Joshua Blue is a project under development by IBM that focuses on advancing the artificial intelligence field by designing and programming computers to emulate human mental functions. == Goals == According to researchers at IBM's Thomas J. Watson Research Center, the main goal of Joshua Blue is "to achieve cognitive flexibility that approaches human functioning". In short, IBM is aiming to design Joshua Blue to 'think like a human', mainly in terms of emotional thought. == How it will work == A model of Joshua Blue's learning pattern has been created. Similar to how young children learn human traits through interacting with their surroundings, Joshua Blue will acquire knowledge through external stimuli present in its environment. IBM believes that if computers evolve to learn in this way and then comprehend and analyze the knowledge gained using reason, computers could begin to possess a "mind", of sorts, capable of demonstrating complex social behaviors similar to those of humans. Thus far, IBM has revealed that Joshua Blue will be a computer with a network of wires and input nodes that function as a computer nervous system. This nervous system will be used by Joshua Blue to perceive affect or personal emotional feelings. Not only will this network of input nodes help Joshua Blue discover things physically, but it will also allow Joshua Blue to interpret the significance of events. The input nodes, or proprioceptors, will enable Joshua Blue to be aware of things that happen around itself, as well as recognize and attach meaning to the emotional effect produced by interacting with an object in a certain way. In addition, Joshua Blue's proprioceptors will function as pain and pleasure sensors, allowing Joshua Blue to employ a similar "reward and punishment" system that humans use to form behaviors.
Robotics
Robotics is the interdisciplinary study and practice of the design, construction, operation, and use of robots. A roboticist is someone who specializes in robotics. Robotics usually combines four aspects of design work: a power source (e.g. a battery), mechanical construction, a control system (electrical circuits), and software (run by remote control or artificial intelligence). The goal of most robotics is to design machines that can assist humans in various fields, such as agriculture, construction, domestic work, food processing, inventory management, manufacturing, medicine, military, mining, space exploration, and transportation. Robots impact humans by displacing workers. Some expect this to occur at an increasing rate, leading to proposed solutions such as basic income. Robotics is itself a lucrative business that creates careers, especially for postgraduates. Roboticists often aim to create machines that seem to interface naturally with humans. The field is under active research and development, with areas of interest including robot kinematics and quantum robotics. == Design == Robotics usually combines four aspects of design work to create a robot: Power source: Potential energy sources include wired electricity, a battery, and/or petrol. Mechanical construction: A physical form or combination of forms is designed to functionally achieve tasks within a given range of environments. This can include locomotive elements such as wheels and caterpillar tracks, as well as hydraulic limbs and manipulators (e.g. hands). Control system: Electrical circuits (utilizing components such as diodes and transistors) are used to run software, govern motor movement, and read sensors. Software: A program is how a robot decides when or how to do something. Robotic programs can be run by remote control, artificial intelligence (AI), or a hybrid of the two. AI programming is an important part of robotic navigation and human–robot interaction. === Power source === Many different types of batteries can be used as a power source. Most are lead–acid batteries, which are safe and have relatively long shelf lives but are rather heavy compared to silver–cadmium batteries, which are much smaller in volume and much more expensive. Designing a battery-powered robot needs to take into account factors such as safety, cycle lifetime, and weight. Generators, often some type of internal combustion engine, can also be used, but are often mechanically complex and inefficient. Additionally, a tether could connect the robot to a power supply, saving weight and space, but requiring a cumbersome cable. Potential power sources include: Flywheel energy storage Hydraulics Nuclear Organic garbage (through anaerobic digestion) Pneumatics (compressed gases) Solar power === Mechanical construction === Actuators are the "muscles" of a robot, the parts which convert stored energy into movement. The most popular actuators are electric motors that rotate a wheel or gear and linear actuators that control factory robots. Most robots use electric motors—often brushed and brushless DC motors in portable robots or AC motors in industrial robots and computer numerical control machines—especially in systems with lighter loads and where the predominant form of motion is rotational. Meanwhile, linear actuators move in and out and often have quicker direction changes, particularly when large forces are needed, such as with industrial robotics. They are typically powered by oil or compressed air, but can also be powered by electricity, usually via a motor