The W3C Device Description Working Group (DDWG), operating as part of the World Wide Web Consortium (W3C) Mobile Web Initiative (MWI), was chartered to "foster the provision and access to device descriptions that can be used in support of Web-enabled applications that provide an appropriate user experience on mobile devices." Mobile devices exhibit the greatest diversity of capabilities, and therefore present the greatest challenge to content adaptation technologies. The group published several documents, including a list of requirements for an interface to a Device Description Repository (DDR) and a standard interface meeting those requirements. The group was rechartered in 2006 to work in public towards the development of the Application Programming Interface (API) for a DDR. Early in 2007, the group launched a wiki and a blog to add to the public mailing list. The group subsequently published a formal vocabulary of core device properties, and an API called the DDR Simple API, which became a W3C Recommendation in December 2008. The group closed at the end of 2008, but with the intention of maintaining the Web pages, blog and wiki through W3C volunteer effort. == Publications == The DDWG published several W3C Working Group Notes and one W3C Recommendation. A W3C WG Note that articulates what the W3C and other organizations are doing or have already done with regard to device information. This document suggests an environment in which these technologies work together to meet the goals of content adaptation. The completed document was published on 31 October 2007. A W3C WG Note describing the ecosystem surrounding creation, maintenance and use of device descriptions. The completed document was published on 31 October 2007. A W3C WG Note describing a set of requirements for a reference repository of device descriptions. The completed document was published on 17 December 2007. A W3C WG Note describing a process to manage contributions to an initial core vocabulary, identification of key device properties, a formal initial core vocabulary and the identification of a maintainer for the core vocabulary. The details were contained in the Working Group Note describing the DDWG Core Vocabulary published on 14 April 2008. A W3C WG Note defining useful grouping and structure patterns in device descriptions. The Device Description Structures document was published as a Working Draft on 5 December 2008. The intention is that this document will be future input to other W3C groups. A W3C Recommendation defining a language-neutral programming interface to a Device Description Repository. The DDR Simple API was published on 5 December 2008. There is the possibility of future publications on the DDWG wiki describing implementations of the API in various languages, including Java, IDL, WSDL, C# etc. Much of the DDWG's material was developed in public via the DDWG Wiki and through their public mailing lists.
Trello
Trello is a web-based, kanban-style list-making application developed by Atlassian. Created in 2011 by Fog Creek Software, it was spun out to form the basis of a separate company in New York City in 2014 and sold to Atlassian in January 2017. == History == The name Trello is derived from the word trellis, which had been a code name for the project at its early stages. Trello was released at a TechCrunch event by Fog Creek founder Joel Spolsky. In September 2011 Wired magazine named the application one of "The 7 Coolest Startups You Haven't Heard of Yet". Lifehacker said "it makes project collaboration simple and kind of enjoyable". In 2014, it raised US$10.3 million in funding from Index Ventures and Spark Capital. Prior to its acquisition, Trello had sold 22% of its shares to investors, with the remaining shares held by founders Michael Pryor and Joel Spolsky. In May 2016, Trello claimed it had more than 1.1 million daily active users and 14 million total signups. In May 2015, Trello expanded internationally with localized interfaces for Brazil, Germany, and Spain. In 2016 Trello launched the Power-Up platform, allowing 3rd party developers to build and distribute extensions known as Power-Ups to Trello. Initial integrations included Zendesk, SurveyMonkey and Giphy. By January 2022 there were a total of 247 power-ups listed in the Power-Up directory. On 9 January 2017, Atlassian announced its intent to acquire Trello for $425 million. The transaction was made with $360 million in cash and $65 million in shares and options. In December 2018, Trello announced its acquisition of Butler, a company that developed a leading power-up for automating tasks within a Trello board. Trello announced 35 million users in March 2019 and 50 million users in October 2019. In 2020 Craig Jones, then cybersecurity operations director at Sophos, found that the company exposed the personally identifiable information (PII) data of its users, exposed through public Trello boards; the researcher first tweeted about this issue in the year 2018. On 16 January 2024 Trello suffered a data breach containing over 15 million unique email addresses, names and usernames, when the data was posted on a popular hacking forum. The data was obtained by enumerating a publicly accessible resource using email addresses from previous breach corpuses; it was then added on 22 January 2024 to the famous website collecting data breaches "Have I Been Pwned?". == Uses == Users can create task boards with different columns and move the tasks between them. Typically columns include task statuses such as To Do, In Progress, Done. The tool can be used for personal and business purposes including real estate management, software project management, school bulletin boards, lesson planning, accounting, web design, gaming, and law office case management. == Architecture == According to a Fog Creek blog post in January 2012, the client was a thin web layer which downloads the main app, written in CoffeeScript and compiled to minified JavaScript, using Backbone.js, HTML5 .pushState(), and the Mustache templating language. The server was built on top of MongoDB, Node.js and a modified version of Socket.io. == Reception == On 26 January 2017, PC Magazine gave Trello a 3.5 / 5 rating, calling it "flexible" and saying that "you can get rather creative", while noting that "it may require some experimentation to figure out how to best use it for your team and the workload you manage."
