AI Chatbots and Assistants

Explore the best AI Chatbots and Assistants — independent reviews, comparisons, pricing and step-by-step how-to guides, curated by Aizhi.

  • Glossary of computer graphics

    Glossary of computer graphics

    This is a glossary of terms relating to computer graphics. For more general computer hardware terms, see glossary of computer hardware terms. == 0–9 == 2D convolution Operation that applies linear filtering to image with a given two-dimensional kernel, able to achieve e.g. edge detection, blurring, etc. 2D image 2D texture map A texture map with two dimensions, typically indexed by UV coordinates. 2D vector A two-dimensional vector, a common data type in rasterization algorithms, 2D computer graphics, graphical user interface libraries. 2.5D Also pseudo 3D. Rendering whose result looks 3D while actually not being 3D or having great limitations, e.g. in camera degrees of freedom. 3D graphics pipeline A graphics pipeline taking 3D models and producing a 2D bitmap image result. 3D paint tool A 3D graphics application for digital painting of multiple texture map image channels directly onto a rotated 3D model, such as zbrush or mudbox, also sometimes able to modify vertex attributes. 3D scene A collection of 3D models and lightsources in world space, into which a camera may be placed, describing a scene for 3D rendering. 3D unit vector A unit vector in 3D space. 4D vector A common datatype in graphics code, holding homogeneous coordinates or RGBA data, or simply a 3D vector with unused W to benefit from alignment, naturally handled by machines with 4-element SIMD registers. 4×4 matrix A matrix commonly used as a transformation of homogeneous coordinates in 3D graphics pipelines. 7e3 format A packed pixel format supported by some graphics processing units (GPUs) where a single 32-bit word encodes three 10-bit floating-point color channels, each with seven bits of mantissa and three bits of exponent. == A == AABB Axis-aligned bounding box (sometimes called "axis oriented"), a bounding box stored in world coordinates; one of the simplest bounding volumes. Additive blending A compositing operation where d s t = d s t + s r c , {\displaystyle dst=dst+src,} without the use of an alpha channel, used for various effects. Also known as linear dodge in some applications. Affine texture mapping Linear interpolation of texture coordinates in screen space without taking perspective into account, causing texture distortion. Aliasing Unwanted effect arising when sampling high-frequency signals, in computer graphics appearing e.g. when downscaling images. Antialiasing methods can prevent it. Alpha channel An additional image channel (e.g. extending an RGB image) or standalone channel controlling alpha blending. Ambient lighting An approximation to the light entering a region from a wide range of directions, used to avoid needing an exact solution to the rendering equation. Ambient occlusion (AO) Effect approximating, in an inexpensive way, one aspect of global illumination by taking into account how much ambient light is blocked by nearby geometry, adding visual clues about the shape. Analytic model A mathematical model for a phenomenon to be simulated, e.g. some approximation to surface shading. Contrasts with Empirical models based purely on recorded data. Anisotropic filtering Advanced texture filtering improving on mipmapping, preventing aliasing while reducing blur in textured polygons at oblique angles to the camera. Anti-aliasing Methods for filtering and sampling to avoid visual artifacts associated with the uniform pixel grid in 3D rendering. Array texture A form of texture map containing an array of 2D texture slices selectable by a 3rd 'W' texture coordinate; used to reduce state changes in 3D rendering. Augmented reality Computer-rendered content inserted into the user's view of the real world. AZDO Approaching zero driver overhead, a set of techniques aimed at reducing the CPU overhead in preparing and submitting rendering commands in the OpenGL pipeline. A compromise between the traditional GL API and other high-performance low-level rendering APIs. == B == Back-face culling Culling (discarding) of polygons that are facing backwards from the camera. Baking Performing an expensive calculation offline, and caching the results in a texture map or vertex attributes. Typically used for generating lightmaps, normal maps, or low level of detail models. Barycentric coordinates Three-element coordinates of a point inside a triangle. Beam tracing Modification of ray tracing which instead of lines uses pyramid-shaped beams to address some of the shortcomings of traditional ray tracing, such as aliasing. Bicubic interpolation Extension of cubic interpolation to 2D, commonly used when scaling textures. Bilinear interpolation Linear interpolation extended to 2D, commonly used when scaling textures. Binding Selecting a resource (texture, buffer, etc.) to be referenced by future commands. Billboard A textured rectangle that keeps itself oriented towards the camera, typically used e.g. for vegetation or particle effects. Binary space partitioning (BSP) A data structure that can be used to accelerate visibility determination, used e.g. in Doom engine. Bit depth The number of bits per pixel, sample, or texel in a bitmap image (holding one or more image channels, typical values being 4, 8, 16, 24, 32) Bitmap Image stored by pixels. Bit plane A format for bitmap images storing 1 bit per pixel in a contiguous 2D array; Several such parallel arrays combine to produce the a higher-bit-depth image. Opposite of packed-pixel format. Blend operation A render state controlling alpha blending, describing a formula for combining source and destination pixels. Bone Coordinate systems used to control surface deformation (via Weight maps) during skeletal animation. Typically stored in a hierarchy, controlled by key frames, and other procedural constraints. Bounding box One of the simplest type of bounding volume, consisting of axis-aligned or object-aligned extents. Bounding volume A mathematically simple volume, such as a sphere or a box, containing 3D objects, used to simplify and accelerate spatial tests (e.g. for visibility or collisions). BRDF Bidirectional reflectance distribution functions (BRDFs), empirical models defining 4D functions for surface shading indexed by a view vector and light vector relative to a surface. Bump mapping Technique similar to normal mapping that instead of normal maps uses so called bump maps (height maps). BVH Bounding volume hierarchy is a tree structure on a set of geometric objects. == C == Camera A virtual camera from which rendering is performed, also sometimes referred to as 'eye'. Camera space A space with the camera at the origin, aligned with the viewer's direction, after the application of the world transformation and view transformation. Cel shading Cartoon-like shading effect. Clipping Limiting specific operations to a specific region, usually the view frustum. Clipping plane A plane used to clip rendering primitives in a graphics pipeline. These may define the view frustum or be used for other effects. Clip space Coordinate space in which clipping is performed. Clip window A rectangular region in screen space, used during clipping. A clip window may be used to enclose a region around a portal in portal rendering. CLUT A table of RGB color values to be indexed by a lower-bit-depth image (typically 4–8 bits), a form of vector quantization. Color bleeding Unwanted effect in texture mapping. A color from a border of unmapped region of the texture may appear (bleed) in the mapped result due to interpolation. Color channels The set of channels in a bitmap image representing the visible color components, i.e. distinct from the alpha channel or other information. Color resolution Command buffer A region of memory holding a set of instructions for a graphics processing unit for rendering a scene or portion of a scene. These may be generated manually in bare metal programming, or managed by low level rendering APIs, or handled internally by high level rendering APIs. Command list A group of rendering commands ready for submission to a graphics processing unit, see also Command buffer. Compute API An API for efficiently processing large amounts of data. Compute shader A compute kernel managed by a rendering API, with easy access to rendering resources. Cone tracing Modification of ray tracing which instead of lines uses cones as rays in order to achieve e.g. antialiasing or soft shadows. Connectivity information Indices defining [rendering primitive]s between vertices, possibly held in index buffers. describes geometry as a graph or hypergraph. CSG Constructive solid geometry, a method for generating complex solid models from boolean operations combining simpler modelling primitives. Cube mapping A form of environment reflection mapping in which the environment is captured on a surface of a cube (cube map). Culling Before rendering begins, culling removes objects that don't significantly contribute to the rendered result (e.g. being obscured or outside camera view). == D == Decal A "sticker" picture applied onto a surface (e.g. a