and a leadscrew. The mechanical rack and pinion is common. Recent alternatives to DC motors are piezoelectric motors, including ultrasonic motors, in which tiny piezoceramic elements vibrate many thousands of times per second, causing linear or rotary motion. One type uses the vibration of the piezo elements to step the motor in a circle or a straight line; another type uses the piezo elements to vibrate a nut or drive a screw. The advantages of these motors are nanometer resolution, speed, and force for their size. Series elastic actuation (SEA) relies on introducing intentional elasticity between the motor actuator and the load for robust force control. Due to the resultant lower reflected inertia, series elastic actuation improves safety during robot interactions or collisions. Further, it provides energy efficiency and shock absorption (mechanical filtering) while reducing excessive wear on the transmission and other components. This approach has successfully been employed in various robots, particularly advanced manufacturing robots and walking humanoid robots. The controller design of a series elastic actuator is most often performed within the passivity framework as it ensures the safety of interaction with unstructured environments. However, this framework suffers from stringent limitations imposed on the controller, which may impact performance. Pneumatic artificial muscles, also known as air muscles, are special tubes that expand (typically up to 42%) when air is forced inside them; they are used in some robot applications. Muscle wire, also known as shape memory alloy, is a material that contracts (under 5%) when electricity is applied; they have been used for some small robots. Electroactive polymers are a plastic material that can contract substantially (up to 380% activation strain) from electricity and have been used in the facial muscles and arms of humanoid robots, as well as to enable new robots to float, fly, swim or walk. Additionally, elastic carbon nanotubes are a promising experimental artificial muscle technology. The absence of defects in carbon nanotubes enables these filaments to deform elastically by several percent, with energy storage levels of perhaps 10 J/cm3 for metal nanotubes. Human biceps could be replaced with wire of this material measuring 8 millimetres (3⁄8 in) in diameter, feasibly allowing future robots to outperform humans. ==== Locomotion ==== Robots with only one or two wheel(s) can have advantages such as greater efficiency, reduced parts, and navigation through confined areas. A one-wheeled robot balances on a round ball; Carnegie Mellon University's Ballbot is the approximate height and width of a person. Several attempts have also been made to build spherical robots (also known as orb bots or ball bots), which move by spinning a weight inside the ball or rotating outer shells. Two-wheeled balancing robots generally use a gyroscope to detect how much a robot is falling and drive the wheels proportionally up to hundreds of times per second to counterbalance the fall, based on inverted pendulum dynamics. NASA's Robonaut has been mounted to a Segway for a similar effect. Most mobile robots have four wheels or continuous tracks. Six wheels can give better traction in outdoor terrain, while tracks provide even more grip. Tracked wheels are common for outdoor off-road robots, but are difficult to use indoors. A small number of skating robots have been developed, one of which is a multimodal walking and skating device with four legs and unpowered wheels. Several robots have been made that can walk on two legs, but not yet as reliably as a human. Many other robots have been built that walk on more than two legs, being significantly easier. Walking robots could be used for uneven terrains, providing a high degree of mobility and efficiency, but two-legged robots can currently only handle flat floors or perhaps stairs. Some approaches have included: The zero moment point (ZMP) is the algorithm used by robots such as Honda's ASIMO. The robot's onboard computer tries to keep the total inertial forces (the combination of Earth's gravity and the acceleration and deceleration of walking) exactly opposed by the floor reaction force (the force of the floor pushing back on the robot's foot). In this way, the two forces cancel out, leaving no moment (force causing the robot to rotate and fall over). Human observers note that this is not exactly how a human walks, with some describing ASIMO's walk as looking like it needs use the bathroom. ASIMO's walking algorithm utilizes some dynamic balancing, but requires a flat surface. Several robots, built in the 1980s by Marc Raibert at the MIT Leg Laboratory, successfully demonstrated very dynamic walking. Initially, a robot with only one leg, and a very small foot could stay upright simply by hopping. The movement is the same as that of a person on a pogo stick. As the robot falls to one side, it would jump slightly in that direction to catch itself. Soon, the algorithm was generalized to two and four legs. A bipedal robot was demonstrated running and even performing somersaults. A quadruped was also demonstrated which could trot, run, pace, and bound. A more advanced approach is a dynamic balancing algorithm, which constantly monitors the robot's motion and places the feet to maintain stability. This technique has been demonstrated by Anybots' Dexter robot (