OpenVINO
OpenVINO is an open-source software toolkit developed by Intel for optimizing and deploying deep learning models. It supports several popular model formats and categories, such as large language models, computer vision, and generative AI. OpenVINO is optimized for Intel hardware, but offers support for ARM/ARM64 processors. It sees great use in AI Sound Processing drivers when tied with Intel's Gaussian & Neural Accelerator (GNA). Based in C++, it extends API support for C and Python, as well as Node.js (in early preview). OpenVINO is cross-platform and free for use under Apache License 2.0. == Workflow == The simplest OpenVINO usage involves obtaining a model and running it as is. Yet for the best results, a more complete workflow is suggested: obtain a model in one of supported frameworks, convert the model to OpenVINO IR using the OpenVINO Converter tool, optimize the model, using training-time or post-training options provided by OpenVINO's NNCF. execute inference, using OpenVINO Runtime by specifying one of several inference modes. == OpenVINO model format == OpenVINO IR is the default format used to run inference. It is saved as a set of two files, .bin and .xml, containing weights and topology, respectively. It is obtained by converting a model from one of the supported frameworks, using the application's API or a dedicated converter. Models of the supported formats may also be used for inference directly, without prior conversion to OpenVINO IR. Such an approach is more convenient but offers fewer optimization options and lower performance, since the conversion is performed automatically before inference. Some pre-converted models can be found in the Hugging Face repository. The supported model formats are: PyTorch TensorFlow TensorFlow Lite ONNX (including formats that may be serialized to ONNX) PaddlePaddle JAX/Flax == OS support == OpenVINO runs on Windows, Linux and MacOS.
Lumpers and splitters
Lumpers and splitters are opposing factions in any academic discipline that has to place individual examples into rigorously defined categories. The lumper–splitter problem occurs when there is the desire to create classifications and assign examples to them, for example, schools of literature, biological taxa, and so on. A "lumper" is a person who assigns examples broadly, judging that differences are not as important as signature similarities. A "splitter" makes precise definitions, and creates new categories to classify samples that differ in key ways. == Origin of the terms == The earliest known use of these terms was thought to be by Charles Darwin, in a letter to Joseph Dalton Hooker in 1857: "It is good to have hair-splitters & lumpers". But according to research done by the deputy director at NCSE, Glenn Branch, the credit is due to naturalist Edward Newman who wrote in 1845, "The time has arrived for discarding imaginary species, and the duty of doing this is as imperative as the admission of new ones when such are really discovered. The talents described under the respective names of 'hair-splitting' and 'lumping' are unquestionably yielding their power to the mightier power of Truth." They were then introduced more widely by George G. Simpson in his 1945 work The Principles of Classification and a Classification of Mammals. As he put it: splitters make very small units – their critics say that if they can tell two animals apart, they place them in different genera ... and if they cannot tell them apart, they place them in different species. ... Lumpers make large units – their critics say that if a carnivore is neither a dog nor a bear, they call it a cat. A later use can be found in the title of a 1969 paper "On lumpers and splitters ..." by the medical geneticist Victor McKusick. Reference to lumpers and splitters in the humanities appeared in a debate in 1975 between J. H. Hexter and Christopher Hill, in the Times Literary Supplement. It followed from Hexter's detailed review of Hill's book Change and Continuity in Seventeenth Century England, in which Hill developed Max Weber's argument that the rise of capitalism was facilitated by Calvinist Puritanism. Hexter objected to Hill's "mining" of sources to find evidence that supported his theories. Hexter argued that Hill plucked quotations from sources in a way that distorted their meaning. Hexter explained this as a mental habit that he called "lumping". According to him, "lumpers" rejected differences and chose to emphasise similarities. Any evidence that did not fit their arguments was ignored as aberrant. Splitters, by contrast, emphasised differences, and resisted simple schemes. While lumpers consistently tried to create coherent patterns, splitters preferred incoherent complexity. == Usage in various fields == === Biology === The categorisation and naming of a particular species should be regarded as a hypothesis