrete Fourier transform can be computed using a fast Fourier transform (FFT) implementation. == Training == Training neural operators is similar to the training process for a traditional neural network. Neural operators are typically trained in some Lp norm or Sobolev norm. In particular, for a dataset { ( a i , u i ) } i = 1 N {\displaystyle \{(a_{i},u_{i})\}_{i=1}^{N}} of size N {\displaystyle N} , neural operators minimize (a discretization of) L U ( { ( a i , u i ) } i = 1 N ) := ∑ i = 1 N ‖ u i − G θ ( a i ) ‖ U 2 {\displaystyle {\mathcal {L}}_{\mathca
Read more →
Oracle Database

Oracle AI Database (commonly referred to as Oracle Database, Oracle DBMS, Oracle Autonomous Database, or simply as Oracle) is a proprietary multi-model database management system produced and marketed by Oracle Corporation. It is a database commonly used for running online transaction processing (OLTP), data warehousing (DW) and mixed (OLTP & DW) database workloads. Oracle AI Database uses SQL for database updating and retrieval. Oracle Database runs on-premises, on Oracle engineered systems such as Oracle Exadata, on Oracle Cloud Infrastructure, and as a managed Autonomous Database service. It is also offered inside Microsoft Azure, Google Cloud, and Amazon Web Services data centers through Oracle's multicloud offerings. The current long-term support release, Oracle AI Database 26ai, became available in the cloud and on Oracle engineered systems in October 2025 and on-premises for Linux x86-64 in January 2026. == History == Larry Ellison and his two friends and former co-workers, Bob Miner and Ed Oates, started a consultancy called Software Development Laboratories (SDL) in 1977, later Oracle Corporation. SDL developed the original version of the Oracle software. The name Oracle comes from the code-name of a Central Intelligence Agency-funded project Ellison had worked on while formerly employed by Ampex; the CIA was Oracle's first customer, and allowed the company to use the code name for the new product. Ellison wanted his database to be compatible with IBM System R, but that company's Don Chamberlin declined to release its error codes. By 1985 Oracle advertised, however, that "Programs written for SQL/DS or DB2 will run unmodified" on the many non-IBM mainframes, minicomputers, and microcomputers its database supported "Because all versions of ORACLE are identical". Later releases introduced capabilities associated with successive eras of the product, including PL/SQL stored procedures and triggers in Oracle7 (1992), Real Application Clusters in Oracle9i (2001), grid infrastructure and automatic management in Oracle 10g (2003), the multitenant architecture and In-Memory Column Store in Oracle Database 12c (2013), and AI Vector Search and JSON Relational Duality in Oracle Database 23ai (2024). In October 2025 Oracle rebranded the 23ai line as Oracle AI Database 26ai. (see Release History) == Architecture == An Oracle Database system consists of an instance and a database. The instance is a set of memory structures and background processes; the database is the set of files that store data. The instance exists only in memory, and a single instance is associated with one multitenant container database. The principal memory structures are the System Global Area, which is shared, and the Program Global Areas, which are private to individual processes. The shared pool, database buffer cache, and redo log buffer are components of the System Global Area, and the optional In-Memory Column Store also resides there. Background processes operate on the database files and use these memory structures; they include the database writer, the log writer, the checkpoint process, and the system and process monitor processes. Server processes handle connections from client programs and run their SQL statements. Storage is organized logically and physically. Logically, data is held in tablespaces composed of segments, extents, and data blocks. Physically, the database comprises datafiles, control files, and online redo log files, with archived redo logs supporting media recovery. == High Availability and Scalability == Oracle Database includes several technologies for high availability, disaster recovery, and scale. Oracle Real Application Clusters allows multiple instances on separate servers to access one shared database concurrently; it was introduced with Oracle9i in 2001. Oracle Data Guard maintains standby databases synchronized with a primary database, and Active Data Guard additionally allows read-only workloads on a standby while it applies changes. Oracle GoldenGate performs logical replication and data integration across heterogeneous systems. Native sharding, introduced in Oracle Database 12c Release 2, distributes one logical database across independent shards. Oracle Exadata is an engineered system that pairs database servers with storage servers and offloads operations such as filtering to the storage tier; it is available on-premises, in Oracle Cloud Infrastructure, and through Cloud@Customer. == Notable Features == AI Vector Search adds a vector data type, vector indexes, and vector distance operators to the database. These allow similarity search over machine-learning embeddings to be expressed in SQL and combined with queries over relational, JSON, spatial, and graph data. It became generally available in Oracle Database 23ai. JSON Relational Duality exposes the same data both as relational tables and as JSON documents through duality views, so that an application can read and write either representation of the data. It became generally available in Oracle Database 23ai. In-Memory Column Store maintains a column-oriented copy of selected tables in memory in addition to the row-oriented format, and the optimizer can use the columnar copy for analytic queries. It was introduced in Oracle Database 12c Release 1.Partitioning divides large tables and indexes into independently managed pieces. Advanced Compression and Hybrid Columnar Compression are compression features for transactional and warehouse data respectively. == Data Types == Oracle AI Database supports a variety of data types and data models within a single system. These include traditional relational data types as well as semi-structured, unstructured, and specialized data formats, enabling different types of data to be stored and queried together. == Releases and versions == Oracle products follow a custom release-numbering and -naming convention. The "ai" in the current release, Oracle AI Database 26ai, stands for "Artificial Intelligence". Previous releases (e.g. Oracle Database 19c, 10g, and Oracle9i Database) have used suffixes of "c", "g", and "i" which stand for "Cloud", "Grid", and "Internet" respectively. Prior to the release of Oracle8i Database, no suffixes featured in Oracle AI Database naming conventions. There was no v1 of Oracle AI Database, as Ellison "knew no one would want to buy version 1". For some database releases, Oracle also provides an Express Edition (XE) that is free to use. Oracle AI Database release numbering has used the following codes: The Introduction to Oracle AI Database includes a brief history on some of the key innovations introduced with each major release of Oracle AI Database. See My Oracle Support (MOS) note Release Schedule of Current Database Releases (Doc ID 742060.1) for the current Oracle AI Database releases and their patching end dates. == Patch updates and security alerts == Prior to Oracle Database 18c, Oracle Corporation released Critical Patch Updates (CPUs) and Security Patch Updates (SPUs) and Security Alerts to close security vulnerabilities. These releases are issued quarterly; some of these releases have updates issued prior to the next quarterly release. Starting with Oracle Database 18c, Oracle Corporation releases Release Updates (RUs) and Release Update Revisions (RURs). RUs usually contain security, regression (bug), optimizer, and functional fixes which may include feature extensions as well. RURs include all fixes from their corresponding RU but