Condensation algorithm
The condensation algorithm (Conditional Density Propagation) is a computer vision algorithm. The principal application is to detect and track the contour of objects moving in a cluttered environment. Object tracking is one of the more basic and difficult aspects of computer vision and is generally a prerequisite to object recognition. Being able to identify which pixels in an image make up the contour of an object is a non-trivial problem. Condensation is a probabilistic algorithm that attempts to solve this problem. The algorithm itself is described in detail by Isard and Blake in a publication in the International Journal of Computer Vision in 1998. One of the most interesting facets of the algorithm is that it does not compute on every pixel of the image. Rather, pixels to process are chosen at random, and only a subset of the pixels end up being processed. Multiple hypotheses about what is moving are supported naturally by the probabilistic nature of the approach. The evaluation functions come largely from previous work in the area and include many standard statistical approaches. The original part of this work is the application of particle filter estimation techniques. The algorithm's creation was inspired by the inability of Kalman filtering to perform object tracking well in the presence of significant background clutter. The presence of clutter tends to produce probability distributions for the object state which are multi-modal and therefore poorly modeled by the Kalman filter. The condensation algorithm in its most general form requires no assumptions about the probability distributions of the object or measurements. == Algorithm overview == The condensation algorithm seeks to solve the problem of estimating the conformation of an object described by a vector x t {\displaystyle \mathbf {x_{t}} } at time t {\displaystyle t} , given observations z 1 , . . . , z t {\displaystyle \mathbf {z_{1},...,z_{t}} } of the detected features in the images up to and including the current time. The algorithm outputs an estimate to the state conditional probability density p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} by applying a nonlinear filter based on factored sampling and can be thought of as a development of a Monte-Carlo method. p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} is a representation of the probability of possible conformations for the objects based on previous conformations and measurements. The condensation algorithm is a generative model since it models the joint distribution of the object and the observer. The conditional density of the object at the current time p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} is estimated as a weighted, time-indexed sample set { s t ( n ) , n = 1 , . . . , N } {\displaystyle \{s_{t}^{(n)},n=1,...,N\}} with weights π t ( n ) {\displaystyle \pi _{t}^{(n)}} . N is a parameter determining the number of sample sets chosen. A realization of p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} is obtained by sampling with replacement from the set s t {\displaystyle s_{t}} with probability equal to the corresponding element of π t {\displaystyle \pi _{t}} . The assumptions that object dynamics form a temporal Markov chain and that observations are independent of each other and the dynamics facilitate the implementation of the condensation algorithm. The first assumption allows the dynamics of the object to be entirely determined by the conditional density p ( x t | x t − 1 ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {x_{t-1}} )} . The model of the system dynamics determined by p ( x t | x t − 1 ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {x_{t-1}} )} must also be selected for the algorithm, and generally includes both deterministic and stochastic dynamics. The algorithm can be summarized by initialization at time t = 0 {\displaystyle t=0} and three steps at each time t: === Initialization === Form the initial sample set and weights by sampling according to the prior distribution. For example, specify as Gaussian and set the weights equal to each other. === Iterative procedure === Sample with replacement N {\displaystyle N} times from the set { s 0 ( n ) , n = 1 , . . . , N } {\displaystyle \{s_{0}^{(n)},n=1,...,N\}} with probability { π 0 ( n ) , n = 1 , . . . , N } {\displaystyle \{\pi _{0}^{(n)},n=1,...,N\}} to generate a realization of p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} . Apply the learned dynamics p ( x t | x t − 1 ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {x_{t-1}} )} to each element of this new set, to generate a new set { s t ( n ) } {\displaystyle \{s_{t}^{(n)}\}} . To take into account the current observation z t {\displaystyle \mathbf {z_{t}} } , set π t ( n ) = p ( z t | s ( n ) ) ∑ j = 1 N p ( z t | s ( j ) ) {\displaystyle \pi _{t}^{(n)}={\frac {p(\mathbf {z_{t}} |s^{(n)})}{\sum _{j=1}^{N}p(\mathbf {z_{t}} |s^{(j)})}}} for each element { s t ( n ) } {\displaystyle \{s_{t}^{(n)}\}} . This algorithm outputs the probability distribution p ( x t | z 1 , . . . , z t ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {z_{1},...,z_{t}} )} which can be directly used to calculate the mean position of the tracked object, as well as the other moments of the tracked object. Cumulative weights can instead be used to achieve a more