about the evolutionary relationships and distinguishability of that group of organisms. As further information comes to hand, the hypothesis may be confirmed or refuted. Sometimes, especially in the past when communication was more difficult, taxonomists working in isolation have given two distinct names to individual organisms later identified as the same species. When two named species are agreed to be of the same species, the older species name is almost always retained dropping the newer species name honouring a convention known as "priority of nomenclature". This form of lumping is technically called synonymisation. Dividing a taxon into multiple, often new, taxa is called splitting. Taxonomists are often referred to as "lumpers" or "splitters" by their colleagues, depending on their personal approach to recognizing differences or commonalities between organisms. For example, the number of genera used in Pteridophyte Phylogeny Group I (PPG I) has proved controversial. PPG I uses 18 lycophyte and 319 fern genera. The earlier system put forward by Smith et al. (2006) had suggested a range of 274 to 312 genera for ferns alone. By contrast, the system of Christenhusz & Chase (2014) used 5 lycophyte and about 212 fern genera. The number of fern genera was further reduced to 207 in a subsequent publication. Defending PPG I, Schuettpelz et al. (2018) argue that the larger number of genera is a result of "the gradual accumulation of new collections and new data" and hence "a greater appreciation of fern diversity and ... an improved ability to distinguish taxa". They also argue that the number of species per genus in the PPG I system is already higher than in other groups of organisms (about 33 species per genus for ferns as opposed to about 22 species per genus for angiosperms) and that reducing the number of genera as Christenhusz and Chase propose yields the excessive number of about 50 species per genus for ferns. In response, Christenhusz and Chase (2018) argue that the excessive splitting of genera destabilises the usage of names and will lead to greater instability in future, and that the highly split genera have few if any characters that can be used to recognise them, making identification difficult, even to generic level. They further argue that comparing numbers of species per genus in different groups is "fundamentally meaningless". === History === In history, lumpers are those who tend to create broad definitions that cover large periods of time and many disciplines, whereas splitters want to assign names to tight groups of inter-relationships. Lumping tends to create a more and more unwieldy definition, with members having less and less mutually in common. This can lead to definitions which are little more than conventionalities, or groups which join fundamentally different examples. Splitting often leads to "distinctions without difference", ornate and fussy categories, and failure to see underlying similarities. For example, in the arts, "Romantic" can refer specifically to a period of German poetry roughly from 1780 to 1810, but would exclude the later work of Goethe, among other writers. In music it can mean every composer from Hummel through Rachmaninoff, plus many that came after. === Software modelling === Software engineering often proceeds by building models (sometimes known as model-driven architecture). A lumper is keen to generalise, and produces models with a small number of broadly defined objects. A splitter is reluctant to generalise, and produces models with a large number of narrowly defined objects. Conversion between the two styles is not necessarily symmetrical. For example, if error messages in two narrowly defined classes behave in the same way, the classes can be easily combined. But if some messages in a broad class behave differently, every object in the class must be examined before the class can be split. This illustrates the principle that "splits can be lumped more easily than lumps can be split". === Language classification === There is no agreement among historical linguists about what amount of evidence is needed for two languages to be safely classified in the same language family. For this reason, many proposed language families have had lumper–splitter controversies, including Altaic, Pama–Nyungan, Nilo-Saharan, and most of the larger families of the Americas. At a completely different level, the splitting of a mutually intelligible dialect continuum into different languages, or lumping them into one, is also an issue that continually comes up, though the consensus in contemporary linguistics is that there is no completely objective way to settle the question. Splitters regard the comparative method (meaning not comparison in general, but only reconstruction of a common ancestor or protolanguage) as the only valid proof of kinship, and consider genetic