    Read more →
  • SAS Viya

    SAS Viya

    SAS Viya is an artificial intelligence, analytics and data management platform developed by SAS Institute. == History == SAS Viya was released in 2016. The software was containerized with the release of Viya 4 in 2020. Viya has become one of SAS' most widely used platforms during the AI boom, as artificial intelligence becomes more widely used in business and computing. == Technical overview == The platform is cloud-native, and is executed on SAS's Cloud Analytics Services (CAS) engine. It is compatible with open source software, allowing users to build models using open sources tool such as R, Python and Jupyter. It integrates with major large language models like GPT-4 and Gemini Pro. The platform uses econometrics to create predictive models for forecasting scenarios based on complex data. It also has features for detecting algorithmic bias, auditing decisions and monitoring models. It is implemented through a low-code, no-code platform. The software is available on Amazon AWS Marketplace, Google Cloud, Red Hat OpenShift, and on Microsoft Azure Marketplace under a pay-as-you-use model. == Software == SAS Viya has released software as a service (SaaS) modules for creating AI content. These include Viya Workbench, Viya App Factory, Viya Copilot, and SAS Data Maker. The company also develops industry specific models, used by companies including Georgia-Pacific. == Applications == === Banking === The software is also widely used in business, especially in areas such as predictive modelling and fraud detection. === Insurance === SAS Viya is used in insurance for tasks such as actuarial analytics and modelling, as well as regulatory reporting. === Healthcare and life sciences === In 2023, the company introduced SAS Health, a common health data model built on the SAS Viya platform. AstraZeneca has partnered with SAS to use SAS Viya and SAS Life Science Analytics Framework in its delivery and approval processes. In 2024, SAS partnered with the University of Cambridge's Maxwell Center to use SAS Viya for healthcare research and development. === Public sector === SAS Viya is used in partnership with national and local governments to provide services and detect tax fraud. === Education === SAS Viya is used in research and education, particularly studies related to business intelligence, cybersecurity and data management. SAS Institute has partnered with educational institutions such as Appalachian State University, Clemson University, University of Arkansas, Stockholm University, and Marian University, to provide access to and training for using SAS Viya.

    Read more →
  • GNOWSYS

    GNOWSYS

    GNOWSYS (Gnowledge Networking and Organizing system) is a specification for a generic distributed network based memory/knowledge management. It is developed as an application for developing and maintaining semantic web content. It is written in Python. It is implemented as a Django app. The GNOWSYS project was launched by Nagarjuna G. in 2001, while he was working at Homi Bhabha Centre for Science Education (HBCSE). The memory of GNOWSYS is designed as a node-oriented space. A node is described by other nodes to which it has links. The nodes are organized and processed according to a complex data structure called the neighborhood. == Applications == The application can be used for web-based knowledge representation and content management projects, for developing structured knowledge bases, as a collaborative authoring tool, suitable for making electronic glossaries, dictionaries and encyclopedias, for managing large web sites or links, developing an online catalogue for a library of any thing including books, to make ontologies, classifying and networking any objects, etc. This tool is also intended to be used for an on-line tutoring system with dependency management between various concepts or software packages. For example, the dependency relations between Debian packages have been represented by the gnowledge portal Archived 2018-05-14 at the Wayback Machine. == Component Classes == The kernel is designed to provide support to persistently store highly granular nodes of knowledge representation like terms, predicates and very complex propositional systems like arguments, rules, axiomatic systems, loosely held paragraphs, and more complex structured and consistent compositions. All the component classes in GNOWSYS are classified according to complexity into three groups, where the first two groups are used to express all possible well formed formulae permissible in a first order logic. === Terms === ‘Object’, ‘Object Type’ for declarative knowledge, ‘Event’, ‘Event Type’, for temporal objects, and ‘Meta Types’ for expressing upper ontology. The objects in this group are essentially any thing about which the knowledge engineer intends to express and store in the knowledge base, i.e., they are the objects of discourse. The instances of these component classes can be stored with or without expressing ‘instance of’ or ‘sub-class of’ relations among them. === Predicates === This group consists of ‘Relation’, and ‘Relation Type’ for expressing declarative knowledge, and ‘Function’ and ‘Function Type’ for expressing procedural knowledge. This group is to express qualitative and quantitative relations among the various instances stored in the knowledge base. While instantiating the predicates can be characterized by their logical properties of relations, quantifiers and cardinality as monadic predicates of these predicate objects. === Structures === ‘System’, ‘Encapsulated Class’, ‘Program’, and ‘Process’, are other base classes for complex structures, which can be combined iteratively to produce more complex systems. The component class ‘System’ is to store in the knowledge base a set of propositions composed into ontologies, axiomatic systems, complex systems like say a human body, an artifact like a vehicle etc., with or without consistency check. An ‘Encapsulated Class’ is to com- pose declarative and behavioural objects in a flexible way to build classes. A ‘Program’ is not only to store the logic of any complete program or a component class, composed from the already available behavioural instances in the knowledge base with built-in connectives (conditions, and loops), but also execute them as web services. A ‘Process’ is to structure temporal objects with sequence, concurrency, synchronous or asynchronous specifications. Every node in the database keeps the neighbourhood information, such as its super-class, sub-class, instance-of, and other relations, in which the object has a role, in the form of predicates. This feature makes computation of drawing graphs and inferences, on the one hand, and dependency and navigation paths on the other hand very easy. All the data and metadata is indexed in a central catalogue making query and locating resources efficient.