efficient sampling. == Implementation considerations == Since object-tracking can be a real-time objective, consideration of algorithm efficiency becomes important. The condensation algorithm is relatively simple when compared to the computational intensity of the Ricatti equation required for Kalman filtering. The parameter N {\displaystyle N} , which determines the number of samples in the sample set, will clearly hold a trade-off in efficiency versus performance. One way to increase efficiency of the algorithm is by selecting a low degree of freedom model for representing the shape of the object. The model used by Isard 1998 is a linear parameterization of B-splines in which the splines are limited to certain configurations. Suitable configurations were found by analytically determining combinations of contours from multiple views, of the object in different poses, and through principal component analysis (PCA) on the deforming object. Isard and Blake model the object dynamics p ( x t | x t − 1 ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {x_{t-1}} )} as a second order difference equation with deterministic and stochastic components: p ( x t | x t − 1 ) ∝ e − 1 2 | | B − 1 ( ( x t − x ¯ ) − A ( x t − 1 − x ¯ ) ) | | 2 ) {\displaystyle p(\mathbf {x_{t}} |\mathbf {x_{t-1}} )\propto e^{-{\frac {1}{2}}||B^{-1}((\mathbf {x_{t}} -\mathbf {\bar {x}} )-A(\mathbf {x_{t-1}} -\mathbf {\bar {x}} ))||^{2})}} where x ¯ {\displaystyle \mathbf {\bar {x}} } is the mean value of the state, and A {\displaystyle A} , B {\displaystyle B} are matrices representing the deterministic and stochastic components of the dynamical model respectively. A {\displaystyle A} , B {\displaystyle B} , and x ¯ {\displaystyle \mathbf {\bar {x}} } are estimated via Maximum Likelihood Estimation while the object performs typical movements. The observation model p ( z | x ) {\displaystyle p(\mathbf {z} |\mathbf {x} )} cannot be directly estimated from the data, requiring assumptions to be made in order to estimate it. Isard 1998 assumes that the clutter which may make the object not visible is a Poisson random process with spatial density λ {\displaystyle \lambda } and that any true target measurement is unbiased and normally distributed with standard deviation σ {\displaystyle \sigma } . The basic condensation algorithm is used to track a single object in time. It is possible to extend the condensation algorithm using a single probability distribution to describe the likely states of multiple objects to track multiple objects in a scene at the same time. Since clutter can cause the object probability distribution to split into multiple peaks, each peak represents a hypothesis about the object configuration. Smoothing is a statistical technique of conditioning the distribution based on both past and future measurements once the tracking is complete in order to reduce the effects of multiple peaks. Smoothing cannot be directly done in real-time since it requires information of future measurements. == Applications == The algorithm can be used for vision-based robot localization of mobile robots. Instead of tracking the position of an object in the scene, however, the position of the camera platform is tracked. This allows the camera platform to be globally localized given a visual map of the environment. Extensions of the condensation algorithm have also been used to recognize human gestures in image sequences. This application of the condensation algorithm impacts the ran
Perplexity AI
Perplexity AI, Inc., or simply Perplexity, is an American privately held software company offering a web search engine that processes user queries and synthesizes responses. Perplexity products use large language models and incorporate real-time web search capabilities, providing responses based on current Internet content, citing sources used. Its real-time search engine is called Sonar and is based on Meta's Llama model. A free public version is available, while a paid Pro subscription offers access to more advanced language models and additional features. Perplexity AI, Inc., was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski. As of September 2025, the company was valued at US$20 billion. Perplexity AI has attracted legal scrutiny over allegations of copyright infringement, unauthorized content use, and trademark issues from several major media organizations, including the BBC, Dow Jones, and The New York Times. According to separate analyses by Wired and later Cloudflare, Perplexity uses undisclosed web crawlers with spoofed user-agent strings to scrape the content of websites which prohibit, or explicitly block, web scraping. == History == In August 2022, Perplexity AI, Inc., was founded by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski, engineers with backgrounds in back-end systems, artificial intelligence (AI) and machine learning. It launched its main search engine on December 7, 2022, and has since released a Google Chrome extension and apps for iOS and Android. In February 2023, Perplexity reported two million unique visitors. By April 2024, Perplexity had raised $165 million in funding, valuing the company at over $1 billion. As of June 2025, Perplexity closed a $500 million round of funding that elevated its valuation to $14 billion. Investors in Perplexity AI have included Jeff Bezos, Tobias Lütke, Nat Friedman, Nvidia, and