relatedness to be the question of interest. American linguists of recent decades tend to be splitters. Lumpers are more willing to admit techniques like mass lexical comparison or lexicostatistics, and mass typological comparison, and to tolerate the uncertainty of whether relationships found by these methods are the result of linguistic divergence (descent from common ancestor) or language convergence (borrowing). Much long-range comparison work has been from Russian linguists belonging to the Moscow School of Comparative Linguistics, most notably Vladislav Illich-Svitych and Sergei Starostin. In the United States, Greenberg and Ruhlen's work has been met with little acceptance from linguists. Earlier American linguists like Morris Swadesh and Edward Sapir also pursued large-scale classifications like Sapir's 1929 scheme for the Americas, accompanied by controversy similar to that today. === Religious studies === Paul F. Bradshaw suggests that the same principles of lumping and splitting apply to the study of early Christian liturgy. Lumpers, who tend to predominate in this field, try to find a single line of successive texts from the apostolic age to the
Capsule neural network
A capsule neural network (CapsNet) is a machine learning system that is a type of artificial neural network (ANN) that can be used to better model hierarchical relationships. The approach is an attempt to more closely mimic biological neural organization. The idea is to add structures called "capsules" to a convolutional neural network (CNN), and to reuse output from several of those capsules to form more stable (with respect to various perturbations) representations for higher capsules. The output is a vector consisting of the probability of an observation, and a pose for that observation. This vector is similar to what is done for example when doing classification with localization in CNNs. Among other benefits, capsnets address the "Picasso problem" in image recognition: images that have all the right parts but that are not in the correct spatial relationship (e.g., in a "face", the positions of the mouth and one eye are switched). For image recognition, capsnets exploit the fact that while viewpoint changes have nonlinear effects at the pixel level, they have linear effects at the part/object level. This can be compared to inverting the rendering of an object of multiple parts. == History == In 2000, Geoffrey Hinton et al. described an imaging system that combined segmentation and recognition into a single inference process using parse trees. So-called credibility networks described the joint distribution over the latent variables and over the possible parse trees. That system proved useful on the MNIST handwritten digit database. A dynamic routing mechanism for capsule networks was introduced by Hinton and his team in 2017. The approach was claimed to reduce error rates on MNIST and to reduce training set sizes. Results were claimed to be considerably better than a CNN on highly overlapped digits. In Hinton's original idea one minicolumn would represent and detect one multidimensional entity. == Transformations == An invariant is an object property that does not change as a result of some transformation. For example, the area of a circle does not change if the circle is shifted to the left. Informally, an equivariant is a property that changes predictably under transformation. For example, the center of a circle moves by the same amount as the circle when shifted. A nonequivariant is a property whose value does not change predictably under a transformation. For example, transforming a circle into an ellipse means that its perimeter can no longer be computed as π times the diameter. In computer vision, the class of an object is expected to be an invariant over many transformations. I.e., a cat is still a cat if it is shifted, turned upside down or shrunken in size. However, many other properties are instead equivariant. The volume of a cat changes when it is scaled. Equivariant properties such as a spatial relationship are captured in a pose, data that describes an object's translation, rotation, scale and reflection. Translation is a change in location in one or more dimensions. Rotation is a change in orientation. Scale is a change in size. Reflection is a mirror image. Unsupervised capsnets learn a global linear manifold between an object and its pose as a matrix of weights. In other words, capsnets can identify an object independent of its pose, rather than having to learn to recognize the object while including its spatial relationships as part of the object. In capsnets, the pose can incorporate properties other than spatial relationships, e.g., color (cats can be of various colors). Multiplying the object by the manifold poses the object (for an object, in space). == Pooling == Capsnets reject the pooling layer