A mashup (computer industry jargon), in web development, is a web page or web application that uses content from more than one source to create a single new service displayed in a single graphical interface. For example, a user could combine the addresses and photographs of their library branches with a Google map to create a map mashup. The term implies easy, fast integration, frequently using open application programming interfaces (open API) and data sources to produce enriched results that were not necessarily the original reason for producing the raw source data. The term mashup originally comes from creating something by combining elements from two or more sources. The main characteristics of a mashup are combination, visualization, and aggregation. It is important to make existing data more useful, for personal and professional use. To be able to permanently access the data of other services, mashups are generally client applications or hosted online. In the past years, more and more Web applications have published APIs that enable software developers to easily integrate data and functions the SOA way, instead of building them by themselves. Mashups can be considered to have an active role in the evolution of social software and Web 2.0. Mashup composition tools are usually simple enough to be used by end-users. They generally do not require programming skills and rather support visual wiring of GUI widgets, services and components together. Therefore, these tools contribute to a new vision of the Web, where users are able to contribute. The term "mashup" is not formally defined by any standard-setting body. == History == The broader context of the history of the Web provides a background for the development of mashups. Under the Web 1.0 model, organizations stored consumer data on portals and updated them regularly. They controlled all the consumer data, and the consumer had to use their products and services to get the information. The advent of Web 2.0 introduced Web standards that were commonly and widely adopted across traditional competitors and which unlocked the consumer data. At the same time, mashups emerged, allowing mixing and matching competitors' APIs to develop new services. The first mashups used mapping services or photo services to combine these services with data of any kind and therefore to produce visualizations of data. In the beginning, most mashups were consumer-based, but recently the mashup is to be seen as an interesting concept useful also to enterprises. Business mashups can combine existing internal data with external services to generate new views on the data. There was also the free Yahoo! Pipes to build mashups for free using the Yahoo! Query Language. == Types of mashup == There are many types of mashup, such as business mashups, consumer mashups, and data mashups. The most common type of mashup is the consumer mashup, aimed at the general public. Business (or enterprise) mashups define applications that combine their own resources, application and data, with other external Web services. They focus data into a single presentation and allow for collaborative action among businesses and developers. This works well for an agile development project, which requires collaboration between the developers and customer (or customer proxy, typically a product manager) for defining and implementing the business requirements. Enterprise mashups are secure, visually rich Web applications that expose actionable information from diverse internal and external information sources. Consumer mashups combine data from multiple public sources in the browser and organize it through a simple browser user interface. (e.g.: Wikipediavision combines Google Map and a Wikipedia API) Data mashups, opposite to the consumer mashups, combine similar types of media and information from multiple sources into a single representation. The combination of all these resources create a new and distinct Web service that was not originally provided by either source. === By API type === Mashups can also be categorized by the basic API type they use but any of these can be combined with each other or embedded into other applications. ==== Data types ==== Indexed data (documents, weblogs, images, videos, shopping articles, jobs ...) used by metasearch engines Cartographic and geographic data: geolocation software, geovisualization Feeds, podcasts: news aggregators ==== Functions ==== Data converters: language translators, speech processing, URL shorteners... Communication: email, instant messaging, notification... Visual data rendering: information visualization, diagrams Security related: electronic payment systems, ID identification... Editors == Mashup enabler == In technology, a mashup enabler is a tool for transforming incompatible IT resources into a form that allows them to be easily combined, in order to create a mashup. Mashup enablers allow powerful techniques and tools (such as mashup platforms) for combining data and services to be applied to new kinds of resources. An example of a mashup enabler is a tool for creating an RSS feed from a spreadsheet (which cannot easily be used to create a mashup). Many mashup editors include mashup enablers, for example, Presto Mashup Connectors, Convertigo Web Integrator or Caspio Bridge. Mashup enablers have also been described as "the service and tool providers, [sic] that make mashups possible". === History === Early mashups were developed manually by enthusiastic programmers. However, as mashups became more popular, companies began creating platforms for building mashups, which allow designers to visually construct mashups by connecting together mashup components. Mashup editors have greatly simplified the creation of mashups, significantly increasing the productivity of mashup developers and even opening mashup development to end-users and non-IT experts. Standard components and connectors enable designers to combine mashup resources in all sorts of complex ways with ease. Mashup platforms, however, have done little to broaden the scope of resources accessible by mashups and have not freed mashups from their reliance on well-structured data and open libraries (RSS feeds and public APIs). Mashup enablers evolved to address this problem, providing the ability to convert other kinds of data and services into mashable resources. === Web resources === Of course, not all