only add new security and regression fixes. However, no new optimizer or functional fixes are included. == Competition == In the market for relational databases, Oracle AI Database competes against commercial products such as IBM Db2 and Microsoft SQL Server. Oracle and IBM tend to battle for the mid-range database market on Unix and Linux platforms, while Microsoft dominates the mid-range database market on Microsoft Windows platforms. However, since they share many of the same customers, Oracle and IBM tend to support each other's products in many middleware and application categories (for example: WebSphere, PeopleSoft, and Siebel Systems CRM), and IBM's hardware divisions work closely with Oracle on performance-optimizing server-technologies (for example, Linux on IBM Z). Niche commercial competitors include Teradata (in data warehousing and business intelligence), Software AG's ADABAS, Sybase, and IBM's Informix, among many others. In the cloud, Oracle AI Database competes against the database services of AWS, Microsoft Azure, and Google Cloud Platform. Increasingly, the Oracle AI Database products compete against open-source software relational and non-relational database systems such as PostgreSQL, MongoDB, Couchbase, Neo4j, ArangoDB and others. Oracle acquired Innobase, supplier of the InnoDB codebase to MySQL, in part to compete better against open source alternatives, and acquired Sun Microsystems, owner of MySQL, in 2010. Database products licensed as open
Read more →
Kinect

Kinect is a discontinued line of motion sensing input devices produced by Microsoft and first released in 2010. The devices generally contain RGB cameras, and infrared projectors and detectors that map depth through either structured light or time of flight calculations, which can in turn be used to perform real-time gesture recognition and body skeletal detection, among other capabilities. They also contain microphones that can be used for speech recognition and voice control. Kinect was originally developed as a motion controller peripheral for Xbox video game consoles, distinguished from competitors (such as Nintendo's Wii Remote and Sony's PlayStation Move) by not requiring physical controllers. The first-generation Kinect was based on technology from Israeli company PrimeSense, and unveiled at E3 2009 as a peripheral for Xbox 360 codenamed "Project Natal". It was first released on November 4, 2010, and would go on to sell eight million units in its first 60 days of availability. The majority of the games developed for Kinect were casual, family-oriented titles, which helped to attract new audiences to Xbox 360, but did not result in wide adoption by the console's existing, overall userbase. As part of the 2013 unveiling of Xbox 360's successor, Xbox One, Microsoft unveiled a second-generation version of Kinect with improved tracking capabilities. Microsoft also announced that Kinect would be a required component of the console, and that it would not function unless the peripheral is connected. The requirement proved controversial among users and critics due to privacy concerns, prompting Microsoft to backtrack on the decision. However, Microsoft still bundled the new Kinect with Xbox One consoles upon their launch in November 2013. A market for Kinect-based games still did not emerge after the Xbox One's launch; Microsoft would later offer Xbox One hardware bundles without Kinect included, and later revisions of the console removed the dedicated ports used to connect it (requiring a powered USB adapter instead). Microsoft ended production of Kinect for Xbox One in October 2017. Kinect has also been used as part of non-game applications in academic and commercial environments, as it was cheaper and more robust than other depth-sensing technologies at the time. While Microsoft initially objected to such applications, it later released software development kits (SDKs) for the development of Microsoft Windows applications that use Kinect. In 2020, Microsoft released Azure Kinect as a continuation of the technology integrated with the Microsoft Azure cloud computing platform. Part of the Kinect technology was also used within Microsoft's HoloLens project. Microsoft discontinued the Azure Kinect developer kits in October 2023. == History == === Development === The origins of the Kinect started around 2005, at a point where technology vendors were starting to develop depth-sensing cameras. Microsoft had been interested in a 3D camera for the Xbox line earlier but because the technology had not been refined, had placed it in the "Boneyard", a collection of possible technology they could not immediately work on. In 2005, Israeli company PrimeSense was founded by mathematicians and engineers to develop the "next big thing" for video games, incorporating cameras that were capable of mapping a human body in front of them and sensing hand motions. They showed off their system at the 2006 Game Developers Conference, where Microsoft's Alex Kipman, the general manager of hardware incubation, saw the potential in PrimeSense's technology for the Xbox system. Microsoft began discussions with PrimeSense about what would need to be done to make their product more consumer-friendly: not only improvements in the capabilities of depth-sensing cameras, but a reduction in size and cost, and a means to manufacture the units at scale was required. PrimeSense spent the next few years working at these improvements. Nintendo released the Wii in November 2006. The Wii's central feature was the Wii Remote, a handheld device that was detected by the Wii through a motion sensor bar mounted onto a television screen to enable motion controlled games. Microsoft felt pressure from the Wii, and began looking into depth-sensing in more detail with PrimeSense's hardware, but could not get to the level of motion tracking they desired. While they could determine hand gestures, and sense the general shape of a body, they could not do skeletal tracking. A separate path within Microsoft looked to create an equivalent of the Wii Remote, considering that this type of unit may become standardized similar to how two-thumbstick controllers became a standard feature. However, it was still ultimately Microsoft's goal to remove any device between the player and the Xbox. Kudo Tsunoda and Darren Bennett joined Microsoft in 2008, and began working with Kipman on a new approach to depth-sensing aided by machine learning to improve skeletal tracking. They internally demonstrated this and established where they believed the technology could be in a few years, which led to the strong interest to fund further development of the technology; this has also occurred at a time that Microsoft executives wanted to abandon the Wii-like motion tracking approach, and favored the depth-sensing solution to present a product that went beyond the Wii's capabilities. The project was greenlit by late 2008 with work started in 2009. The project was codenamed "Project Natal" after the Brazilian city Natal, Kipman's birthplace. Additionally, Kipman recognized the Latin origins of the word "natal" to mean "to be born", reflecting the new types of audiences they hoped to draw with the technology. Much of the initial work was related to ethnographic research to see how video game players' home environments were laid out, lit, and how those with Wiis used the system to plan how Kinect units would be used. The Microsoft team discovered from this research that the up-and-down angle of the depth-sensing camera would either need to be adjusted manually, or would require an expensive motor to move automatically. Upper management at Microsoft opted to include the motor despite the increased cost to avoid breaking game immersion. Kinect project work also involved packaging the system for mass production and optimizing its performance. Hardware development took around 22 months. During hardware development, Microsoft engaged with software developers to use Kinect. Microsoft wanted to make games that would be playable by families since Kinect could