strategy of conventional CNNs that reduces the amount of detail to be processed at the next higher layer. Pooling allows a degree of translational invariance (it can recognize the same object in a somewhat different location) and allows a larger number of feature types to be represented. Capsnet proponents argue that pooling: violates biological shape perception in that it has no intrinsic coordinate frame; provides invariance (discarding positional information) instead of equivariance (disentangling that information); ignores the linear manifold that underlies many variations among images; routes statically instead of communicating a potential "find" to the feature that can appreciate it; damages nearby feature detectors, by deleting the information they rely upon. == Capsules == A capsule is a set of neurons that individually activate for various properties of a type of object, such as position, size and hue. Formally, a capsule is a set of neurons that collectively produce an activity vector with one element for each neuron to hold that neuron's instantiation value (e.g., hue). Graphics programs use instantiation value to draw an object. Capsnets attempt to derive these from their input. The probability of the entity's presence in a specific input is the vector's length, while the vector's orientation quantifies the capsule's properties. Artificial neurons traditionally output a scalar, real-valued activation that loosely represents the probability of an observation. Capsnets replace scalar-output feature detectors with vector-output capsules and max-pooling with routing-by-agreement. Because capsules are independent, when multiple capsules agree, the probability of correct detection is much higher. A minimal cluster of two capsules considering a six-dimensional entity would agree within 10% by chance only once in a million trials. As the number of dimensions increase, the likelihood of a chance agreement across a larger cluster with higher dimensions decreases exponentially. Capsules in higher layers take outputs from capsules at lower layers, and accept those whose outputs cluster. A cluster causes the higher capsule to output a high probability of observation that an entity is present and also output a high-dimensional (20-50+) pose. Higher-level capsules ignore outliers, concentrating on clusters. This is similar to the Hough transform, the RHT and RANSAC from classic digital image processing. == Routing by agreement == The outputs from one capsule (child) are routed to capsules in the next layer (parent) according to the child's ability to predict the parents' outputs. Over the course of a few iterations, each parents' outputs may converge with the predictions of some children and diverge from those of others, meaning that that parent is present or absent from the scene. For each possible parent, each child computes a prediction vector by multiplying its output by a weight matrix (trained by backpropagation). Next the output of the parent is computed as the scalar product of a prediction with a coefficient representing the probability that this child belongs to that parent. A child whose predictions are relatively close to the resulting output successively increases the coefficient between that parent and child and decreases it for parents that it matches less well. This increases the contribution that that child makes to that parent, thus increasing the scalar product of the capsule's prediction with the parent's output. After a few iterations, the coefficients strongly connect a parent to its most likely children, indicating that the presence of the children imply the presence of the parent in the scene. The more children whose predictions are close to a parent's output, the more quickly the coefficients grow, driving convergence. The pose of the parent (reflected in its output) progressively becomes compatible with that of its children. The coefficients' initial logits are the log prior probabilities that a child belongs to a parent. The priors can be trained discriminatively along with the weights. The priors depend on the location and type of the child and parent capsules, but not on the current input. At each iteration, the coefficients are adjusted via a "routing" softmax so that they continue to sum to 1 (to express the probability that a given capsule is the parent of a given child.) Softmax amplifies larger values and diminishes smaller values beyond their proportion of the total. Similarly, the probability that a feature is present in the input is exaggerated by a nonlinear "squashing" function that reduces values (smaller ones drastically and larger ones such that they are less than 1). This dynamic routing mechanism provides the necessary deprecation of alternatives ("explaining away") that is needed for segmenting overlapped objects. This learned routing of signals has no clear biological equivalent. Some operations can be found in cortical layers, but they do not seem to relate this technique. === Math/code === The pose vector u i {\textstyle \mathbf {u} _{i}} is rotated and translated by a matrix W i j {\textstyle \mathbf {W} _{ij}} into a vector u ^ j | i {\textstyle \mathbf {\hat {u}} _{j|i}} that predicts the output of the parent capsule. u ^ j | i = W i j u i {\displaystyle \mathbf {
