A wide-column store (or extensible record store) is a type of NoSQL database. It uses tables, rows, and columns, but unlike a relational database, the names and format of the columns can vary from row to row in the same table. A wide-column store can be interpreted as a two-dimensional key–value store. Google's Bigtable is one of the prototypical examples of a wide-column store. == Wide-column stores versus columnar databases == Wide-column stores such as Bigtable and Apache Cassandra are not column stores in the original sense of the term, since their two-level structures do not use a columnar data layout. In genuine column stores, a columnar data layout is adopted such that each column is stored separately on disk. Wide-column stores do often support the notion of column families that are stored separately. However, each such column family typically contains multiple columns that are used together, similar to traditional relational database tables. Within a given column family, all data is stored in a row-by-row fashion, such that the columns for a given row are stored together, rather than each column being stored separately. Wide-column stores that support column families are also known as column family databases. == Notable examples == Notable wide-column stores include: Apache Accumulo Apache Cassandra Apache HBase Bigtable DataStax Enterprise (uses Apache Cassandra) DataStax Astra DB (uses Apache Cassandra) Hypertable Azure Tables ScyllaDB
Large language model
A large language model (LLM) is a neural network trained on a vast amount of text for natural language processing tasks, especially language generation. LLMs can typically generate, summarize, translate and analyze text in many contexts, and are a foundational technology behind modern chatbots. Biased or inaccurate training data can make an LLM's output less reliable. As of 2026, the most capable LLMs are based on transformer architectures, which, according to the 2017 paper "Attention Is All You Need", can be more efficient and parallelizable than earlier statistical and recurrent neural network models. Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety. == History == Before the emergence of transformer-based models in 2017, some language models were considered large relative to the computational and data constraints of their time. In the early 1990s, IBM's statistical models pioneered word alignment techniques for machine translation, laying the groundwork for corpus-based language modeling. In 2001, a smoothed n-gram model, such as those employing Kneser–Ney smoothing, trained on 300 million words, achieved state-of-the-art perplexity on benchmark tests. During the 2000s, with the rise of widespread internet access, researchers began compiling massive text datasets from the web ("web as corpus") to train statistical language models. Moving beyond n-gram models, researchers started in 2000 to use neural networks as language models. Following the breakthrough of deep neural networks in image classification around 2012, similar architectures were adapted for language tasks. This shift was marked by the development of word embeddings (e.g., Word2Vec by Mikolov in 2013) and sequence-to-sequence (seq2seq) models using LSTM. In 2016, Google transitioned its translation service to neural machine translation (NMT), replacing statistical phrase-based models with deep recurrent neural networks. These early NMT systems used LSTM-based encoder-decoder architectures, as they preceded the invention of transformers. At the 2017 NeurIPS conference, Google researchers introduced the transformer architecture in their landmark paper "Attention Is All You Need". This paper's goal was to improve upon 2014 seq2seq technology, and was based mainly on the attention mechanism developed by Bahdanau et al. in 2014. The following year in 2018, BERT was introduced and quickly became "ubiquitous". Though the original transformer has both encoder and decoder blocks, BERT is an encoder-only model. Academic and research usage of BERT began to decline in 2023, following rapid improvements in the abilities of decoder-only models (such as GPT) to solve tasks via prompting. Although decoder-only GPT-1 was introduced in 2018, it was GPT-2 in 2019 that caught widespread attention because OpenAI claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use. GPT-3 in 2020 went a step further and as of 2025 is available only via API with no offering of downloading the model to execute locally. But it was the consumer-facing chatbot ChatGPT in late 2022 that received extensive media coverage and public attention by 2023. The 2023 GPT-4 was praised for its increased accuracy and as a "holy grail" for its multimodal capabilities. OpenAI did not reveal the high-level architecture and the number of parameters of GPT-4. The release of ChatGPT led to an uptick in LLM usage across several research subfields of computer science, including robotics, software engineering, and societal impact work. In 2024, OpenAI released the reasoning model OpenAI o1, which generates long chains of thought before returning a final answer. Many LLMs with parameter counts comparable to those of OpenAI's GPT series have been developed. Since 2022, weights-available models have been gaining popularity, especially at first with BLOOM and LLaMA, though both have restrictions on usage and deployment. Mistral AI's open-weight models Mistral 7B and Mixtral 8x7B have a more permissive Apache License. In January 2025, DeepSeek released DeepSeek R1, a 671-billion-parameter open-weight model that performs comparably to OpenAI o1 but at a much lower price per token for users. Since 2023, many LLMs have been trained to be multimodal, having the ability to also process or generate other types of data, such as images, audio, or 3D meshes. Open-weight LLMs