valuable data is located within organizations. In fact, the most valuable information for business intelligence and decision support is often external to the organization. With the emergence of rich web applications and online Web portals, a wide range of business-critical processes (such as ordering) are becoming available online. Unfortunately, very few of these data sources syndicate content in RSS format and very few of these services provide publicly accessible APIs. Mashup editors therefore solve this problem by providing enablers or connectors. == Mashups versus portals == Mashups and portals are both content aggregation technologies. Portals are an older technology designed as an extension to traditional dynamic Web applications, in which the process of converting data content into marked-up Web pages is split into two phases: generation of markup "fragments" and aggregation of the fragments into pages. Each markup fragment is generated by a "portlet", and the portal combines them into a single Web page. Portlets may be hosted locally on the portal server or remotely on a separate server. Portal technology defines a complete event model covering reads and updates. A request for an aggregate page on a portal is translated into individual read operations on all the portlets that form the page ("render" operations on local, JSR 168 portlets or "getMarkup" operations on remote, WSRP portlets). If a submit button is pressed on any portlet on a portal page, it is translated into an update operation on that portlet alone (processAction on a local portlet or performBlockingInteraction on a remote, WSRP portlet). The update is then immediately followed by a read on all portlets on the page. Portal technology is about server-side, presentation-tier aggregation. It cannot be used to drive more robust forms of application integration such as two-phase commit. Mashups differ from portals in the following respects: The portal model has been around longer and has had greater investment and product research. Portal technology is therefore more standardized and mature. Over time, increasing maturity and standardization of mashup technology will likely make it more popular than portal technology because it is more closely associated with Web 2.0 and lately Service-oriented Architectures (SOA). New versions of portal products are expected to eventually add mashup support while still supporting legacy portlet applications. Mashup technologies, in contrast, are not expected to provide support for portal standards. == Business mashups
Avizo (software)
Avizo (pronounce: 'a-VEE-zo') is a general-purpose commercial software application for scientific and industrial data visualization and analysis. Avizo is developed by Thermo Fisher Scientific and was originally designed and developed by the Visualization and Data Analysis Group at Zuse Institute Berlin (ZIB) under the name Amira. Avizo was commercially released in November 2007. For the history of its development, see the Wikipedia article about Amira. == Overview == Avizo is a software application which enables users to perform interactive visualization and computation on 3D data sets. The Avizo interface is modelled on the visual programming. Users manipulate data and module components, organized in an interactive graph representation (called Pool), or in a Tree view. Data and modules can be interactively connected together, and controlled with several parameters, creating a visual processing network whose output is displayed in a 3D viewer. With this interface, complex data can be interactively explored and analyzed by applying a controlled sequence of computation and display processes resulting in a meaningful visual representation and associated derived data. == Application areas == Avizo has been designed to support different types of applications and workflows from 2D and 3D image data processing to simulations. It is a versatile and customizable visualization tool used in many fields: Scientific visualization Materials Research Tomography, Microscopy, etc. Nondestructive testing, Industrial Inspection, and Visual Inspection Computer-aided Engineering and simulation data post-processing Porous medium analysis Civil Engineering Seismic Exploration, Reservoir Engineering, Microseismic Monitoring, Borehole Imaging Geology, Digital Rock Physics (DRP), Earth Sciences Archaeology Food technology and agricultural science Physics, Chemistry Climatology, Oceanography, Environmental Studies Astrophysics == Features == Data import: 2D and 3D image stack and volume data: from microscopes (electron, optical), X-ray tomography (CT, micro-/nano-CT, synchrotron), neutron tomography and other acquisition devices (MRI, radiography, GPR) Geometric models (such as point sets, line sets, surfaces, grids) Numerical simulation data (such as Computational fluid dynamics or Finite element analysis data) Molecular data Time series and animations Seismic data Well logs 4D Multivariate Climate Models 2D/3D data visualization: Volume rendering Digital Volume Correlation Visualization of sections, through various slicing and clipping methods Isosurface rendering Polygonal meshes Scalar fields, Vector fields, Tensor representations, Flow visualization (Illuminated Streamlines, Stream Ribbons) Image processing: 2D/3D Alignment of image slices, Image registration Image filtering Mathematical Morphology (erode, dilate, open, close, tophat) Watershed Transform, Distance Transform Image segmentation 3D models reconstruction: Polygonal surface generation from segmented objects Generation of tetrahedral grids Surface reconstruction from point clouds Skeletonization (reconstruction of dendritic, porous or fracture network) Surface model simplification Quantification and analysis: Measurements and statistics Analysis spreadsheet and charting Material properties computation, based on 3D images: Absolute permeability Thermal conductivity Molecular diffusivity Electrical resistivity/formation factor 3D image-based meshing for CFD and FEA: From 3D imaging modalities (CT, micro-CT, MRI, etc.) Surface and volume meshes generation Export to FEA and CFD solvers for simulation Post-processing for simulation analysis Presentation, automation: MovieMaker, Multiscreen, Video wall, collaboration, and VR support TCL Scripting, C++ extension API Avizo is based on Open Inventor 3D graphics toolkits (FEI Visualization Sciences Group).