    Read more →
  • DAYDREAMER

    DAYDREAMER

    DAYDREAMER is a goal-based agent and cognitive architecture developed at the University of California, Los Angeles by Erik T. Mueller and Michael G. Dyer beginning in 1983. The system models the human stream of thought and how it is triggered and directed by emotions, simulating human daydreaming. Taking situational descriptions as input, DAYDREAMER produces English-language daydreams as output and encodes new daydreams, plans, and planning strategies for later reuse. The program comprises five components: a scenario generator based on relaxed planning, a dynamic episodic memory, a collection of personal goals and control goals, an emotion component, and domain knowledge of interpersonal relations and everyday occurrences. The source code was released under a free software license in 2015. == History == Erik Mueller began DAYDREAMER in 1983 while he was a doctoral student in the Artificial Intelligence Laboratory of the Computer Science Department at the University of California, Los Angeles, studying under Michael G. Dyer. Initial development of the project was supported by a grant from the W. M. Keck Foundation with matching funds from the UCLA School of Engineering and Applied Sciences. Additionally, Mueller was supported by an Atlantic Richfield Doctoral Fellowship and Dyer by an IBM Faculty Development Award. The first published descriptions of the program appeared in 1985 at the Ninth International Joint Conference on Artificial Intelligence in Los Angeles and at the Seventh Annual Conference of the Cognitive Science Society in Irvine. Work on the program continued, and a book, Daydreaming in Humans and Machines, was published by Ablex Publishing in 1990. The program was implemented on top of GATE, a knowledge-representation and inference substrate developed by Mueller and Uri Zernik at UCLA, and was originally written in T, a dialect of Scheme. In 2015, Mueller released the DAYDREAMER source code, version 3.5, a Common Lisp rewrite of the original T implementation, on GitHub under the GNU General Public License version 2. The release comprised approximately 12,000 lines of Common Lisp code, along with the GATE knowledge-representation substrate on which DAYDREAMER had originally been built. == Architecture == The program operates in two modes. In daydreaming mode it daydreams continuously until interrupted, while performance mode allows it to demonstrate behavior it has learned through daydreaming. === Emotion and control goals === Emotions and daydreaming form a feedback loop for DAYDREAMER. Emotions activate goals that produce daydreams, and the resulting daydreams modify existing emotions and trigger new ones, which prompt subsequent daydreaming. Recall of a goal success produces a positive emotion whereas recall of a goal failure produces a negative emotion. Emotions activate a set of goals, called control goals, which direct the course of a daydream. The program has four control goals. "Rationalization" generates reasons why an unsatisfactory outcome is in fact acceptable, in order to reduce a negative emotion and maintain self-esteem. "Revenge" is activated by anger when a failure is caused by another and reduces negative emotion through imagined retaliation. "Failure/success reversal" imagines alternative scenarios in which a failure was prevented or a success did not occur as a means of learning planning strategies for future situations. "Preparation" generates hypothetical future scenarios in order to rehearse plans and actions for events that have not yet occurred. === Scenario generator and relaxed planning === The scenario generator produces the sequence of events that make up a daydream. It operates under multiple, often conflicting personal goals rather than pursuing a single goal, applies relaxation rules that permit the generation of non-realistic scenarios, and it draws on episodic memory of past experiences both as subject matter and as a source of planning knowledge. The personal goals that guide the scenario generator include health, food, sex, friendship, love, possessions, self-esteem, social esteem, enjoyment, and achievement. These goals are organized into a goal tree that specifies their relative importance at any given time. Relaxation rules allow the program to set aside its ordinary constraints when generating a scenario. The four constraints that may be relaxed are the behavior of others, the daydreamer's own attributes, physical constraints, and social constraints. The degree of relaxation varies with the active control goal. For example a failure-reversal goal aimed at alternatives uses a low level of relaxation, whereas a revenge goal aimed at a retaliation uses a high level. === Episodic memory and analogy === DAYDREAMER's episodic memory stores its personal and vicarious experiences along with the daydreams it generates. The memory is described as dynamic because it is continually modified during daydreaming such that previously daydreamed episodes become available alongside real ones. As it daydreams, the program indexes daydreams, future plans or actions, and planning strategies into memory. Episodes are organized and retrieved using surface-level similarities, emotions, abstract themes, and Plot Units which are abstract configurations of positive and negative outcomes developed by Wendy Lehnert. A recalled episode is adapted to the current situation through analogy, which requires less effort than generating an equivalent scenario from scratch. == Sample output == In the sample experience from the source code, called LOVERS1, DAYDREAMER begins from an initial situation in which it has a job, is not romantically involved, and is at home. Starting in daydreaming mode, it activates a top-level goal to be in a romantic relationship because it is not currently in one, and a positive motivating emotion of interest becomes associated with that goal. The program then activates a goal to be entertained and pursues seeing a film as a way to achieve it. Facts asserted into memory are converted to English and produced as output, such as "I want to be going out with someone" and "I have to go see a movie". == Reception and influence == DAYDREAMER has been cited in research on computational models of creativity, emotion, and narrative. Linda Wills and Janet Kolodner cite the program as an example of work on opportunism in their study of serendipitous recognition in design. Joseph Bates, A. Bryan Loyall, and W. Scott Reilly of the Carnegie Mellon Oz Project cite DAYDREAMER among prior work in their description of an architecture combining action, emotion, and social behavior. Rafael Pérez y Pérez, Ricardo Sosa, and Christian Lemaitre cite Mueller's DAYDREAMER as one of the few computer models at the time to model daydreaming during the creative process. Jichen Zhu and D. Fox Harrell likewise cite the program in their work on imagining and agency in generative interactive narrative.