have become more influential since 2023. Per Vake et al. (2025), community-driven contributions to open-weight models improve their efficiency and performance via collaborative platforms such as Hugging Face. == Dataset preprocessing == === Tokenization === As machine learning algorithms process numbers rather than text, the text must be converted to numbers. In the first step, a vocabulary is decided upon, then integer indices are arbitrarily but uniquely assigned to each vocabulary entry, and finally, an embedding is associated with the integer index. Algorithms include byte-pair encoding (BPE) and WordPiece. There are also special tokens serving as control characters, such as [MASK] for masked-out token (as used in BERT), and [UNK] ("unknown") for characters not appearing in the vocabulary. Also, some special symbols are used to denote special text formatting. For example, "Ġ" denotes a preceding whitespace in RoBERTa and GPT and "##" denotes continuation of a preceding word in BERT. For example, the BPE tokenizer used by the legacy version of GPT-3 would split tokenizer: texts -> series of numerical "tokens" as Tokenization also compresses the datasets. Because LLMs generally require input to be an array that is not jagged, the shorter texts must be "padded" until they match the length of the longest one. ==== Byte-pair encoding ==== As an example, consider a tokenizer based on byte-pair encoding. In the first step, all unique characters (including blanks and punctuation marks) are treated as an initial set of n-grams (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) n-grams that most frequently occur together are then again merged into even lengthier n-gram, until a vocabulary of prescribed size is obtained. After a tokenizer is trained, any text can be tokenized by it, as long as it does not contain characters not appearing in the initial-set of uni-grams. === Dataset cleaning === In the context of training LLMs, datasets are typically cleaned by removing low-quality, duplicated, or toxic data. Cleaned datasets can increase training efficiency and lead to improved downstream performance. A trained LLM can be used to clean datasets for training a further LLM. With the increasing proportion of LLM-generated content on the web, data cleaning in the future may include filtering out such content. LLM-generated content can pose a problem if the content is similar to human text (making filtering difficult) but of lower quality (degrading performance of models trained on it). === Synthetic data === Training of largest language models might need more linguistic data than naturally available, or that the naturally occurring data is of insufficient quality. In these cases, synthetic data might be used. == Training == An LLM is a type of foundation model (large X model) trained on language. LLMs can be trained in different ways. In particular, GPT models are first pretrained to predict the next word on a large amount of data, before being fine-tuned. === Cost === Substantial infrastructure is necessary for training the largest models. The tendency towards larger models is visible in the list of large language models. For example, the training of GPT-2 (i.e. a 1.5-billion-parameter model) in 2019 cost $50,000, while training of the PaLM (i.e. a 540-billion-parameter model) in 2022 cost $8 million, and Megatron-Turing NLG 530B (in 2021) cost around $11 million. The qualifier "large" in "large language model" is inherently vague, as there is no definitive threshold for the number of parameters required to qualify as "large". === Fine-tuning === Before being fine-tuned, most LLMs are next-token predictors. The fine-tuning shapes the LLM's behavior via techniques like reinforcement learning from human feedback (RLHF) or constitutional AI. Instruction fine-tuning is a form of supervised learning used to teach LLMs to follow user instructions. In 2022, OpenAI demonstrated InstructGPT, a version of GPT-3 similarly fine-tuned to follow instructions. Reinforcement learning from human feedback (RLHF) involves training a reward model to predict which text humans prefer. Then, the LLM can be fine-tuned through reinforcement learning to better satisfy this reward model. Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers. == Architecture == LLMs are generally based on the tra
Cognitive Technologies
Cognitive Technologies is a Russian software corporation that develops corporate business applications, AI-based advanced driver assistance systems. Founded in 1993 in Moscow (Russia), the company has offices in Eastern Europe, with R&D Centers in Russia. == History == Cognitive Technologies was founded in 1993 by Olga Uskova and Vladimir Arlazarov. The first employees previously worked in the team that developed the first world computer chess champion "Kaissa". The first programs developed by Cognitive Technologies were optical image and character recognition software – Tiger and CuneiForm. In February 2015 Cognitive Technologies and Kamaz, Russian Dakar Rally-winning truck manufacturer, started working on the self-driving Kamaz truck project. The first field tests took place in June 2015. In 2015 Andrey Chernogorov was appointed CEO of the company. == Products == Cognitive Technologies develops business application software and self-driving vehicle artificial intelligence. The main products are: C-pilot, AI-based ADAS E1 Evfrat – electronic workflow system CognitiveLot – e-purchasing systems == Cooperation with global companies == Under the contract signed between Cognitive Technologies and Hewlett-Packard, all scanners sold in Russia had text recognition software developed by Cognitive Technologies. It was the first contract with HP for an Eastern European company. Afterwards, Cognitive Technologies signed OEM contracts and business agreements with several