Hallucination (artificial intelligence)
In the field of artificial intelligence (AI), a hallucination or artificial hallucination (also called bullshitting, confabulation, or delusion) is a response generated by AI that contains false or misleading information presented as fact. This term draws a loose analogy with human psychology, where a hallucination typically involves false percepts. For example, a chatbot powered by large language models (LLMs), like ChatGPT, may embed plausible-sounding random falsehoods within its generated content. Detecting and mitigating errors and hallucinations pose significant challenges for practical deployment and reliability of LLMs in high-stakes scenarios, such as chip design, supply chain logistics, and medical diagnostics. Some software engineers and statisticians have criticized the specific term "AI hallucination" for unreasonably anthropomorphizing computers. Symbolic artificial intelligence models generally do not produce hallucinations, unlike large language models. == Term == === Origin === Since the 1980s, the term "hallucination" has been used in computer vision with a positive connotation to describe the process of adding detail to an image. For example, the task of generating high-resolution face images from low-resolution inputs is called face hallucination. The first documented use of the term "hallucination" in this sense is in the PhD thesis of Eric Mjolsness in 1986. A notable work is the face hallucination algorithm by Simon Baker and Takeo Kanade published in 1999. In the 2000s, hallucinations were described in statistical machine translation as a failure mode. Since the 2010s, the term has undergone a semantic shift to signify the generation of factually incorrect or misleading outputs by AI systems in tasks like machine translation and object detection. In 2015, hallucinations were identified in visual semantic role labeling tasks by Saurabh Gupta and Jitendra Malik. In 2015, computer scientist Andrej Karpathy used the term "hallucinated" in a blog post to describe his recurrent neural network (RNN) language model generating an incorrect citation link. In 2017, Google researchers used the term to describe the responses generated by neural machine translation (NMT) models when they are not related to the source text, and in 2018, the term was used in computer vision to describe instances where non-existent objects are erroneously detected because of adversarial attacks. In July 2021, Meta warned during its release of BlenderBot 2 that the system is prone to "hallucinations", which Meta defined as "confident statements that are not true". Following OpenAI's ChatGPT release in beta version in November 2022, some users complained that such chatbots often seem to pointlessly embed plausible-sounding random falsehoods within their generated content. Many news outlets, including The New York Times, started to use the term "hallucinations" to describe these models' frequently incorrect or inconsistent responses. In 2023, the Cambridge dictionary updated its definition of hallucination to include this new sense specific to the field of AI. Some researchers have highlighted a lack of consistency in how the term is used, but also identified several alternative terms in the literature, such as confabulations, fabrications, and factual errors. === Definitions and alternatives === Uses, definitions and characterizations of the term "hallucination" in the context of LLMs include: "a tendency to invent facts in moments of uncertainty" (OpenAI, May 2023) "a model's logical mistakes" (OpenAI, May 2023) "fabricating information entirely, but behaving as if spouting facts" (CNBC, May 2023) "making up information" (The Verge, February 2023) "probability distributions" (in scientific contexts) Journalist Benj Edwards, in Ars Technica, writes that the term "hallucination" is controversial, but that some form of metaphor remains necessary; Edwards suggests "confabulation" as an analogy for processes that involve "creative gap-filling". In July 2024, a White House report on fostering public trust in AI research mentioned hallucinations only in the context of reducing them. Notably, when acknowledging David Baker's Nobel Prize-winning work with AI-generated proteins, the Nobel committee avoided the term entirely, instead referring to "imaginative protein creation". Hicks, Humphries, and Slater, in their article in Ethics and Information Technology, argue