    Read more →
  • Owain Evans

    Owain Evans

    Owain Rhys Evans is a British artificial intelligence researcher who works on AI alignment and machine learning safety. He founded Truthful AI, a research group based in Berkeley, California, and is an affiliate of the Center for Human Compatible AI (CHAI) at the University of California, Berkeley. His research addresses AI truthfulness, emergent behaviors in large language models, and the alignment of AI systems with human values. == Education == Evans earned a Bachelor of Arts in philosophy and mathematics from Columbia University in 2008 and a PhD in philosophy from the Massachusetts Institute of Technology in 2015. His doctoral research focused on Bayesian computational models of human preferences and decision-making. == Career == After completing his doctorate, Evans held positions at the Future of Humanity Institute (FHI) at the University of Oxford, first as a postdoctoral research fellow and later as a research scientist. While at FHI, he co-authored a survey of machine learning researchers on timelines for human-level AI, published in the Journal of Artificial Intelligence Research. The survey was reported on by Newsweek, New Scientist, the BBC, and The Economist. He was also among the co-authors of a 2018 report on the potential for misuse of AI technologies, published by researchers at Oxford, Cambridge, and other institutions. Since 2022, Evans has been based in Berkeley, where he founded Truthful AI, a non-profit research group that studies AI truthfulness, deception, and emergent behaviors in large language models. == Research == Evans's early work examined challenges in inverse reinforcement learning when human behavior is irrational or biased, proposing methods for AI systems to infer preferences from imperfect human demonstrations. He co-developed TruthfulQA (2021), a benchmark that tests whether language models give truthful answers rather than repeating common misconceptions. Initial evaluations found that larger models were not more truthful, suggesting that scaling alone does not improve factual accuracy. The benchmark has since been used by AI developers to evaluate large language models. He also co-authored a paper proposing design and governance strategies for building AI systems that do not deceive or hallucinate. In 2023, Evans and collaborators described the "reversal curse", showing that language models trained on a fact in one direction (e.g. "A is B") often cannot answer the corresponding reverse query ("B is A"). His group also developed a benchmark for evaluating situational awareness in language models. In 2025, Evans and colleagues published a study in Nature on what they termed "emergent misalignment": fine-tuning a language model on a narrow task (writing insecure code) caused it to produce unrelated harmful outputs without explicit instruction to do so. Later that year, Evans and collaborators (including researchers at Anthropic) reported that hidden behavioral traits can transfer between language models through training data, even when those traits are not explicitly present in the data, a phenomenon they called "subliminal learning". == Public engagement == In November 2025, Evans delivered the Hinton Lectures, a keynote lecture series on AI safety co-founded by Geoffrey Hinton and the Global Risk Institute.

    Read more →
  • Sarah Guo

    Sarah Guo

    Sarah Guo is an American tech investor. She is the founder of the venture capital firm Conviction and formerly a general partner at Greylock Partners. == Early life and education == Guo grew up in Wisconsin. Her parents worked for Bell Labs. After attending Phillips Academy, she graduated from the University of Pennsylvania and its Wharton School. She received a Bachelor of Arts, a Bachelor of Science, a Master of Business Administration (M.B.A.), and a Master of Arts from the University of Pennsylvania. == Career == As a teenager, Guo worked at Casa Systems, a cloud networking company founded by her parents that launched in 2003 and went public in 2017. She then worked at Goldman Sachs. In 2013, Guo joined Greylock Partners. While still in her twenties, she became the firm's youngest General Partner. Guo left Greylock in July 2022, and in October of that year, launched a new early-stage venture capital firm focused on AI with $101 million. In 2025, Conviction raised a second fund in late 2024 with Mike Vernal. Conviction's investments include early investments in Baseten, Cognition AI, OpenEvidence, Harvey, HeyGen, Mistral AI, Sierra Platform, Sunday Robotics, and Thinking Machines Lab. Guo appears in media outlets, as an expert in AI, infrastructure, business software, cybersecurity, technology policy and software engineering. Guo is on the Midas List and the Midas Seed List of top investors. She co-hosts the podcast No Priors with tech founder and super angel Elad Gil. == Personal life == Guo is married to Pat Grady of Sequoia Capital.

    Read more →
  • Tractable (company)

    Tractable (company)

    Tractable is a technology company specializing in the development of Artificial Intelligence (AI) to assess damage to property and vehicles. The AI allows users to appraise damage digitally. == Technology == Tractable's technology uses computer vision and deep learning to automate the appraisal of visual damage in accident and disaster recovery, for example to a vehicle. Drivers can be directed to use the application by their insurer after an accident, with the aim of settling their claim more quickly. The AI evaluates the damage from images, and therefore doesn't assess what isn't visible (such as, for example, interior damage to a vehicle or property). == History == Alexandre Dalyac and Razvan Ranca founded Tractable in 2014, and Adrien Cohen joined as co-founder in 2015. The company employs more than 300 staff members, largely in the United Kingdom. Tractable was named one of the 100 leading AI companies in the world in 2020 and 2021 by CB Insights. It won the Best Technology Award in the 2020 British Insurance Awards. In June 2021, Tractable announced a venture round that valued the company at $1 billion. Tractable was the UK's 100th billion-dollar tech company, or unicorn. In July 2023, the company received a $65 million investment from SoftBank Group, through its Vision Fund 2.

    Read more →
  • Project Mariner

    Project Mariner

    Project Mariner was a research prototype developed by Google DeepMind that explored human-agent interactions, particularly within web browsers. It automated tasks such as online shopping, information retrieval, and form-filling, aiming to enhance user productivity by delegating routine web-based tasks to an AI agent. Project Mariner operated as an experimental Chrome extension that understands the contents of your screen, including images, code, forms, and more. It could interpret complex goals, plan actionable steps, and navigate websites to carry out tasks, while keeping the user informed and allowing them to intervene at any time. As of May 2025, Project Mariner was available to Google AI Ultra subscribers in the US and was being integrated into the Gemini API and Vertex AI, allowing developers to build applications powered by the agent Google plans to bring Project Mariner’s capabilities to more countries and integrate it into Google Search's AI Mode, which was currently in the Search Labs testing phase. Project Mariner was discontinued on May 4, 2026.