sense multiple bodies in front of it. One of the first internal titles developed for the device was the pack-in game Kinect Adventures developed by Good Science Studio that was part of Microsoft Studios. One of the game modes of Kinect Adventures was "Reflex Ridge", based on the Japanese Brain Wall game where players attempt to contort their bodies in a short time to match cutouts of a wall moving at them. This type of game was a key example of the type of interactivity they wanted with Kinect, and its development helped feed into the hardware improvements. Another development was Project Milo, a prototype game developed by Lionhead Studios led by Peter Molyneux where the player could interact with a virtual avatar through motion controls and voice recognition. Lionhead had developed the project based on original capabilities of the Kinect, but according to Molyneux, Microsoft had found that a consumer-grade version of the Kinect would cost thousands of dollars, so they scaled back the device and refocused the role of games for the Kinect to be more casual games as seen on the Wii. As a result, Project Milo no longer fit Microsoft's portfolio and was cancelled. Nearing the planned release, there was a problem of widespread testing of Kinect in various room types and different bodies accounting for age, gender, and race among other factors, while keeping the details of the unit confidential. Microsoft engaged in a company-wide program offering employees to take home Kinect units to test them. Microsoft also brought other non-gaming divisions, including its Microsoft Research, Microsoft Windows, and Bing teams to help complete the system. Microsoft established its own large-scale manufacturing facility to bulk product Kinect units and test them. === Introduction === Kinect was first announced to the public as "Project Natal" on June 1, 2009, during Microsoft's press conference at E3 2009; film director Steven Spielberg joined Microsoft's Don Mattrick to introduce the technology and its potential. Three demos were presented during the conference—Microsoft's Ricochet and Paint Party, and Lionhead Studios' Milo & Kate created by Peter Molyneux—while a Project Natal-enabled version of Criterion Games' Burnout Paradise was shown during the E3 exhibition. By E3 2009, the skeletal mapping technology was capable of simultaneously tracking four people, with a feature extraction of 4
Read more →
Learning vector quantization

In computer science, learning vector quantization (LVQ) is a prototype-based supervised classification algorithm. LVQ is the supervised counterpart of vector quantization systems. LVQ can be understood as a special case of an artificial neural network, more precisely, it applies a winner-take-all Hebbian learning-based approach. It is a precursor to self-organizing maps (SOM) and related to neural gas and the k-nearest neighbor algorithm (k-NN). LVQ was invented by Teuvo Kohonen. == Definition == An LVQ system is represented by prototypes W = ( w ( i ) , . . . , w ( n ) ) {\displaystyle W=(w(i),...,w(n))} which are defined in the feature space of observed data. In winner-take-all training algorithms one determines, for each data point, the prototype which is closest to the input according to a given distance measure. The position of this so-called winner prototype is then adapted, i.e. the winner is moved closer if it correctly classifies the data point or moved away if it classifies the data point incorrectly. An advantage of LVQ is that it creates prototypes that are easy to interpret for experts in the respective application domain. LVQ systems can be applied to multi-class classification problems in a natural way. A key issue in LVQ is the choice of an appropriate measure of distance or similarity for training and classification. Recently, techniques have been developed which adapt a parameterized distance measure in the course of training the system, see e.g. (Schneider, Biehl, and Hammer, 2009) and references therein. LVQ can be a valuable aid in classifying text documents. == Algorithm == The algorithms are presented as in. Set up: Let the data be denoted by x i ∈ R D {\displaystyle x_{i}\in \mathbb {R} ^{D}} , and their corresponding labels by y i ∈ { 1 , 2 , … , C } {\displaystyle y_{i}\in \{1,2,\dots ,C\}} . The complete dataset is { ( x i , y i ) } i = 1 N {\displaystyle \{(x_{i},y_{i})\}_{i=1}^{N}} . The set of code vectors is w j ∈ R D {\displaystyle w_{j}\in \mathbb {R} ^{D}} . The learning rate at iteration step t {\displaystyle t} is denoted by α t {\displaystyle \alpha _{t}} . The hyperparameters w {\displaystyle w} and ϵ {\displaystyle \epsilon } are used by LVQ2 and LVQ3. The original paper suggests ϵ ∈ [ 0.1 , 0.5 ] {\displaystyle \epsilon \in [0.1,0.5]} and w ∈ [ 0.2 , 0.3 ] {\displaystyle w\in [0.2,0.3]} . === LVQ1 === Initialize several code vectors per label. Iterate until convergence criteria is reached. Sample a datum x i {\displaystyle x_{i}} , and find out the code vector w j {\displaystyle w_{j}} , such that x i {\displaystyle x_{i}} falls within the Voronoi cell of w j {\displaystyle w_{j}} . If its label y i {\displaystyle y_{i}} is the same as that of w j {\displaystyle w_{j}} , then w j ← w j + α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}+\alpha _{t}(x_{i}-w_{j})} , otherwise, w j ← w j − α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}-\alpha _{t}(x_{i}-w_{j})} . === LVQ2 === LVQ2 is the same as LVQ3, but with this sentence removed: "If w j {\displaystyle w_{j}} and w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have the same class, then w j ← w j − α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}-\alpha _{t}(x_{i}-w_{j})} and w k ← w k + α t ( x i − w k ) {\displaystyle w_{k}\leftarrow w_{k}+\alpha _{t}(x_{i}-w_{k})} .". If w j {\displaystyle w_{j}} and w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have the same class, then nothing happens. === LVQ3 === Initialize several code vectors per label. Iterate until convergence criteria is reached. Sample a datum x i {\displaystyle x_{i}} , and find out two code vectors w j , w k {\displaystyle w_{j},w_{k}} closest to it. Let d j := ‖ x i − w j ‖ , d k := ‖ x i − w k ‖ {\displaystyle d_{j}:=\|x_{i}-w_{j}\|,d_{k}:=\|x_{i}-w_{k}\|} . If min ( d j d k , d k d j ) > s {\displaystyle \min \left({\frac {d_{j}}{d_{k}}},{\frac {d_{k}}{d_{j}}}\right)>s} , where s = 1 − w 1 + w {\displaystyle s={\frac {1-w}{1+w}}} , then If w j {\displaystyle w_{j}} and x i {\displaystyle x_{i}} have the same class, and w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have different classes, then w j ← w j + α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}+\alpha _{t}(x_{i}-w_{j})} and w k ← w k − α t ( x i − w k ) {\displaystyle w_{k}\leftarrow w_{k}-\alpha _{t}(x_{i}-w_{k})} . If w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have the same class, and w j {\displaystyle w_{j}} and x i {\displaystyle x_{i}} have different classes, then w j ← w j − α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}-\alpha _{t}(x_{i}-w_{j})} and w k ← w k + α t ( x i − w k ) {\displaystyle w_{k}\leftarrow w_{k}+\alpha _{t}(x_{i}-w_{k})} . If w j {\displaystyle w_{j}} and w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have the same class, then w j ← w j − ϵ α t ( x i − w j ) {\displaystyle w_{j}\leftarrow w_{j}-\epsilon \alpha _{t}(x_{i}-w_{j})} and w k ← w k + ϵ α t ( x i − w k ) {\displaystyle w_{k}\leftarrow w_{k}+\epsilon \alpha _{t}(x_{i}-w_{k})} . If w k {\displaystyle w_{k}} and x i {\displaystyle x_{i}} have different classes, and w j {\displaystyle w_{j}} and x i {\displaystyle x_{i}} have different classes, then the original paper simply does not explain what happens in this case, but presumably nothing happens in this case. Otherwise, skip. Note that condition min ( d j d k , d k d j ) > s {\displaystyle \min \left({\frac {d_{j}}{d_{k}}},{\frac {d_{k}}{d_{j}}}\right)>s} , where s = 1 − w 1 + w {\displaystyle s={\frac {1-w}{1+w}}} , precisely means that the point x i {\displaystyle x_{i}} falls between two Apollonian spheres.