that the output of LLMs is "bullshit" under Harry Frankfurt's definition of the term, and that the models are "in an important way indifferent to the truth of their outputs", with true statements only accidentally true, and false ones accidentally false. Some researchers also use the derogatory term "botshit", often referring to uncritical use of AI. === Criticism === In the scientific community, some researchers avoid the term "hallucination", seeing it as potentially misleading. It has been criticized by Usama Fayyad, executive director of the Institute for Experimental Artificial Intelligence at Northeastern University, on the grounds that it misleadingly personifies large language models and is vague. Mary Shaw said, "The current fashion for calling generative AI's errors 'hallucinations' is appalling. It anthropomorphizes the software, and it spins actual errors as somehow being idiosyncratic quirks of the system even when they're objectively incorrect." In Salon, statistician Gary Smith argues that LLMs "do not understand what words mean" and consequently that the term "hallucination" unreasonably anthropomorphizes the machine. Murray Shanahan argues that anthropomorphic framing of LLM capabilities, including terms like "hallucination", encourages users and researchers to attribute cognitive processes to systems that operate through statistical pattern completion, and advocates for more careful linguistic practices when discussing LLM behavior. Kristina Šekrst argues that applying psychological vocabulary to LLM outputs obscures the difference between the appearance of mental properties and their genuine presence. Förster & Skop assert that tech companies use the hallucination metaphor to anthropomorphize models and deflect responsibility for non-factual outputs. Some see the AI outputs not as illusory but as prospective—that is, having some chance of being true, similar to early-stage scientific conjectures. The term has also been criticized for its association with psychedelic drug experiences. == In natural language generation == In natural language generation, there are several reasons why natural language models hallucinate: === Hallucination from data === Hallucinations can stem from incomplete, inaccurate or unrepresentative data sets. === Modeling-related causes === The pre-training of generative pretrained transformers (GPT) involves predicting the next word. It incentivizes GPT models to "give a guess" about what the next word is, even when they lack information. Some researchers take an anthropomorphic perspective and posit that hallucinations arise from a tension between novelty and usefulness. For instance, Amabile and Pratt define human creativity as the production of novel and useful ideas. By extension, a focus on novelty in machine creativity can lead to the production of original but inaccurate responses—that is, falsehoods—whereas a focus on usefulness may result in memorized content lacking originality. By 2022, newspapers such as The New York Times expressed concern that, as the adoption of bots based on large language models continued to grow, unwarranted user confidence in bot output could lead to problems. === Interpretability research === In 2025, interpretability research by Anthropic on the LLM Claude identified internal circuits that cause it to decline to answer questions unless it knows the answer. By default, the circuit is active and the LLM doesn't answer. When the LLM has sufficient information, these circuits are inhibited and the LLM answers the question. Hallucinations were found to occur when this inhibition happens incorrectly, such as when Claude recognizes a name but lacks sufficient information about that person, causing it to generate plausible but untrue responses. === Examples === On 15 November 2022, researchers from Meta AI published Galactica, designed to "store, combine and reason about scientific knowledge". Content generated by Galactica came with the warning: "Outputs may be unreliable! Language Models are prone to hallucinate text." In one case, when asked to draft a paper on creating avatars, Galactica cited a fictitious paper from a real author who works in the relevant area. Meta withdrew Galactica on 17 November due to offensiveness and inaccuracy. OpenAI's ChatGPT, released in beta version to the public on November 30, 2022, was based on the foundation model GPT-3.5 (a revision of GPT-3). Professor Ethan Mollick of Wharton called it an "omniscient, eager-to-please intern who sometimes lies to you". Data scientist Teresa Kuba
Mistral AI