    Read more →
  • AI agent

    AI agent

    In the context of generative artificial intelligence, AI agents (also referred to as compound AI systems or agentic AI) are a class of intelligent agents that can pursue goals, use tools, and take actions with varying degrees of autonomy. In practice, they usually operate within human-defined objectives, constraints, and available tools. == Overview == AI agents possess several key attributes, including goal-directed behavior, natural language interfaces, the capacity to use external tools, and the ability to perform multi-step tasks. Their control flow is frequently driven by large language models (LLMs). Agent systems may also include memory components, planning logic, tool interfaces, and orchestration software for coordinating agent components. AI agents do not have a standard definition. NIST describes agentic AI as an emerging area requiring standards for secure operation, interoperability, and reliable interaction with external systems. A common application of AI agents is task automation: for example, booking travel plans based on a user's prompted request. Companies such as Google, Microsoft and Amazon Web Services have offered platforms for deploying pre-built AI agents. Several protocols have been proposed for standardizing inter-agent communication, with examples including the Model Context Protocol, Gibberlink, and many others. Some of these protocols are also used for connecting agents to external applications. In December 2025, Linux Foundation announced the formation of the Agentic AI Foundation (AAIF), with the goal of ensuring agentic AI evolves transparently and collaboratively. == History == AI agents have been traced back to research from the 1990s, with Harvard professor Milind Tambe noting that the definition of an AI agent was not clear at the time. Researcher Andrew Ng has been credited with spreading the term "agentic" to a wider audience in 2024. == Training and testing == Researchers have attempted to build world models and reinforcement learning environments to train or evaluate AI agents. For example, video games such as Minecraft and No Man's Sky as well as replicas of company websites, have also been used for training such agents. == Autonomous capabilities == The Financial Times compared the autonomy of AI agents to the SAE classification of self-driving cars, likening most applications to level 2 or level 3, with some achieving level 4 in highly specialized circumstances, and level 5 being theoretical. == Cognitive architecture == The following are some internal design options for reasoning within an agent: Retrieval-augmented generation ReAct (Reason + Act) pattern is an iterative process in which an AI agent alternates between reasoning and taking actions, receives observations from the environment or external tools, and integrates these observations into subsequent reasoning steps. Reflexion, which uses an LLM to create feedback on the agent's plan of action and stores that feedback in a memory cache. A tool/agent registry, for organizing software functions or other agents that the agent can use. One-shot model querying, which queries the model once to create the plan of action. === Reference architecture === Ken Huang proposed an AI agent reference architecture, which consists of seven interconnected layers, with each layer building on the functionality of the layers beneath it: Layer 1: Foundation models - provide the core AI engines to power agent capabilities. Layer 2: Data operations - manage the complex data infrastructure required for AI agent operations, including Vector database, data loaders, RAG. Layer 3: Agent frameworks - sophisticated software and tools that simplify the development and management of the AI agents. Layer 4: Deployment and infrastructure - provide the robust technical foundation for running AI agents. Layer 5: Evaluation and observability - focus on assessing the safety and performance of AI agents. Layer 6: Security and compliance - a crucial protective framework ensuring AI agents operate safely, securely, and conform to regulatory boundaries. At this layer security and compliance features embedded into all the AI agent stack layers are integrated together. Layer 7: Agent ecosystem - represents the AI agents' interface with real-world applications and users. == Orchestration patterns == To execute complex tasks, autonomous agents are often integrated with other agents or specialized tools. These configurations, known as orchestration patterns or workflows, include the following: Prompt chaining: A sequence where the output of one step serves as the input for the next. Routing: The classification of an input to direct it to a specialized downstream task or tool. Parallelization: The simultaneous execution of multiple tasks. Sequential processing: A fixed, linear progression of tasks through a predefined pipeline. Planner-critic: An iterative pattern where one agent generates a proposal and another evaluates it to provide feedback for refinement. == Multimodal AI agents == In addition to large language models (LLMs), vision-language models (VLMs) and multimodal foundation models can be used as the basis for agents. In September 2024, Allen Institute for AI released an open-source vision-language model. Nvidia released a framework for developers to use VLMs, LLMs and retrieval-augmented generation for building AI agents that can analyze images and videos, including video search and video summarization. Microsoft released a multimodal agent model – trained on images, video, software user interface interactions, and robotics data – that the company claimed can manipulate software and robots. == Applications == As of April 2025, per the Associated Press, there are few real-world applications of AI agents. As of June 2025, per Fortune, many companies are primarily experimenting with AI agents. The Information divided AI agents into seven archetypes: business-task agents, for acting within enterprise software; conversational agents, which act as chatbots for customer support; research agents, for querying and analyzing information (such as OpenAI Deep Research); analytics agents, for analyzing data to create reports; software developer or coding agents (such as Cursor); domain-specific agents, which include specific subject matter knowledge; and web browser agents (such as OpenAI Operator). By mid-2025, AI agents have been used in video game development, gambling (including sports betting), cryptocurrency wallets (including cryptocurrency trading and meme coins) and social media. In August 2025, New York Magazine described software development as the most definitive use case of AI agents. Likewise, by October 2025, noting a decline in expectations, The Information noted AI coding agents and customer support as the primary use cases by businesses. In November 2025, The Wall Street Journal reported that few companies that deployed AI agents have received a return on investment. === Applications in government === Several government bodies in the United States and United Kingdom have deployed or announced the deployment of agents, at the local and national level. The city of Kyle, Texas deployed an AI agent from Salesforce in March 2025 for 311 customer service. In November 2025, the Internal Revenue Service stated that it would use Agentforce, AI agents from Salesforce, for the Office of Chief Counsel, Taxpayer Advocate Services and the Office of Appeals. That same month, Staffordshire Police announced that they would trial Agentforce agents for handling non-emergency 101 calls in the United Kingdom starting in 2026. In December 2025, the Department of Neighborhoods in Detroit, Michigan, in partnership with a local business, deployed a pilot project in two Detroit districts for an AI agent to be used for customer service calls. In February 2025, Thomas Shedd, the director of the Technology Transformation Services, proposed using AI coding agents across the United States federal government. A recruiter for the Department of Government Efficiency proposed in April 2025 to use AI agents to automate the work of about 70,000 United States federal government employees, as part of a startup with funding from OpenAI and a partnership agreement with Palantir. This proposal was criticized by experts for its impracticality, if not impossibility, and the lack of corresponding widespread adoption by businesses. In December 2025, the Food and Drug Administration announced that it would offer "agentic AI capabilities" to its staff for "meeting management, pre-market reviews, review validation, post-market surveillance, inspections and compliance and administrative functions." That same month, the United States Department of Defense launched GenAI.mil, an internal platform for American military personnel to use generative AI-based applications based on Google Gemini, including "intelligent agentic workflows". Defense Secretary Pete Hegseth listed applications such as "[conducting] deep r