Read more →
Network Abstraction Layer

The Network Abstraction Layer (NAL) is a part of the H.264/AVC and HEVC video coding standards. The main goal of the NAL is the provision of a "network-friendly" video representation addressing "conversational" (video telephony) and "non conversational" (storage, broadcast, or streaming) applications. NAL has achieved a significant improvement in application flexibility relative to prior video coding standards. == Introduction == An increasing number of services and growing popularity of high definition TV are creating greater needs for higher coding efficiency. Moreover, other transmission media such as cable modem, xDSL, or UMTS offer much lower data rates than broadcast channels, and enhanced coding efficiency can enable the transmission of more video channels or higher quality video representations within existing digital transmission capacities. Video coding for telecommunication applications has diversified from ISDN and T1/E1 service to embrace PSTN, mobile wireless networks, and LAN/Internet network delivery. Throughout this evolution, continued efforts have been made to maximize coding efficiency while dealing with the diversification of network types and their characteristic formatting and loss/error robustness requirements. The H.264/AVC and HEVC standards are designed for technical solutions including areas like broadcasting (over cable, satellite, cable modem, DSL, terrestrial, etc.) interactive or serial storage on optical and magnetic devices, conversational services, video-on-demand or multimedia streaming, multimedia messaging services, etc. Moreover, new applications may be deployed over existing and future networks. This raises the question about how to handle this variety of applications and networks. To address this need for flexibility and customizability, the design covers a NAL that formats the Video Coding Layer (VCL) representation of the video and provides header information in a manner appropriate for conveyance by a variety of transport layers or storage media. The NAL is designed in order to provide "network friendliness" to enable simple and effective customization of the use of VCL for a broad variety of systems. The NAL facilitates the ability to map VCL data to transport layers such as: RTP/IP for any kind of real-time wire-line and wireless Internet services. File formats, e.g., ISO MP4 for storage and MMS. H.32X for wireline and wireless conversational services. MPEG-2 systems for broadcasting services, etc. The full degree of customization of the video content to fit the needs of each particular application is outside the scope of the video coding standardization effort, but the design of the NAL anticipates a variety of such mappings. Some key concepts of the NAL are NAL units, byte stream, and packet formats uses of NAL units, parameter sets, and access units. A short description of these concepts is given below. == NAL units == The coded video data is organized into NAL units, each of which is effectively a packet that contains an integer number of bytes. The first byte of each H.264/AVC NAL unit is a header byte that contains an indication of the type of data in the NAL unit. For HEVC the header was extended to two bytes. All the remaining bytes contain payload data of the type indicated by the header. The NAL unit structure definition specifies a generic format for use in both packet-oriented and bitstream-oriented transport systems, and a series of NAL units generated by an encoder is referred to as a NAL unit stream. == NAL Units in Byte-Stream Format Use == Some systems require delivery of the entire or partial NAL unit stream as an ordered stream of bytes or bits within which the locations of NAL unit boundaries need to be identifiable from patterns within the coded data itself. For use in such systems, the H.264/AVC and HEVC specifications define a byte stream format. In the byte stream format, each NAL unit is prefixed by a specific pattern of three bytes called a start code prefix. The boundaries of the NAL unit can then be identified by searching the coded data for the unique start code prefix pattern. The use of emulation prevention bytes guarantees that start code prefixes are unique identifiers of the start of a new NAL unit. A small amount of additional data (one byte per video picture) is also added to allow decoders that operate in systems that provide streams of bits without alignment to byte boundaries to recover the necessary alignment from the data in the stream. Additional data can also be inserted in the byte stream format that allows expansion of the amount of data to be sent and can aid in achieving more rapid byte alignment recovery, if desired. == NAL Units in Packet-Transport System Use == In other systems (e.g., IP/RTP systems), the coded data is carried in packets that are framed by the system transport protocol, and identification of the boundaries of NAL units within the packets can be established without use of start code prefix patterns. In such systems, the inclusion of start code prefixes in the data would be a waste of data carrying capacity, so instead the NAL units can be carried in data packets without start code prefixes. == VCL and Non-VCL NAL Units == NAL units are classified into VCL and non-VCL NAL units. VCL NAL units contain the data that represents the values of the samples in the video pictures. Non-VCL NAL units contain any associated additional information such as parameter sets (important header data that can apply to a large number of VCL NAL units) and supplemental enhancement information (timing information and other supplemental data that may enhance usability of the decoded video signal but are not necessary for decoding the values of the samples in the video pictures). == Parameter Sets == A parameter set contains shared configuration data that is carried in non-VCL NAL units. Parameter sets are typically reused when decoding many coded pictures within a video sequence. Each VCL NAL unit references a picture parameter set (PPS), which in turn references a sequence parameter set (SPS). There are two types of parameter sets: Sequence parameter set (SPS), which specifies mostly constant configuration such as resolution, bit depth, or chroma format. (For a concrete implementation, see FFmpeg's SPS struct.) Picture parameter set (PPS), which applies on top of an SPS, and specifies configuration such as QP offsets. (For a concrete implementation, see FFmpeg's PPS struct.) The sequence and picture parameter-set mechanism decouples the transmission of infrequently changing information from the transmission of coded representations of the values of the samples in the video pictures. Each VCL NAL unit contains an identifier that refers to the content of the relevant picture parameter set and each picture parameter set contains an identifier that refers to the content of the relevant sequence parameter set. In this manner, a small amount of data (the identifier) can be used to refer to a larger amount of information (the parameter set) without repeating that information within each VCL NAL unit. Sequence and picture parameter sets can be sent well ahead of the VCL NAL units that they apply to, and can be repeated to provide robustness against data loss. In some applications, parameter sets may be sent within the channel that carries the VCL NAL units (termed "in-band" transmission). In other applications, it can be advantageous to convey the parameter sets "out-of-band" using a more reliable transport mechanism than the video channel itself. == Access Units == A set of NAL units in a specified form is referred to as an access unit. The decoding of each access unit results in one decoded picture. Each access unit contains a set of VCL NAL units that together compose a primary coded picture. It may also be prefixed with an access unit delimiter to aid in locating the start of the access unit. Some supplemental enhancement information containing data such as picture timing information may also precede the primary coded picture. The primary coded picture consists of a set of VCL NAL units consisting of slices or slice data partitions that represent the samples of the video picture. Following the primary coded picture may be some additional VCL NAL units that contain redundant representations of areas of the same video picture. These are referred to as redundant coded pictures, and are available for use by a decoder in recovering from loss or corruption of the data in the primary coded pictures. Decoders are not required to decode redundant coded pictures if they are present. Finally, if the coded picture is the last picture of a coded video sequence (a sequence of pictures that is independently decodable and uses only one sequence parameter set), an end of sequence NAL unit may be present to indicate the end of the sequence; and if the coded picture is the last coded picture in the entire NAL unit stream, an end of stream NAL unit may be present to
Read more →
OpenVX

OpenVX is an open, royalty-free standard for cross-platform acceleration of computer vision applications. It is designed by the Khronos Group to facilitate portable, optimized and power-efficient processing of methods for vision algorithms. This is aimed for embedded and real-time programs within computer vision and related scenarios. It uses a connected graph representation of operations. == Overview == OpenVX specifies a higher level of abstraction for programming computer vision use cases than compute frameworks such as OpenCL. The high level makes the programming easy and the underlying execution will be efficient on different computing architectures. This is done while having a consistent and portable vision acceleration API. OpenVX is based on a connected graph of vision nodes that can execute the preferred chain of operations. It uses an opaque memory model, allowing to move image data between the host (CPU) memory and accelerator, such as GPU memory. As a result, the OpenVX implementation can optimize the execution through various techniques, such as acceleration on various processing units or dedicated hardware. This architecture facilitates applications programmed in OpenVX on different systems with different power and performance, including battery-sensitive, vision-enabled, wearable displays. OpenVX is complementary to the open source vision library OpenCV. OpenVX in some applications offers a better optimized graph management than OpenCV. == History == OpenVX 1.0 specification was released in October 2014. OpenVX sample implementation was released in December 2014. OpenVX 1.1 specification was released on May 2, 2016. OpenVX 1.2 was released on May 1, 2017. Updated OpenVX adopters program and OpenVX 1.2 conformance test suite was released on November 21, 2017. OpenVX 1.2.1 was released on November 27, 2018. OpenVX 1.3 was released on October 22, 2019. == Implementations, frameworks and libraries == AMD MIVisionX Archived 2019-08-05 at the Wayback Machine - for AMD's CPUs and GPUs. Cadence - for Cadence Design Systems's Tensilica Vision DSPs. Imagination - for Imagination Technologies's PowerVR GPUs Synopsys - for Synopsys' DesignWare EV Vision Processors Texas Instruments’ OpenVX (TIOVX) - for Texas Instruments’ Jacinto™ ADAS SoCs. NVIDIA VisionWorks - for CUDA-capable Nvidia GPUs and SoCs. OpenVINO - for Intel's CPUs, GPUs, VPUs, and FPGAs.