Mistral AI SAS (French: [mistʁal]) is a French artificial intelligence (AI) company, headquartered in Paris. Founded in 2023, it has open-weight large language models (LLMs), with both open-source and proprietary AI models. As of 2025 the company has a valuation of more than US$14 billion. == Namesake == The company is named after the mistral, a powerful, cold wind in southern France, a term which originates from the Occitan language. == History == Mistral AI was established in April 2023 by three French AI researchers, Arthur Mensch, Guillaume Lample and Timothée Lacroix. Mensch, an expert in advanced AI systems, is a former employee of Google DeepMind; Lample and Lacroix, meanwhile, are large-scale AI models specialists who had worked for Meta Platforms. The trio originally met during their studies at École Polytechnique. == Company operation == === Funding === In June 2023, the start-up carried out a first fundraising of €105 million ($117 million) with investors including the American fund Lightspeed Venture Partners, Eric Schmidt, Xavier Niel and JCDecaux. The valuation was then estimated by the Financial Times at €240 million ($267 million). On 10 December 2023, Mistral AI announced that it had raised €385 million ($428 million) as part of its second fundraising. This round of financing involves the Californian fund Andreessen Horowitz, BNP Paribas and the software publisher Salesforce. It was valued at over €2 billion. On 26 February 2024, Microsoft announced an investment of $16 million in Mistral AI. On 16 April 2024, reporting revealed that Mistral was in talks to raise €500 million, a deal that would more than double its current valuation to at least €5 billion. In June 2024, Mistral AI secured a €600 million ($645 million) funding round, increasing its valuation to €5.8 billion ($6.2 billion). Based on valuation, as of June 2024, the company was ranked fourth globally in the AI industry, and first outside the San Francisco Bay Area. In April 2025, Mistral AI announced a €100 million partnership with the shipping company CMA CGM. In August 2025, the Financial Times reported that Mistral was in talks to raise $1 billion at a $10 billion valuation. In September 2025, Bloomberg announced that Mistral AI has secured a €2 billion investment valuing it at €12 billion ($14 billion). This comes after $1.5 billion investment from Dutch company ASML, which owns 11% of Mistral. In February 2026, Mistral acquired Koyeb, a Paris-based AI startup. Later that month, Mistral AI announced a multi-year strategic partnership with Accenture to help enterprises deploy sovereign AI solutions at scale. In March 2026 Mistral raised $830 million in order to build new datacenters near Paris and in Sweden. == Services == On 19 November, 2024, the company announced updates for Le Chat (pronounced /lə ʃa/ in French, like the French word for "cat"). It added the ability to create images, using Black Forest Labs' Flux Pro model. On 6 February 2025, Mistral AI released Le Chat on iOS and Android mobile devices. Mistral AI also introduced a Pro subscription tier, priced at $14.99 per month, which provides access to more advanced models, unlimited messaging, and web browsing. At the end of May 2026, Le Chat was renamed Vibe, and new features were introduced at the same time. == Models == The following table lists the main model versions of Mistral, describing the significant changes included with each version: === Mistral 7B === Mistral AI claimed in the Mistral 7B release blog post that the model outperforms LLaMA 2 13B on all benchmarks tested, and is on par with LLaMA 34B on many benchmarks tested, despite having only 7 billion parameters, a small size compared to its competitors. === Mixtral 8x7B === Mistral AI claimed in 2023 that its model beat both LLaMA 70B, and GPT-3.5 in most benchmarks. In March 2024, research conducted by Patronus AI comparing performance of LLMs on a 100-question test with prompts to generate text from books protected under U.S. copyright law found that OpenAI's GPT-4, Mixtral, Meta AI's LLaMA-2, and Anthropic's Claude 2 generated copyrighted text verbatim in 44%, 22%, 10%, and 8% of responses respectively. === Mistral Small 3.1 === On 17 March 2025, Mistral released Mistral Small 3.1 as a smaller, more efficient model. === Mistral Medium 3 === On 7 May 2025, Mistral AI released Mistral Medium 3. === Magistral Small and Magistral Medium === On 10 June 2025, Mistral AI released their first AI reasoning models: Magistral Small (open-source), and Magistral Medium, models which are purported to have chain-of-thought capabilities. === Mistral Large 3 and Ministral 3 === On 2 December 2025, Mistral AI released Mistral Large 3, a sparse, mixture-of-experts model with 41 billion active parameters and 675 billion total parameters, and Ministral 3, three small, dense models with 3 billion, 7 billion and 14 billion parameters. === Devstral 2 and Devstral Small 2 === On 10 December 2025, Mistral AI released Devstral 2 and Devstral Small 2. Devstral Small 2, a 24B parameter model is claimed to achieve better performance at coding than Qwen 3 Coder Flash model which is a 30B parameter model.