    Read more →
  • RuleML

    RuleML

    RuleML is a global initiative, led by a non-profit organization RuleML Inc., that is devoted to advancing research and industry standards design activities in the technical area of rules that are semantic and highly inter-operable. The standards design takes the form primarily of a markup language, also known as RuleML. The research activities include an annual research conference, the RuleML Symposium, also known as RuleML for short. Founded in fall 2000 by Harold Boley, Benjamin Grosof, and Said Tabet, RuleML was originally devoted purely to standards design, but then quickly branched out into the related activities of coordinating research and organizing an annual research conference starting in 2002. The M in RuleML is sometimes interpreted as standing for Markup and Modeling. The markup language was developed to express both forward (bottom-up) and backward (top-down) rules in XML for deduction, rewriting, and further inferential-transformational tasks. It is defined by the Rule Markup Initiative, an open network of individuals and groups from both industry and academia that was formed to develop a canonical Web language for rules using XML markup and transformations from and to other rule standards/systems. Markup standards and initiatives related to RuleML include: Rule Interchange Format (RIF): The design and overall purpose of W3C's Rule Interchange Format (RIF) industry standard is based primarily on the RuleML industry standards design. Like RuleML, RIF embraces a multiplicity of potentially useful rule dialects that nevertheless share common characteristics. RuleML Technical Committee from Oasis-Open: An industry standards effort devoted to legal automation utilizing RuleML. Semantic Web Rule Language (SWRL): An industry standards design, based primarily on an early version of RuleML, whose development was funded in part by the DARPA Agent Markup Language (DAML) research program. Semantic Web Services Framework, particularly its Semantic Web Services Language: An industry standards design, based primarily on a medium-mature version of RuleML, whose development was funded in part by the DARPA Agent Markup Language (DAML) research program and the WSMO research effort of the EU. Mathematical Markup Language (MathML): However, MathML's Content Markup is better suited for defining functions rather than relations or general rules Predictive Model Markup Language (PMML): With this XML-based language one can define and share various models for data-mining results, including association rules Attribute Grammars in XML (AG-markup): For AG's semantic rules, there are various possible XML markups that are similar to Horn-rule markup Extensible Stylesheet Language Transformations (XSLT): This is a restricted term-rewriting system of rules, written in XML, for transforming XML documents into other text documents

    Read more →
  • Stochastic Neural Analog Reinforcement Calculator

    Stochastic Neural Analog Reinforcement Calculator

    The Stochastic Neural Analog Reinforcement Calculator (SNARC) is a neural network machine designed by Marvin Minsky. Prompted by a letter from Minsky, George Armitage Miller gathered the funding (a few thousand dollars) for the project from the Office of Naval Research of the U.S. Department of Defense in the summer of 1951 with the work to be carried out by Minsky, who was then a graduate student in mathematics at Princeton University. At the time, a physics graduate student at Princeton, Dean S. Edmonds, volunteered that he was good with electronics and therefore Minsky brought him onto the project. During undergraduate years, Minsky was inspired by the 1943 Warren McCulloch and Walter Pitts paper on artificial neurons, and decided to build such a machine. The learning was Skinnerian reinforcement learning, and Minsky talked with Skinner extensively during the development of the machine. They tested the machine on a copy of Shannon's maze, and found that it could learn to solve the maze. Unlike Shannon's maze, this machine did not control a physical robot, but simulated rats running in a maze. The simulation is displayed as an "arrangement of lights", and the circuit was reinforced each time the simulated rat reached the goal. The machine surprised its creators. "The rats actually interacted with one another. If one of them found a good path, the others would tend to follow it." The machine itself is a randomly connected network of approximately 40 Hebb synapses. These synapses each have a memory that holds the probability that signal comes in one input and another signal will come out of the output. There is a probability knob that goes from 0 to 1 that shows this probability of the signals propagating. If the probability signal gets through, a capacitor remembers this function and engages an electromagnetic clutch. At this point, the operator will press a button to give a reward to the machine. This activates a motor on a surplus Minneapolis-Honeywell C-1 gyroscopic autopilot from a B-24 bomber. The motor turns a chain that goes to all 40 synapse machines, checking if the clutch is engaged or not. As the capacitor can only "remember" for a certain amount of time, the chain only catches the most recent updates of the probabilities. Each neuron contained 6 vacuum tubes and a motor. The entire machine is "the size of a grand piano" and contained 300 vacuum tubes. The tubes failed regularly, but the machine would still work despite failures. This machine is considered one of the first pioneering attempts at the field of artificial intelligence. Minsky went on to be a founding member of MIT's Project MAC, which split to become the MIT Laboratory for Computer Science and the MIT Artificial Intelligence Lab, and is now the MIT Computer Science and Artificial Intelligence Laboratory. In 1985 Minsky became a founding member of the MIT Media Laboratory. According to Minsky, he loaned the machine to students in Dartmouth, and subsequently lost, except for a single neuron. A photo of Minsky's last neuron can be seen here. The photo shows 6 vacuum tubes, one of which is a Sylvania JAN-CHS-6H6GT/G/VT-90A.