Six Little Dragons
Six Little Dragons (Chinese: 杭州六小龙), or Six Little Dragons of Hangzhou, are an informal grouping of the tech startups Game Science, DeepSeek, Unitree Robotics, DEEP Robotics, BrainCo and Manycore Tech. All six were established in Hangzhou, They are active in artificial intelligence, robotics, gaming, and brain-computer interface technology. Hangzhou is referred to as the China’s “e-commerce capital” (电商之都). The nickname "Six Little Dragons" originated from the Chinese internet. == Background == === Chinese government investments (2002 — 2010s) === From 2002 to 2007, under Xi Jinping's leadership as party secretary of Zhejiang, provincial spending on technology research grew over four times to 28 billion RMB. The province launched "Digital Zhejiang" (数字浙江) to advance modernization and the "Eight Eight Strategy" (八八战略), focusing on eight advantages and actions to boost industrial development, including specialized industries. In 2010, Hangzhou's government started "Project Eagle" (雏鹰计划) to aid science and technology startups. The project works with incubators and accelerators to find promising tech companies and offers public funding and other help, especially for startups by graduates and returning students. Unitree received support in the initial phase, along with government subsidies from Binjiang District. === AI-startups and further investments (2025 — present) === In January 2025, the Chinese government created the "Hangzhou AI Industry Chain High-Quality Development Action Plan" which focuses on computing power, LLM technologies, and AI applications. The plan was made to certify over 2,000 new high-tech enterprises, initiate over 300 major tech projects, and invest more than 300 billion RMB (US$40 billion) annually. The Chinese government also renewed "Project Eagle" and to allocate 15% of industrial policy funds for future industries. Hangzhou aimed to become a center for tech startups, highlighting the "six little dragons of Hangzhou," a nickname popularized in early 2025. This group includes DeepSeek, Game Science, Unitree Robotics, Manycore Tech, BrainCo, and DEEP Robotics, companies in gaming, robotics, and software development. Earlier in 2025, DeepSeek, one of the six dragons, launched an AI system at a much lower cost than those from Silicon Valley. Since then, DeepSeek and Alibaba have produced top-performing open source AI models. Game Science launched the successful video game Black Myth: Wukong in 2024, while Unitree gained attention for their dancing robots in the 2025 annual spring gala broadcast by Chinese state media. The group was acknowledged by Chinese authorities in Hangzhou in a New Years message for local businesses in January 2025. Hangzhou’s universities were given credit for the development of Chinese technological industry. Zhejiang University alumni founded three of the "Six Little Dragons". By September 2024, the university produced 102 executives in Chinese AI start-ups, ranking third among China's top institutions. On February 20, 2025, Alibaba's Eddie Wu stated that the company would focus on artificial generative intelligence and plans significant investment in AI. The company also sought to boost foreign investment to China's "Six Little Dragons" following Alibaba's founder Jack Ma attended General Secretary of the Chinese Communist Party Xi Jinping's business symposium with corporate leaders and entrepreneurs that same month. == Challenges == China's net foreign direct investment (FDI) fell by US$168 billion in 2024, marking the largest capital flight since 1990. Foreign investment peaked at US$344 billion in 2021 but has since declined according to the State Administration of Foreign Exchange. In 2024, foreign investors put in only US$4.5 billion while Chinese firms invested US$173 billion abroad. According to interviews conducted by The New York Times, some start-up company founders believe that Chinese government's support for Hangzhou's technological sector has deterred foreign investors. Tensions with the United States led many international companies to adopt a China Plus One strategy, while Chinese firms build factories overseas to avoid potential Trump tariffs. China also faced US restrictions on its access of advanced chips, forcing Chinese tech companies to stockpile Nvidia chips while Chinese producers like Huawei and Semiconductor Manufacturing International Corporation (SMIC) were competing to produce their own.