Databricks. Perplexity has also received funding from 1789 Capital, a venture capital firm notable for its association with Donald Trump Jr. During Bloomberg’s Tech Summit 2025, Srinivas shared that the company processed 780 million queries in May 2025, experiencing more than 20% month-over-month growth, processing around 30 million queries daily. In July 2024, Perplexity announced the launch of a new publishers' program to share advertising revenue with partners. On January 18, 2025, the day before the impending U.S. ban on the social media app TikTok, Perplexity submitted a proposal for a merger with TikTok US. On August 12, 2025, Perplexity made a bid to buy Chrome from Google for $34.5 billion. Perplexity stated that the sale could remedy anti-trust litigation against Google, in which a judge was considering compelling the sale of Chrome. In December 2025, Cristiano Ronaldo took an undisclosed stake in Perplexity AI and entered a global brand partnership with the company. === Business Strategy and Finance (2026) === As of early 2026, Perplexity AI reached a valuation of $21.21 billion following its Series E-6 funding round. The company's Annual Recurring Revenue (ARR) grew from $80 million in late 2024 to an estimated $200 million by February 2026. In January 2026, the company entered into a three-year, $750 million commitment with Microsoft Azure to secure the GPU capacity required for its advanced "Deep Research" and "Model Council" features. In February 2026, Perplexity transitioned to a subscription-first model by discontinuing its AI-integrated advertising strategy. Leadership stated the move was intended to preserve user trust in the "answer engine," prioritizing objective results over ad revenue. The company also introduced the "Model Council" feature on February 5, 2026, which allows users to compare outputs from multiple large language models, such as GPT-5.2 and Claude 4.6, simultaneously. To expand its user base, Perplexity began offering a free year of Pro access to students, U.S. Military Veterans, and government employees. == Products and services == === Search engine web portal === Perplexity’s primary offering is an online information retrieval system (search engine) that uses large language models to generate responses to user queries by searching and summarizing web-based content. Perplexity offers a feature known as Perplexity Pages that generates structured summaries and report-like content from user queries by aggregating cited sources. Perplexity is available without charge or registration to Web users, a freemium model. === Perplexity Pro === Perplexity Pro is a subscription tier, a more capable paid "enterprise" service, including stronger security and data protection and additional tools, including the ability to search uploaded documents alongside web content and access to a programmatic application programming interface (API). It allows the user to select between backend models such as GPT-5.4, Claude 4.6 and Gemini 3.1 Pro. The company has also developed its own models, Sonar (based on Llama 3.3) and R1 1776 (based on DeepSeek R1). === Internal Knowledge Search === Internal Knowledge Search enables Pro and Enterprise Pro users to simultaneously search across web content and internal documents. Users can upload and search through Excel, Word, PDF, and other common file formats. Enterprise Pro users can upload and index up to 500 files. === Search API === Perplexity's Search API provides AI developers with programmatic access to the company's search infrastructure. The September 2025 release includes a software development kit, an open-source evaluation framework called search_evals, and documentation detailing the API's design and optimization. === Shopping hub === Perplexity's Shopping Hub is an online shopping platform that provides AI-generated product recommendations, and enables users to purchase products directly through Perplexity's interface. It was launched in November 2024 with backing by Amazon and Nvidia. === Finance === In October 2024, Perplexity AI introduced new finance-related features, including looking up stock prices and company earnings data. The tool provides real-time stock quotes and price tracking, industry peer comparisons and basic financial analysis tools. The platform sources its financial data from Financial Modeling Prep. === Assistant === In January 2025, Perplexity launched the Perplexity Assistant, an AI-powered tool designed to enhance the functionality of its search engine. It can perform tasks across multiple apps, such as hailing a ride or searching for a song, and can maintain context across actions. The assistant is also multi-modal, meaning it can use a phone's camera to provide answers about the user's surroundings or on-screen content. Perplexity has acknowledged that the assistant is still in development and may not always function as expected. For instance, certain features, such as summarizing unread emails or upcoming calendar events, require users to enable a workaround based on notifications. === Comet === In July 2025, Perplexity launched Comet, an AI browser based on Chromium. Initially, access to the browser was limited to users subscribed to the most expensive subscription tier. The browser was later released for free download in October 2025. A key feature is integration of the Perplexity search engine, which can perform a variety