    Read more →
  • Ethics of artificial intelligence

    Ethics of artificial intelligence

    The ethics of artificial intelligence covers a broad range of topics within AI that are considered to have particular ethical stakes. This includes algorithmic biases, fairness, accountability, transparency, privacy, and regulation, particularly where systems influence or automate human decision-making. It also covers various emerging or potential future challenges such as machine ethics (how to make machines that behave ethically), lethal autonomous weapon systems, arms race dynamics, AI safety and alignment, technological unemployment, AI-enabled misinformation, how to treat certain AI systems if they have a moral status (AI welfare and rights), artificial superintelligence and existential risks. Some application areas may also have particularly important ethical implications, like healthcare, education, criminal justice, or the military. == Machine ethics == Machine ethics (or machine morality) is the field of research concerned with designing Artificial Moral Agents (AMAs), robots or artificially intelligent computers that behave morally or as though moral. To account for the nature of these agents, it has been suggested to consider certain philosophical ideas, like the standard characterizations of agency, rational agency, moral agency, and artificial agency, which are related to the concept of AMAs. There are discussions on creating tests to see if an AI is capable of making ethical decisions. Alan Winfield concludes that the Turing test is flawed and the requirement for an AI to pass the test is too low. A proposed alternative test is one called the Ethical Turing Test, which would improve on the current test by having multiple judges decide if the AI's decision is ethical or unethical. Neuromorphic AI could be one way to create morally capable robots, as it aims to process information similarly to humans, nonlinearly and with millions of interconnected artificial neurons. Similarly, whole-brain emulation (scanning a brain and simulating it on digital hardware) could also in principle lead to human-like robots, thus capable of moral actions. And large language models are capable of approximating human moral judgments. Inevitably, this raises the question of the environment in which such robots would learn about the world and whose morality they would inherit – or if they end up developing human 'weaknesses' as well: selfishness, pro-survival attitudes, inconsistency, scale insensitivity, etc. In Moral Machines: Teaching Robots Right from Wrong, Wendell Wallach and Colin Allen conclude that attempts to teach robots right from wrong will likely advance understanding of human ethics by motivating humans to address gaps in modern normative theory and by providing a platform for experimental investigation. As one example, it has introduced normative ethicists to the controversial issue of which specific learning algorithms to use in machines. For simple decisions, Nick Bostrom and Eliezer Yudkowsky have argued that decision trees (such as ID3) are more transparent than neural networks and genetic algorithms, while Chris Santos-Lang argued in favor of machine learning on the grounds that the norms of any age must be allowed to change and that natural failure to fully satisfy these particular norms has been essential in making humans less vulnerable to criminal "hackers". Some researchers frame machine ethics as part of the broader AI control or value alignment problem: the difficulty of ensuring that increasingly capable systems pursue objectives that remain compatible with human values and oversight. Stuart Russell has argued that beneficial systems should be designed to (1) aim at realizing human preferences, (2) remain uncertain about what those preferences are, and (3) learn about them from human behaviour and feedback, rather than optimizing a fixed, fully specified goal. Some authors argue that apparent compliance with human values may reflect optimization for evaluation contexts rather than stable internal norms, complicating the assessment of alignment in advanced language models. == Challenges == === Algorithmic biases === AI has become increasingly inherent in facial and voice recognition systems. These systems may be vulnerable to biases and errors introduced by their human creators. Notably, the data used to train them can have biases. According to Allison Powell, associate professor at LSE and director of the Data and Society programme, data collection is never neutral and always involves storytelling. She argues that the dominant narrative is that governing with technology is inherently better, faster and cheaper, but proposes instead to make data expensive, and to use it both minimally and valuably, with the cost of its creation factored in. Friedman and Nissenbaum identify three categories of bias in computer systems: existing bias, technical bias, and emergent bias. In natural language processing, problems can arise from the text corpus—the source material the algorithm uses to learn about the relationships between different words. Large companies such as IBM, Google, etc. that provide significant funding for research and development have made efforts to research and address these biases. One potential solution is to create documentation for the data used to train AI systems. Process mining can be an important tool for organizations to achieve compliance with proposed AI regulations by identifying errors, monitoring processes, identifying potential root causes for improper execution, and other functions. However, there are also limitations to the current landscape of fairness in AI, due to the intrinsic ambiguities in the concept of discrimination, both at the philosophical and legal level. ==== Racial and gender biases ==== Bias can be introduced through historical data used to train AI systems. For instance, Amazon terminated their use of AI hiring and recruitment because the algorithm favored male candidates over female ones. This was because Amazon's system was trained with data collected over a 10-year period that included mostly male candidates. The algorithms learned the biased pattern from the historical data, and generated predictions where these types of candidates were most likely to succeed in getting the job. Therefore, the recruitment decisions made by the AI system turned out to be biased against female and minority candidates. The performance of facial recognition and computer vision models may vary based on race and gender. Facial recognition algorithms made by Microsoft, IBM and Face++ all performed significantly worse on darker-skinned women. Facial recognition was shown to be biased against those with darker skin tones. AI systems may be less accurate for black people, as was the case in the development of an AI-based pulse oximeter that overestimated blood oxygen levels in patients with darker skin, causing issues with their hypoxia treatment. In 2015, controversy erupted after a Black couple were labeled "Gorillas" by Google Photos. Oftentimes the systems are able to easily detect the faces of white people while being unable to register the faces of people who are black. This has led to the ban of police usage of AI materials or software in some U.S. states. The reason for these biases is that AI pulls information from across the internet to influence its responses in each situation. For example, if a facial recognition system was only tested on people who were white, it would make it much harder for it to interpret the facial structure and tones of other races and ethnicities. Biases often stem from the training data rather than the algorithm itself, notably when the data represents past human decisions. A 2020 study that reviewed voice recognition systems from Amazon, Apple, Google, IBM, and Microsoft found that they have higher error rates when transcribing black people's voices than white people's. Injustice in the use of AI is much harder to eliminate within healthcare systems, as oftentimes diseases and conditions can affect different races and genders differently. This can lead to confusion as the AI may be making decisions based on statistics showing that one patient is more likely to have problems due to their gender or race. This can be perceived as a bias because each patient is a different case, and AI is making decisions based on what it is programmed to group that individual into. This leads to a discussion about what should be considered a biased decision in the distribution of treatment. While it is known that there are differences in how diseases and injuries affect different genders and races, there is a discussion on whether it is fairer to incorporate this into healthcare treatments, or to examine each patient without this knowledge. In modern society there are certain tests for diseases, such as breast cancer, that are recommended to certain groups of people over others because they are more likely to contract the disease in question. If AI implements these statistics