global IT-companies, including IBM, Canon, Corel, Samsung, Xerox, Brother, Epson, and Olivetti. In 1998 Cognitive Technologies became the first company in Eastern Europe to get the Oracle Complementary Software Provider status. In 2001 Cognitive Technologies sold its Russian language speech corpus to Intel. In 2010 Cognitive Technologies sold its text parsing module to Yandex. The company also signed an agreement with NVIDIA join efforts in the development of intelligent document recognition technologies. == Self-driving car project == The system developed by Cognitive Technologies does not require building smart cities and smart roads equipped with multiple sensors – it works the opposite way, trying to understand the situation on the road like humans do. The system uses a video camera like a driver who uses his eyes, analyzing the information and focusing on the relevant data. For this purpose the system uses a special type of computer vision – foveal computer vision. Only 5–7% of the data gathered by the video cameras and sensors is processed by the system as relevant. The prototype is being tested in Russia on rough roads, on roads without marking, with the goal to prepare the system for work in difficult situations and on bad roads all around the world. == C-Pilot ADAS project == In August 2016 Cognitive Technologies started its own ADAS development project C-Pilot for ground transport control automation. == Self-driving tractors and harvesters project == The experts from Cognitive Technologies claim that the system will track stones, poles, and other obstacles that might be dangerous for the vehicles. This data will enable the engineers to develop an interactive field map, with GPS coordinates for stones and other obstacles. Eventually, this will result in an alteration of the harvester's movement pattern preventing it from running into stones or other objects that may inflict damage. Harvesters will work autonomously on the field, on the territory that is narrowed by radio beacons. == Present international activities == In 2016 Cognitive Technologies has joined the international community OpenPower Foundation, a consortium of open source solutions to developers based on POWER technology from IBM, which includes the world's leading IT map of Google, NVidia, Mellanox, etc. Within the consortium Cognitive Technologies is the initiator of forming of an international working group to develop a single software standard for the self-driving vehicle control. == Awards == In 2016, the leading Russian business newspaper Kommersant, announced that Cognitive Technologies is the TOP-2 Russian software company. TOP-6 Russian software company in 2015 according to Russoft TOP-500 biggest Russian companies according to RBC TOP-2 company of the Russian EDMS market in 2014 according to IDC TOP-20 Russian biggest IT-companies in 2013 according to Cnews Analytics
Nabil Ali
Nabil Ali Mohammed Abd AL Azeez (Arabic:نبيل علي) (3 January 1938 – 27 January 2016) was an Egyptian scientist, writer, and intellectual who worked in the field of natural language processing and computational linguistics. Ali is considered a pioneer of Arabic language computing, making significant innovations in early computational linguistics. == Education and career == Ali earned a bachelor's degree in Aeronautical Engineering in 1960, and a master's degree in 1967. In 1971, he earned a PhD in Aeronautics. From 1961 to 1972 Ali worked as an engineering officer in the Egyptian Air Force, specializing in maintenance and training. In 1972, he shifted focus to computing, and from 1972 to 1977 he worked as a computer manager at Egyptair. While in this position, Ali introduced the first automated reservation system for airlines in the Arab world. He later held various computing positions in Egypt, Kuwait, Europe, Canada and the US. Ali started working for Sakhr Software, an Arabic language technology company, in 1983. From 1985 to 1999, he was vice president of Sakhr's council for Research and Development. As a director of the Multilingual Advanced Systems Foundation and project manager at the Egyptian National Company for Scientific and Technical Information, Ali did extensive research on information culture and artificial intelligence relating to the Arabic language. Over the course of his career, Ali developed more than 20 educational programs relating to computational linguistics. He developed the first Arabic lexical database and the first knowledge base for Arabic poetry, as well as many other pieces of Arabic language software. == Awards == 1994: General Book Authority Award for Best Book (in the field of future studies). 2003: General Book Authority Award for Best Culture Book (in the field of "Challenges of the Information Age"). 2007: General Book Authority "Innovation in Information Technology" Award. 2012: King Faisal International Award, with Professor Ali Helmy Mousa, in the field of computer processing of the Arabic Language. == Works == Arabic Language and Computer (Research study), Dar Localization, 1988. Al Arab and the Information Age, Knowledge World Series No. 184, April 1994. Arab Culture and the Information Age: A Vision for the Future of Arab Culture Discourse, World of Knowledge Series, No. 265 January 2001. The Digital Gap: an Arab Vision for a Knowledge Society (in partnership with Dr. Nadia Hegazy), World of Knowledge Series, No. 318 August 2005. The Arab Mind and the Knowledge Society: Manifestations of the Crisis and Suggestions for Solutions, Part 1, The World of Knowledge Series, No. 369, November 2009. The Arab Mind and the Knowledge Society: Manifestations of the Crisis and Suggestions for Solutions, Part 2, The World of Knowledge Series, No. 370, December 2009. == Tribute == On 3 January 2020, Google Doodle celebrated Nabil Ali Mohamed's 82nd Birthday.