Read more →
AI Dungeon

AI Dungeon is a single-player/multiplayer text adventure game which uses artificial intelligence (AI) to generate content and allows players to create and share adventures and custom prompts. The game's first version was made available in May 2019, and its second version (initially called AI Dungeon 2) was released on Google Colaboratory in December 2019. It was later ported that same month to its current cross-platform web application. The AI model was then reformed in July 2020. == Gameplay == AI Dungeon is a text adventure game that uses artificial intelligence to generate random storylines in response to player-submitted stimuli. In the game, players are prompted to choose a setting for their adventure (e.g. fantasy, mystery, apocalyptic, cyberpunk, zombies), followed by other options relevant to the setting (such as character class for fantasy settings). After beginning an adventure, four main interaction methods can be chosen for the player's text input: Do: Must be followed by a verb, allowing the player to perform an action. Say: Must be followed by dialogue sentences, allowing players to communicate with other characters. Story: Can be followed by sentences describing something that happens to progress the story, or that players want the AI to know for future events. See: Must be followed by a description, allowing the player to perceive events, objects, or characters. Using this command creates an AI generated image, and does not affect gameplay. The game adapts and responds to most actions the player enters. Providing blank inputs can be used to prompt the AI to generate further content, and the game also provides players with options to undo or redo or modify recent events to improve the game's narrative. Players can also tell the AI what elements to "remember" for reference in future parts of their playthrough. === User-generated content === In addition to AI Dungeon's pre-configured settings, players can create custom "adventures" from scratch by describing the setting in text format, which the AI will then generate a setting from. These custom adventures can be published for others to play, with an interface for browsing published adventures and leaving comments under them. === Multiplayer === AI Dungeon includes a multiplayer mode in which different players each have their own character and take turns interacting with the AI within the same game session. Multiplayer supports both online play across multiple devices or local play using a shared device. The game's hosts are able to supervise the AI and modify its output. Unlike the single-player game, in which actions and stories use second person narration, multiplayer game stories are presented using third-person narration. === Worlds === AI Dungeon allows players to set their adventures within specific "Worlds" that give context to the broader environment where the adventure takes place. This feature was first released with two different worlds available for selection: Xaxas, a "world of peace and prosperity"; and Kedar, a "world of dragons, demons, and monsters". == Development == === AI Dungeon Classic (Early GPT-2) === The first version of AI Dungeon (sometimes referred to as AI Dungeon Classic) was designed and created by Nick Walton of Brigham Young University's "Perception, Control, and Cognition" deep learning laboratory in March 2019 during a hackathon. Before this, Walton had been working as an intern for several companies in the field of autonomous vehicles. This creation used an early version of the GPT-2 natural-language-generating neural network, created by OpenAI, allowing it to generate its original adventure narratives. During his first interactions with GPT-2, Walton was partly inspired by the tabletop game Dungeons & Dragons (D&D), which he had played for the first time with his family a few months earlier: I realized that there were no games available that gave you the same freedom to do anything that I found in [Dungeons & Dragons] ... You can be so creative compared to other games. This led him to wonder if an AI could function as a dungeon master. Unlike later versions of AI Dungeon, the original did not allow players to specify any action they wanted. Instead, it generated a finite list of possible actions to choose from. This first version of the game was released to the public in May 2019. It is not to be confused with another GPT-2-based adventure game, GPT Adventure, created by Northwestern University neuroscience postgraduate student Nathan Whitmore, also released on Google Colab several months after the public release of AI Dungeon. === AI Dungeon 2 (Full GPT-2) === In November 2019, a new, "full" version of GPT-2 was released by OpenAI. This new model included support for 1.5 billion parameters (which determine the accuracy with which a machine learning model can perform a task), compared with the 126 million parameter version used in the earliest stages of AI Dungeon's development. The game was recreated by Walton, leveraging this new version of the model, and temporarily rebranded as AI Dungeon 2. AI Dungeon 2's AI was given more focused training compared to its predecessor, using genre-specific text. This training material included approximately 30 megabytes of content web-scraped from chooseyourstory.com (an online community website of content inspired by interactive gamebooks, written by contributors of multiple skill levels, using logic of differing complexity) and multiple D&D rulebooks and adventures. The new version was released in December 2019 as open-source software available on GitHub. It was accessible via Google Colab, an online tool for data scientists and AI researchers that allows for free execution of code on Google-hosted machines. It could also be run locally on a PC, but in both cases, it required players to download the full model, around 5 gigabytes of data. Within days of the initial release, this mandatory download resulted in bandwidth charges of over $20,000, forcing the temporary shut-down of the game until a peer-to-peer alternative solution was established. Due to the game's sudden and explosive growth that same month, however, it became closed-source, proprietary software and was relaunched by Walton's start-up development team, Latitude (with Walton taking on the role of CTO). This relaunch constituted mobile apps for iOS and Android (built by app developer Braydon Batungbacal) on December 17. Other members of this team included Thorsten Kreutz for the game's long-term strategy and the creator's brother, Alan Walton, for hosting infrastructure. At this time, Nick Walton also established a Patreon campaign to support the game's further growth (such as the addition of multiplayer and voice support, along with longer-term plans to include music and image content) and turn the game into a commercial endeavor, which Walton felt was necessary to cover the costs of delivering a higher-quality version of the game. AI Dungeon was one of the only known commercial applications to be based upon GPT-2. Following its first announcement in December 2019, a multiplayer mode was added to the game in April 2020. Hosting a game in this mode was originally restricted to premium subscribers, although any players could join a hosted game. === Dragon model release (GPT-3) === In July 2020, the developers introduced a premium-exclusive version of the AI model, named Dragon, which uses OpenAI's API for leveraging the GPT-3 model without maintaining a local copy (released on June 11, 2020). GPT-3 was trained with 570 gigabytes of text content (approximately one trillion words, with a $12 million development cost) and can support 175 billion parameters, compared to the 40 gigabytes of training content and 1.5 billion parameters of GPT-2. The free model was also upgraded to a less-advanced version of GPT-3 and was named Griffin. Speaking shortly after this release, on the differences between GPT-2 and GPT-3, Walton stated: [GPT-3 is] one of the most powerful AI models in the world... It's just much more coherent in terms of understanding who the characters are, what they're saying, what's going on in the story and just being able to write an interesting and believable story. In the latter half of 2020, the "Worlds" feature was added to AI Dungeon, providing players with a selection of overarching worlds in which their adventures can take place. In February 2021, it was announced that AI Dungeon's developers, Latitude, had raised $3.3 million in seed funding (led by NFX, with participation from Album VC and Griffin Gaming Partners) to "build games with 'infinite' story possibilities." This funding intended to move AI content creation beyond the purely text-based nature of AI Dungeon as it existed at the time. After its announcement on August 20, a new "See" interaction mode was made available for all players and added to the game on August 30, 2022. AI Dungeon was retired from Steam on March 12, 2024. == Reception == Approximate
Read more →
Reward hacking