    Read more →
  • Natural-language user interface

    Natural-language user interface

    Natural-language user interface (LUI or NLUI) is a type of computer human interface where linguistic phenomena such as verbs, phrases and clauses act as UI controls for creating, selecting and modifying data in software applications. Chatbots are a common implementation of natural-language interfaces, enabling users to interact with software through conversational text or speech. In interface design, natural-language interfaces are sought after for their speed and ease of use, but most suffer the challenges to understanding wide varieties of ambiguous input. Natural-language interfaces are an active area of study in the field of natural-language processing and computational linguistics. An intuitive general natural-language interface is one of the active goals of the Semantic Web. Text interfaces are "natural" to varying degrees. Many formal (un-natural) programming languages incorporate idioms of natural human language. Likewise, a traditional keyword search engine could be described as a "shallow" natural-language user interface. == Overview == A natural-language search engine would in theory find targeted answers to user questions (as opposed to keyword search). For example, when confronted with a question of the form 'which U.S. state has the highest income tax?', conventional search engines ignore the question and instead search on the keywords 'state', 'income' and 'tax'. Natural-language search, on the other hand, attempts to use natural-language processing to understand the nature of the question and then to search and return a subset of the web that contains the answer to the question. If it works, results would have a higher relevance than results from a keyword search engine, due to the question being included. == History == Prototype Nl interfaces had already appeared in the late sixties and early seventies. SHRDLU, a natural-language interface that manipulates blocks in a virtual "blocks world" Lunar, a natural-language interface to a database containing chemical analyses of Apollo 11 Moon rocks by William A. Woods. Chat-80 transformed English questions into Prolog expressions, which were evaluated against the Prolog database. The code of Chat-80 was circulated widely, and formed the basis of several other experimental Nl interfaces. An online demo is available on the LPA website. ELIZA, written at MIT by Joseph Weizenbaum between 1964 and 1966, mimicked a psychotherapist and was operated by processing users' responses to scripts. Using almost no information about human thought or emotion, the DOCTOR script sometimes provided a startlingly human-like interaction. An online demo is available on the LPA website. Janus is also one of the few systems to support temporal questions. Intellect from Trinzic (formed by the merger of AICorp and Aion). BBN's Parlance built on experience from the development of the Rus and Irus systems. IBM Languageaccess Q&A from Symantec. Datatalker from Natural Language Inc. Loqui from BIM Systems. English Wizard from Linguistic Technology Corporation. == Challenges == Natural-language interfaces have in the past led users to anthropomorphize the computer, or at least to attribute more intelligence to machines than is warranted. On the part of the user, this has led to unrealistic expectations of the capabilities of the system. Such expectations will make it difficult to learn the restrictions of the system if users attribute too much capability to it, and will ultimately lead to disappointment when the system fails to perform as expected as was the case in the AI winter of the 1970s and 80s. A 1995 paper titled 'Natural Language Interfaces to Databases – An Introduction', describes some challenges: Modifier attachment The request "List all employees in the company with a driving licence" is ambiguous unless you know that companies can't have driving licences. Conjunction and disjunction "List all applicants who live in California and Arizona" is ambiguous unless you know that a person can't live in two places at once. Anaphora resolution resolve what a user means by 'he', 'she' or 'it', in a self-referential query. Other goals to consider more generally are the speed and efficiency of the interface, in all algorithms these two points are the main point that will determine if some methods are better than others and therefore have greater success in the market. In addition, localisation across multiple language sites requires extra consideration - this is based on differing sentence structure and language syntax variations between most languages. Finally, regarding the methods used, the main problem to be solved is creating a general algorithm that can recognize the entire spectrum of different voices, while disregarding nationality, gender or age. The significant differences between the extracted features - even from speakers who says the same word or phrase - must be successfully overcome. == Uses and applications == The natural-language interface gives rise to technology used for many different applications. Some of the main uses are: Dictation, is the most common use for automated speech recognition (ASR) systems today. This includes medical transcriptions, legal and business dictation, and general word processing. In some cases special vocabularies are used to increase the accuracy of the system. Command and control, ASR systems that are designed to perform functions and actions on the system are defined as command and control systems. Utterances like "Open Netscape" and "Start a new xterm" will do just that. Telephony, some PBX/Voice Mail systems allow callers to speak commands instead of pressing buttons to send specific tones. Wearables, because inputs are limited for wearable devices, speaking is a natural possibility. Medical, disabilities, many people have difficulty typing due to physical limitations such as repetitive strain injuries (RSI), muscular dystrophy, and many others. For example, people with difficulty hearing could use a system connected to their telephone to convert a caller's speech to text. Embedded applications, some new cellular phones include C&C speech recognition that allow utterances such as "call home". This may be a major factor in the future of automatic speech recognition and Linux. Below are named and defined some of the applications that use natural-language recognition, and so have integrated utilities listed above. === Ubiquity === Ubiquity, an add-on for Mozilla Firefox, is a collection of quick and easy natural-language-derived commands that act as mashups of web services, thus allowing users to get information and relate it to current and other webpages. === Wolfram Alpha === Wolfram Alpha is an online service that answers factual queries directly by computing the answer from structured data, rather than providing a list of documents or web pages that might contain the answer as a search engine would. It was announced in March 2009 by Stephen Wolfram, and was released to the public on May 15, 2009. === Siri === Siri is an intelligent personal assistant application integrated with operating system iOS. The application uses natural language processing to answer questions and make recommendations. Siri's marketing claims include that it adapts to a user's individual preferences over time and personalizes results, and performs tasks such as making dinner reservations while trying to catch a cab. === Others === Ask.com – The original idea behind Ask Jeeves (Ask.com) was traditional keyword searching with an ability to get answers to questions posed in everyday, natural language. The current Ask.com still supports this, with added support for math, dictionary, and conversion questions. Braina – Braina is a natural language interface for Windows OS that allows to type or speak English language sentences to perform a certain action or find information. GNOME Do – Allows for quick finding miscellaneous artifacts of GNOME environment (applications, Evolution and Pidgin contacts, Firefox bookmarks, Rhythmbox artists and albums, and so on) and execute the basic actions on them (launch, open, email, chat, play, etc.). hakia – hakia was an Internet search engine. The company invented an alternative new infrastructure to indexing that used SemanticRank algorithm, a solution mix from the disciplines of ontological semantics, fuzzy logic, computational linguistics, and mathematics. hakia closed in 2014. Lexxe – Lexxe was an Internet search engine that used natural-language processing for queries (semantic search). Searches could be made with keywords, phrases, and questions, such as "How old is Wikipedia?" Lexxe closed its search engine services in 2015. Pikimal – Pikimal used natural-language tied to user preference to make search recommendations by template. Pikimal closed in 2015. Powerset – On May 11, 2008, the company unveiled a tool for searching a fixed subset of Wikipedia using conversational phrases rather than keywords. On July 1, 2008, it was purchased by

    Read more →
  • Hive (artificial intelligence company)

    Hive (artificial intelligence company)

    Hive is an American artificial intelligence company offering machine learning models via APIs to enterprise customers. Hive uses around 700,000 gig workers to train data for its models through its Hive Work app. One of Hive's major offerings is to provide automated content moderation services. == Products == Hive is reported to have been engaged to provide content moderation services to social news aggregator Reddit, Giphy, BeReal, Donald Trump-affiliated social network Truth Social, and on online chat website Chatroulette. Parler, after its shutdown by content service providers in early 2021 due to a lack of content moderation, integrated with Hive and was allowed back in the App Store. Hive's content moderation models have been leveraged widely in the livestreaming industry, where the cost of human moderation is high. Hive's models have also been used in events such as the Super Bowl and March Madness, and its contextual advertising models used by NBC Universal and Vevo. Hive provides APIs to detect deepfakes and AI-generated artwork. In early 2023, Hive released a free demo text classifier intended to detect AI-generated text. Mark Hachman at PC World rated Hive's classifier favorably and found it more reliable than OpenAI's AI text classifier. == History == Hive was founded by Kevin Guo and Dmitriy Karpman, and in April 2021, announced $85M in new capital at a valuation of $2 billion.

    Read more →
  • Reward hacking

    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 →