Ofer Dekel (researcher)
Ofer Dekel (Hebrew: עופר דקל) is a computer science researcher in the Machine Learning Department of Microsoft Research. He obtained his PhD in computer science from the Hebrew University of Jerusalem and is an affiliate faculty at the Computer Science & Engineering department at the University of Washington. == Areas of research == Dekel's research topics include machine learning, online prediction, statistical learning theory, and stochastic optimization. He is currently engaged in the application of machine learning techniques in the development of the Bing search engine.
Trazzler
Trazzler is a travel destination app that specializes in unique and local destinations. The initial concept was developed by Adam Rugel and Biz Stone in 2006 at Twitter's original offices under the name "71 miles". More than 10,000 writers and photographers have contributed and more than $350,000 in freelance contracts have been issued as a result of Trazzeler's weekly writing and photography contests. Investors in the company include SV Angel, AOL Founder Steve Case, and the Twitter founders, Evan Williams, Jack Dorsey, and Biz Stone. The company's partners are the City of Chicago, Hawaii Tourism Authority, Fairmont Hotels & Resorts, Salon.com, and Air New Zealand. Trazzler is designed for use on the iOS, Android, and Facebook.
Universal Networking Language
Universal Networking Language (UNL) is a declarative formal language specifically designed to represent semantic data extracted from natural language texts. It can be used as a pivot language in interlingual machine translation systems or as a knowledge representation language in information retrieval applications. == Structure == In UNL, the information conveyed by the natural language is represented sentence by sentence as a hypergraph composed of a set of directed binary labeled links between nodes or hypernodes. As an example, the English sentence "The sky was blue?!" can be represented in UNL as follows: In the example above, sky(icl>natural world) and blue(icl>color), which represent individual concepts, are UW's attributes of an object directed to linking the semantic relation between the two UWs; "@def", "@interrogative", "@past", "@exclamation" and "@entry" are attributes modifying UWs. UWs are expressed in natural language to be humanly readable. They consist of a "headword" (the UW root) and a "constraint list" (the UW suffix between parentheses), where the constraints are used to disambiguate the general concept conveyed by the headword. The set of UWs is organized in the UNL Ontology. Relations are intended to represent semantic links between words in every existing language. They can be ontological (such as "icl" and "iof"), logical (such as "and" and "or"), or thematic (such as "agt" = agent, "ins" = instrument, "tim" = time, "plc" = place, etc.). There are currently 46 relations in the UNL Specs that jointly define the UNL syntax. Within the UNL program, the process of representing natural language sentences in UNL graphs is called UNLization, and the process of generating natural language sentences out of UNL graphs is called NLization. UNLization is intended to be carried out semi-automatically (i.e., by humans with computer aids), and NLization is intended to be carried out automatically. == History == The UNL program started in 1996 as an initiative of the Institute of Advanced Studies (IAS) of the United Nations University (UNU) in Tokyo, Japan. In January 2001, the United Nations University set up an autonomous and non-profit organization, the UNDL Foundation, to be responsible for the development and management of the UNL program. It inherited from the UNU/IAS the mandate of implementing the UNL program. The overall architecture of the UNL System has been developed with a set of basic software and tools. It was recognized by the Patent Cooperation Treaty (PCT) for the "industrial applicability" of the UNL, which was obtained in May 2002 through the World Intellectual Property Organization (WIPO); the UNL acquired the US patents 6,704,700 and 7,107,206.