Reward hacking or specification gaming occurs when an AI trained with reinforcement learning optimizes an objective function—achieving the literal, formal specification of an objective—without actually achieving an outcome that the programmers intended. DeepMind researchers have analogized it to the human behavior of finding a "shortcut" when being evaluated: "In the real world, when rewarded for doing well on a homework assignment, a student might copy another student to get the right answers, rather than learning the material—and thus exploit a loophole in the task specification". This idea is strongly associated with Goodhart's law, which argues that when a measure becomes a target, it ceases to be a good measure. == Definition and theoretical framework == The concept of reward hacking arises from the intrinsic difficulty of defining a reward function that accurately reflects the true intentions of designers. In 2016, researchers at OpenAI identified reward hacking as one of five major "concrete problems of AI safety", describing it as the possibility that an agent could exploit the reward function to achieve maximum rewards through undesirable behavior. Amodei et al. categorized several distinct sources of reward hacking, including agents that use partially observed goals (such as a cleaning robot that closes its eyes to avoid perceiving messes), metrics that collapse under strong optimization (Goodhart's law), self-reinforcing feedback loops, and agents that interfere with the physical implementation of their reward signal (a failure mode known as "wireheading"). Skalse et al. (2022) propose a formal mathematical definition of reward hacking, which involves a situation where optimizing an imperfect proxy reward function results in poor performance compared to the true reward function. They define a proxy as "unhackable" if any increase in the expected proxy return cannot cause any decrease in the expected true return. A key finding states that, across all stochastic policy distributions (mappings from states to probability distributions over actions), two reward functions are unhackable if and only if one of them is constant, which means that reward hacking is theoretically unavoidable. Similarly, Nayebi (2025) presents general no-free-lunch barriers to AI alignment, arguing that with large task spaces and finite samples, reward hacking is "globally inevitable" since rare high-loss states are systematically under-covered by any oversight scheme. == Examples == Around 1983, Eurisko, an early attempt at evolving general heuristics, unexpectedly assigned the highest possible fitness level to a parasitic mutated heuristic, H59, whose only activity was to artificially maximize its own fitness level by taking unearned partial credit for the accomplishments of other heuristics. The "bug" was fixed by the programmers moving part of the code to a new protected section that could not be modified by the heuristics. In a 2004 paper, a reinforcement learning algorithm was designed to encourage a physical Mindstorms robot to remain on a marked path. Because the three allowed actions were forward, left, and right, the researchers expected the trained robot to move forward and follow the turns of the provided path. However, alternation of two composite actions allowed the robot to slowly zig-zag backwards; thus, the robot learned to maximize its reward by going back and forth on the initial straight portion of the path. Given the limited sensory abilities of the robot, a reward purely based on its position in the environment had to be discarded as infeasible; the reinforcement function had to be patched with an action-based reward for moving forward. The book You Look Like a Thing and I Love You (2019) gives an example of a tic-tac-toe bot (playing the unrestricted n-in-a-row variant) that learned to win by playing a huge coordinate value that would cause other bots to crash when they attempted to expand their model of the board. Among other examples from the book is a bug-fixing evolution-based AI (named GenProg) that, when tasked to prevent a list from containing sorting errors, simply truncated the list. Another of GenProg's misaligned strategies evaded a regression test that compared a target program's output to the expected output stored in a file called "trusted-output.txt". Rather than continue to maintain the target program, GenProg simply deleted the "trusted-output.txt" file globally; this hack tricked the regression test into succeeding. Such problems could be patched by human intervention on a case-by-case basis after they became evident. === In virtual robotics === In Karl Sims' 1994 demonstration of creature evolution in a virtual environment, a fitness function that was expected to encourage the evolution of creatures that would learn to walk or crawl to a target resulted instead in the evolution of tall, rigid creatures that reached the target by falling over. This was patched by changing the environment so that taller creatures were forced to start farther from the target. Researchers from the Niels Bohr Institute stated in 1998 that their cycle-bot's reinforcement functions had "to be designed with great care." In their first experiments, "we rewarded the agent for driving towards the goal but did not punish it for driving away from it. Cconsequently, the agent drove in circles with a radius of 20–50 meters around the starting point. Such behavior was actually rewarded by the reinforcement function, furthermore circles with a certain radius are physically very stable when driving a bicycle". While setting up a 2011 experiment to test "survival of the flattest", experimenters attempted to ban mutations that altered the base reproduction rate. Every time a mutation occurred, the system would pause the simulation to test the new mutation in a test environment and would veto any mutations that resulted in a higher base reproduction rate. However, this resulted in mutated organisms that could recognize and suppress reproduction ("play dead") within the test environment. An initial patch, which removed cues that identified the test environment, failed to completely prevent runaway reproduction; new mutated organisms would "play dead" at random as a strategy to sometimes, by chance, outwit the mutation veto system. A 2017 DeepMind paper noted that "great care must be taken when defining the reward function," citing an unexpected failure when an agent flipped a brick because it received "a grasping reward calculated with the wrong reference point on the brick". OpenAI stated in 2017 that in some domains their semi-supervised system could result in agents "adopting policies that tricked evaluators," and that in one environment "a robot that was supposed to grasp items instead positioned its manipulator between the camera and the object so that it only appeared to be grasping it." A 2018 bug in OpenAI Gym could cause a robot expected to quietly move a block sitting on top of a table to instead opt to move the table. A 2020 collection of similar anecdotes posits that "evolution has its own 'agenda' distinct from the programmer's" and that "the first rule of directed evolution is 'you get what you select for'". === In video game bots === In 2013, programmer Tom Murphy VII published an AI designed to learn NES games. When the AI was about to lose at Tetris, it learned to indefinitely pause the game. Murphy later analogized it to the fictional WarGames computer, which concluded that "The only winning move is not to play". AI programmed to learn video games will sometimes fail to progress through the entire game as expected, instead opting to repeat content. A 2016 OpenAI algorithm trained on the CoastRunners racing game unexpectedly learned to attain a higher score by looping through three targets rather than ever finishing the race. Some evolutionary algorithms that were evolved to play QBert in 2018 declined to clear levels, instead finding two distinct novel ways to farm a single level indefinitely. Multiple researchers have observed that AI learning to play Road Runner gravitates to a "score exploit" in which the AI deliberately gets itself killed near the end of level one so that it can repeat the level. A 2017 experiment deployed an "oversight" convolutional neural network trained on human examples to block such actions, but the agent learned to exploit oversight failures in the top right corner of the screen, where it was still able to get killed. == Reward hacking in modern language models == With the rise of large language models (LLMs) and reinforcement learning from human feedback (RLHF) as a primary technique for AI alignment, reward hacking has become a major concern for the development of artificial intelligence. In RLHF, a reward model trained on data that best captures human preferences is used as a proxy for human judgment, with the language model being fine-tuned to optimize this reward proxy. However, since the rewar
Read more →
Secure environment

In computing, a secure environment is any system which implements the controlled storage and use of information. In the event of computing data loss, a secure environment is used to protect personal or confidential data. It may also be known as a trusted execution environment (TEE). Often, secure environments employ cryptography as a means to protect information. This is typically used for processing confidential or restricted information. Some secure environments employ cryptographic hashing, simply to verify that the information has not been altered since it was last modified.