of tasks such as generating article summaries, describing an image, conducting research about a topic and composing emails. === Truth Social chatbot === Perplexity has been contracted to produce a chatbot for Donald Trump's social media platform Truth Social. == Leadership == Aravind Srinivas is the CEO and co-founder of Perplexity AI. He previously held research positions at OpenAI, Google DeepMind, and other AI research institutions focusing on machine learning and artificial intelligence. In a March 2026 All-In episode, Srinivas said the incoming AI-related layoffs were "glorious future" to "look forward", as it freed people from jobs they didn't like and gave them opportunities to pursue entrepreneurship. == Controversies == === Copyright and trademark infringement allegations === In June 2024, Forbes publicly criticized Perplexity for using their content. According to Forbes, Perplexity published a story largely copied from a proprietary Forbes article without mentioning or prominently citing Forbes. In response, Srinivas said that the feature had some "rough edges" and accepted feedback but maintained that Perplexity only "aggregates" rather than plagiarizes information. In October 2024, The New York Times sent a cease-and-desist notice to Perplexity to stop accessing and using NYT content, claiming that Perplexity is violating its copyright by scraping data from its website. In June 2024, Dow Jones and New York Post filed a lawsuit against Perplexity, alleging copyright infringement. The lawsuit also alleged that Perplexity harmed their brand by attributing hallucinated quotes, for example on F-16 jets for Ukraine, to artic
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
Machine unlearning
Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin
Automated essay scoring
Automated essay scoring (AES) is the use of specialized computer programs to assign grades to essays written in an educational setting. It is a form of educational assessment and an application of natural language processing. Its objective is to classify a large set of textual entities into a small number of discrete categories, corresponding to the possible grades, for example, the numbers 1 to 6. Therefore, it can be considered a problem of statistical classification. Several factors have contributed to a growing interest in AES. Among them are cost, accountability, standards, and technology. Rising education costs have led to pressure to hold the educational system accountable for results by imposing standards. The advance of information technology promises to measure educational achievement at reduced cost. The use of AES for high-stakes testing in education has generated significant backlash, with opponents pointing to research that computers cannot yet grade writing accurately and arguing that their use for such purposes promotes teaching writing in reductive ways (i.e. teaching to the test). == History == Most historical summaries of AES trace the origins of the field to the work of Ellis Batten Page. In 1966, he argued for the possibility of scoring essays by computer, and in 1968 he published his successful work with a program called Project Essay Grade (PEG). Using the technology of that time, computerized essay scoring would not have been cost-effective, so Page abated his efforts for about two decades. Eventually, Page sold PEG to Measurement Incorporated. By 1990, desktop computers had become so powerful and so widespread that AES was a practical possibility. As early as 1982, a UNIX program called Writer's Workbench was able to offer punctuation, spelling and grammar advice. In collaboration with several companies (notably Educational Testing Service), Page updated PEG and ran some successful trials in the early 1990s. Peter Foltz and Thomas Landauer developed a system using a scoring engine called the Intelligent Essay Assessor (IEA). IEA was first used to score essays in 1997 for their undergraduate courses. It is now a product from Pearson Educational Technologies and used for scoring within a number of commercial products and state and national exams. IntelliMetric is Vantage Learning's AES engine. Its development began in 1996. It was first used commercially to score essays in 1998. Educational Testing Service offers "e-rater", an automated essay scoring program. It was first used commercially in February 1999. Jill Burstein was the team leader in its development. ETS's Criterion Online Writing Evaluation Service uses the e-rater engine to provide both scores and targeted feedback. Lawrence Rudner has done some work with Bayesian scoring, and developed a system called BETSY (Bayesian Essay Test Scoring sYstem). Some of his results have been published in print or online, but no commercial system incorporates BETSY as yet. Under the leadership of Howard Mitzel and Sue Lottridge, Pacific Metrics developed a constructed response automated scoring engine, CRASE. Currently utilized by several state departments of education and in a U.S. Department of Education-funded Enhanced Assessment Grant, Pacific Metrics’ technology has been used in large-scale formative and summative assessment environments since 2007. Measurement Inc. acquired the rights to PEG in 2002 and has continued to develop it. In 2012, the Hewlett Foundation sponsored a competition on Kaggle called the Automated Student Assessment Prize (ASAP). 