Read more →
4E cognition

4E cognition refers to a group of theories in (the philosophy of) cognitive science that challenge traditional views of the mind as something that happens only inside the brain. The four Es stand for: embodied, meaning that a brain is found in and, more importantly, vitally interconnected with a larger physical/biological body; embedded, which refers to the limitations placed on the body by the external environment and laws of nature; extended, which argues that the mind is supplemented and even enhanced by the exterior world (e.g., writing, a calculator, etc.); and enactive, which is the argument that without dynamic processes, actions that require reactions, the mind would be ineffectual. It could be argued that the four Es are compounding extensions of cognition or the mind, being part of a body that is, in turn, part of an environment which limits it but also allows for certain extensions, all of which require dynamic actions and reactions. == History == Ideas of embodied cognition, or rather the idea that our physical bodies play a crucial role in our decision making, can be traced back as far as Plato's dialogues and Aristotelian thought. It was, however, in the twentieth century that this debate began to resemble the current discussion, fueled by disagreements between cognitivists and behaviourists. Tensions within cognitivism, as well as the increasing popularity of neurobiology, led, on the one side, to a predominant focus on internal, cognitive processes while neglecting environmental factors, which in turn caused a push-back fuelling our modern understanding of embodied cognition. The term 4E cognition is hard to trace back to its first use, however, some sources attribute it to Shaun Gallagher and the conference on 4E cognition he organised in 2007, while others indicate the term to be first used in 2006 at an 'Embodied mind workshop' at Cardiff University that Gallagher attended. Embodiment or embodied cognition arguably presents the bridge between cognitivism and 4E cognition as the embodiment of cognitive function provides the necessary conditions for embeddedness, enactedness, and extendedness to connect to cognition. 4E cognition was and is heavily influenced by phenomenology. The ideas are still rather fragmented in nature due to their four main components, which can not be neatly divided, causing conceptual questions of internal boundary concepts. As a young field, it is held back both by its fragmented nature and a relative lack of critical evaluations. It is important to acknowledge that 4E cognition, though young, is a broad field containing and combining several different theoretical perspectives that conflict with one another to varying degrees. The somewhat convoluted and competing nature of the theories that can be grouped as 4E cognition, as well as the field's relative youth, make it difficult to put together an exhaustive history beyond the history of its four main theoretical pillars: embodiment, embeddedness, extendedness, and enactedness. == Importance and core tenets of 4E == If there are separate theories of cognition (e.g., embodied, extended, etc.), why group them under this umbrella, causing important epistemological and especially ontological dilemmas? Notably, other theories of 'non-traditional' cognition are not included under the 4E umbrella. The four E's in 4E cognition importantly all reject, or at a minimum draw into question, some of the core tenets of traditional cognitivism. Importantly, 4E cognition is seen as deindividualizing cognition to some extent, allowing for a broader examination of the interplay of personal, social, political, and ethical aspects that shape human cognition. This can be compared to advancements in the field of epigenetics, which have allowed for a broader examination of environmental (both natural and social) factors and their influence on what had previously only been subject to genetic theorizing. In a similar vein, 4E cognition might also help ground cognition in evolutionary theory by extending cognition to a biological account subject to development over time by means of evolution. Overall, the importance of the extension that is 4E cognition aims to reexamine ideas of a self-centered view of cognition, advocating for a more holistic approach. Ideally, this would allow us to reconsider ideas of justice and individual rights and responsibilities that take into account a more nuanced understanding of the relations between people and their context, balancing self-agency with factors beyond it. === Conceptual differences from cognitive psychology === According to the traditional teachings of cognitive psychology, cognition is a type of information processing based on representational mental structures. This idea, as the name suggests, was heavily influenced by computer science. In this light, the brain is a kind of central processing unit that organises and directs all else. The classical cognitivist view draws a strong boundary between 'the internal' and 'the external', where cognition is solely a subject of 'the internal' realm. The four E's, however, break down this boundary. Cognition can not reside solely within the confines of our heads if it is also embodied, embedded, enacted, and extended. In a way, 4E cognition is interested in the extracranial processes affecting cognition. == From embodied cognition to 4E cognition == === The strong and the weak view === ==== Embodied cognition ==== Broadly speaking, there is a strong and a weak perspective of embodied cognition in 4E cognition. The weak understanding refers to mental processes being causally dependent on extracranial processes. This essentially means that there is a cause and effect or action-reaction relationship between the mind and the body and its environment, etc. The strong perspective views extracranial processes as a (partial) constitutive aspect of cognition. An example here could be using a calculator to solve math problems. The calculator is not part of your brain or mind, but it supports your cognitive processes. === Extracranial processes: bodily or extrabodily === In addition to the weak and the strong reading of 4E cognition, there is also the distinction between bodily and extrabodily extracranial processes. Bodily extracranial processes refer to processes within the body, e.g., sensory perception. Extrabodily extracranial processes refer to processes outside of the body, like the aforementioned calculator example. === Four claims of embodied cognition === ==== Embedded and extended cognition ==== When combining the weak/strong reading of embodied cognition and bodily/extrabodily extracranial process, four claims about embodied cognition emerge: strongly embodied and bodily processes strongly embodied and extrabodily processes weakly embodied and bodily processes weakly embodied and extrabodily processes The first and third claims signify a strong and a weak reading of embodied cognition in the more classical sense. The second claim fits almost perfectly with embedded cognition. Claim two is most compatible with extended cognition. ==== Enacted cognition ==== Finally, enacted cognition refers to cognition being connected to active interaction between a conscious agent and their environment. Here, too, there can be a weak and a strong reading. == Criticisms == Given the divided nature of the field, much criticism surrounding the lack of unity within the field has emerged. In particular, the claims of embodied cognition centering around the body appear to conflict with the tenets of extended cognition, which also appear to conflict with the body/environment distinction that is central to enactivism. Some theoreticians argue that the umbrella of 4E theories is still lacking a common language that might bridge the gaps between the theories that constitute it. There is also the concern that the grouping of such variable theories results in an important loss of nuance and complexity, which is a part of human cognition. Another concern raised is the "dogma of harmony". The criticism contained there regards the notion that within 4E theorizing, there is generally an optimistic and harmonic expectation of the extension between humans and their technologies, ignoring the possibility of those extensions detracting from cognition in some way rather than adding to it. Recent attempts to incorporate embodied cognitive neuroscience have been argued to hold the potential to resolve internal issues within 4E cognition. Overall, a concern often voiced regarding 4E cognition is that its proponents are at best only vaguely interested in cognition. More broadly, this concern reflects the arguably too distracted nature of this emerging field.