201 challenge participants attempted to predict, using AES, the scores that human raters would give to thousands of essays written to eight different prompts. The intent was to demonstrate that AES can be as reliable as human raters, or more so. The competition also hosted a separate demonstration among nine AES vendors on a subset of the ASAP data. Although the investigators reported that the automated essay scoring was as reliable as human scoring, this claim was not substantiated by any statistical tests because some of the vendors required that no such tests be performed as a precondition for their participation. Moreover, the claim that the Hewlett Study demonstrated that AES can be as reliable as human raters has since been strongly contested, including by Randy E. Bennett, the Norman O. Frederiksen Chair in Assessment Innovation at the Educational Testing Service. Some of the major criticisms of the study have been that five of the eight datasets consisted of paragraphs rather than essays, four of the eight data sets were graded by human readers for content only rather than for writing ability, and that rather than measuring human readers and the AES machines against the "true score", the average of the two readers' scores, the study employed an artificial construct, the "resolved score", which in four datasets consisted of the higher of the two human scores if there was a disagreement. This last practice, in particular, gave the machines an unfair advantage by allowing them to round up for these datasets. In 1966, Page hypothesized that, in the future, the computer-based judge will be better correlated with each human judge than the other human judges are. Despite criticizing the applicability of this approach to essay marking in general, this hypothesis was supported for marking free text answers to short questions, such as those typical of the British GCSE system. Results of supervised learning demonstrate that the automatic systems perform well when marking by different human teachers is in good agreement. Unsupervised clustering of answers showed that excellent papers and weak papers formed well-defined clusters, and the automated marking rule for these clusters worked well, whereas marks given by human teachers for the third cluster ('mixed') can be controversial, and the reliability of any assessment of works from the 'mixed' cluster can often be questioned (both human and computer-based). == Different dimensions of essay quality == According to a recent survey, modern AES systems try to score different dimensions of an essay's quality in order to provide feedback to users. These dimensions include the following items: Grammaticality: following grammar rules Usage: using of prepositions, word usage Mechanics: following rules for spelling, punctuation, capitalization Style: word choice, sentence structure variety Relevance: how relevant of the content to the prompt Organization: how well the essay is structured Development: development of ideas with examples Cohesion: appropriate use of transition phrases Coherence: appropriate transitions between ideas Thesis Clarity: clarity of the thesis Persuasiveness: convincingness of the major argument == Procedure == From the beginning, the basic procedure for AES has been to start with a training set of essays that have been carefully hand-scored. The program evaluates surface features of the text of each essay, such as the total number of words, the number of subordinate clauses, or the ratio of uppercase to lowercase letters—quantities that can be measured without any human insight. It then constructs a mathematical model that relates these quantities to the scores that the essays received. The same model is then applied to calculate scores of new essays. Recently, one such mathematical model was created by Isaac Persing and Vincent Ng. which not only evaluates essays on the above features, but also on their argument strength. It evaluates various features of the essay, such as the agreement level of the author and reasons for the same, adherence to the prompt's topic, locations of argument components (major claim, claim, premise), errors in the arguments, cohesion in the arguments among various other features. In contrast to the other models mentioned above, this model is closer in duplicating human insight while grading essays. Due to the growing popularity of deep neural networks, deep learning approaches have been adopted for automated essay scoring, generally obtaining superior results, often surpassing inter-human agreement levels. The various AES programs differ in what specific surface features they measure, how many essays are required in the training set, and most significantly in the mathematical modeling technique. Early attempts used linear regression. Modern systems may use linear regression or other machine learning techniques often in combination with other statistical techniques such as latent semantic analysis and Bayesian inference. The automated essay scoring task has also been studied in the cross-domain setting using machine learning models, where the models are trained on essays written for one prompt (topic) and tested on essays written for another prompt. Successful approaches in the cross-domain scenario are based on deep neural networks or models that combine deep and shallow features. == Criteria for success == Any method of a