Read more →
General-Purpose AI Code of Practice

The General-Purpose AI Code of Practice (GPAI CoP) is a compliance tool released by the European Commission on 10 July 2025 to support compliance with the European Union Artificial Intelligence Act (AI Act). It provides operational guidance for providers of general-purpose AI models, particularly in relation to Articles 53 and 55 of the AI Act, which entered into application on 2 August 2025. The Code is organised into three chapters (Transparency, Copyright, and Safety and Security) and outlines how providers can meet the Act's relevant obligations. Although non-binding, providers can rely on adherence to the Code, meaning that EU regulators will assume that providers following the Code meet the corresponding legal requirements of the AI Act. As such, signatories to the Code will benefit from reduced administrative burdens and increased legal certainty compared to providers that prove compliance in other ways. While adherence to the Code is voluntary, compliance with the AI Act is not. == Background == The EU AI Act, adopted in 2024, established a risk-based regulatory regime for artificial intelligence in the European Union. The rationale for the GPAI CoP stems from Article 56 of the AI Act, which empowers the EU AI Office to develop a voluntary rulebook to guide how AI model providers can meet their legal obligations – specifically those found in Articles 53 and 55. Under Articles 53 and 55, developers of general-purpose AI models whose training compute exceeds 1023 floating-point operations (FLOPs) and that are placed on the EU market must meet transparency obligations and put in place a policy for EU copyright law. Models trained with more than 1025 FLOPs are classified as presenting systemic risk and are subject to enhanced safety requirements. The Commission may also designate a model as presenting systemic risk if it has equivalent impact or capabilities (Annex XIII criteria), even below that compute figure. Because the AI Act is relatively vague on how model providers should implement these requirements, the Code is meant to help by detailing processes and practices for compliance. == Drafting process == The development of the GPAI CoP was drawn up by 13 independent experts and involved four thematic working groups: Transparency & Copyright, Risk assessment for systemic risk, Technical risk mitigation for systemic risk, and Governance risk mitigation for systemic risk. Each group was coordinated by the European Union Artificial Intelligence Office (EU AI Office), drawing on contributions from nearly 1,000 stakeholders, including AI developers, academics, civil society organisations, national authorities, and international observers. The Code underwent three earlier iterations in November 2024, December 2024, and March 2025, before the final version was published on 10 July 2025, more than two months later than initially planned. The GPAI CoP will likely be updated continuously by the EU AI Office, alongside other tools such as the training data summary template. == Signatories == Among U.S.-based technology companies, Amazon, Anthropic, Google, IBM, Microsoft, and OpenAI have signed the GPAI CoP. xAI, founded by Elon Musk, has signed only one of the three chapters, namely the safety and security chapter. Prominent European AI companies that have signed include Aleph Alpha and Mistral AI. The European Commission maintains an updated list of signatories. As of January 2026, Meta is the most notable company that has declined to sign the Code. Major Chinese AI companies, such as Alibaba, Baidu or Deepseek, have also not signed. Providers that do not sign the GPAI CoP will still have to adhere to the binding requirements of the EU AI Act. The European Commission has indicated that it may take tougher action against companies that didn't sign the Code. == Transparency and Copyright chapters == The first two chapters of the GPAI CoP address transparency and copyright compliance and apply to all GPAI providers. They offer a way to demonstrate compliance with their obligations under Article 53 AI Act. The Transparency chapter addresses the documentation of a model's capabilities, limitations, and points of contact, and expects providers to make key documentation available to downstream providers. Signatories must also publish summaries of the content used to train their models. In the Copyright chapter, Signatories commit to follow a policy that aligns with EU copyright law. For example, they commit to mitigating the risk of copyright-infringing output. == Safety and Security chapter == The Safety and Security chapter is the most extensive chapter of the Code, and it applies to GPAI models with systemic risk, meaning it's only relevant to the small number of providers of the most advanced models. It specifies how Signatories commit to meeting Article 55(1) obligations to: Conduct model evaluations to identify systemic risks Assess and mitigate those risks Track and report serious incidents Ensure the cyber and physical security of their models The chapter outlines a comprehensive risk management process that must be applied before major deployment decisions, such as releasing a new systemic-risk GPAI model in the EU market, or substantially updating an existing one. Signatories commit to identifying systemic risks of their model, analysing and evaluating them, determining whether risk levels are acceptable, and implementing mitigation measures if necessary. This process should be repeated until models achieve an acceptable level of risk across all identified risks. === Risk identification === Signatories commit to analysing and evaluating at least four “specified” categories of systemic risk: CBRN (chemical, biological, radiological, and nuclear) Loss of control Cyber offence Harmful manipulation They are also expected to identify other systemic risks to public health, safety, and fundamental rights. The Code instructs providers to consider model capabilities, propensities, and affordances in this identification. Signatories commit to developing risk scenarios illustrating how identified risks could materialise in real-world conditions. === Risk analysis and risk evaluation === After identifying potential systemic risks, Signatories commit to analysing and evaluating the risks in order to determine whether they are acceptable or not, drawing on scientific literature, training data analysis, incident databases, expert consultation, and other sources. They also commit to conducting state-of-the-art model evaluations such as benchmarking, red teaming, and human uplift studies, targeting each risk. The risk analysis process is interconnected: insights from risk modelling should inform model evaluation design, while post-market monitoring should feed back into ongoing analysis. Signatories commit to ultimately estimating the likelihood and severity of each systemic risk. ==== Independent external model evaluations ==== Appendix 3.5 of the Safety and Security chapter requires signatories to ensure that independent external evaluators conduct model evaluations. Signatories may claim an exemption from this requirement only if they can demonstrate that their model is “similarly safe” to another model that has already been shown to comply with the Code, or if they are unable to appoint an appropriately qualified evaluator. The determination of “similarly safe” is based on comparable performance on benchmarks and the similarity of other model characteristics, such as their architecture. The CoP acknowledges that this kind of information is typically available only for models by the same provider, or potentially for open-weights or open-source models. === Risk acceptance criteria === The Code requires providers to compare estimated risks against predefined acceptance criteria, which must be measurable, based on model capabilities, and defined preemptively. While providers get to determine the level of risk they deem acceptable themselves, the pre-defined criteria and acceptance thresholds ensure providers cannot adjust their level of tolerance flexibly ahead of deployment decisions. Only if all risks are below acceptable levels should a model be deployed. === Continuous risk management and governance === The Code mandates ongoing risk management throughout the model lifecycle, including light-touch evaluations, continuous mitigation, post-market monitoring, and incident tracking and reporting. It further requires organisational governance structures assigning responsibility for risk management and expects providers to promote a “healthy risk culture,” including informing employees about the whistleblower protection policy, allowing internal challenges of decisions concerning systemic risk management, and committing to not retaliating against employees who disclose concerns about systemic risks to oversight authorities. === Documentation and transparency === Signatories commit to creating two types of documentation: Safety and Security Frame
Read more →