TensorFlow
TensorFlow is a software library for machine learning and artificial intelligence. It can be used across a range of tasks, but is used mainly for training and inference of neural networks. It is one of the most popular deep learning frameworks, alongside others such as PyTorch. It is free and open-source software released under the Apache License 2.0. It was developed by the Google Brain team for Google's internal use in research and production. The initial version was released under the Apache License 2.0 in 2015. Google released an updated version, TensorFlow 2.0, in September 2019. TensorFlow can be used in a wide variety of programming languages, including Python, JavaScript, C++, and Java, facilitating its use in a range of applications in many sectors. == History == === DistBelief === Starting in 2011, Google Brain built DistBelief as a proprietary machine learning system based on deep learning neural networks. Its use grew rapidly across diverse Alphabet companies in both research and commercial applications. Google assigned multiple computer scientists, including Jeff Dean, to simplify and refactor the codebase of DistBelief into a faster, more robust application-grade library, which became TensorFlow. In 2009, the team, led by Geoffrey Hinton, had implemented generalized backpropagation and other improvements, which allowed generation of neural networks with substantially higher accuracy, for instance a 25% reduction in errors in speech recognition. === TensorFlow === TensorFlow is Google Brain's second-generation system. Version 1.0.0 was released on February 11, 2017. While the reference implementation runs on single devices, TensorFlow can run on multiple CPUs and GPUs (with optional CUDA and SYCL extensions for general-purpose computing on graphics processing units). TensorFlow is available on 64-bit Linux, macOS, Windows, and mobile computing platforms including Android and iOS. Its flexible architecture allows for easy deployment of computation across a variety of platforms (CPUs, GPUs, TPUs), and from desktops to clusters of servers to mobile and edge devices. TensorFlow computations are expressed as stateful dataflow graphs. The name TensorFlow derives from the operations that such neural networks perform on multidimensional data arrays, which are referred to as tensors. During the Google I/O Conference in June 2016, Jeff Dean stated that 1,500 repositories on GitHub mentioned TensorFlow, of which only 5 were from Google. In March 2018, Google announced TensorFlow.js version 1.0 for machine learning in JavaScript. In Jan 2019, Google announced TensorFlow 2.0. It became officially available in September 2019. In May 2019, Google announced TensorFlow Graphics for deep learning in computer graphics. === Tensor processing unit (TPU) === In May 2016, Google announced its Tensor processing unit (TPU), an application-specific integrated circuit (ASIC, a hardware chip) built specifically for machine learning and tailored for TensorFlow. A TPU is a programmable AI accelerator designed to provide high throughput of low-precision arithmetic (e.g., 8-bit), and oriented toward using or running models rather than training them. Google announced they had been running TPUs inside their data centers for more than a year, and had found them to deliver an order of magnitude better-optimized performance per watt for machine learning. In May 2017, Google announced the second-generation, as well as the availability of the TPUs in Google Compute Engine. The second-generation TPUs deliver up to 180 teraflops of performance, and when organized into clusters of 64 TPUs, provide up to 11.5 petaflops. In May 2018, Google announced the third-generation TPUs delivering up to 420 teraflops of performance and 128 GB high bandwidth memory (HBM). Cloud TPU v3 Pods offer 100+ petaflops of performance and 32 TB HBM. In February 2018, Google announced that they were making TPUs available in beta on the Google Cloud Platform. === Edge TPU === In July 2018, the Edge TPU was announced. Edge TPU is Google's purpose-built ASIC chip designed to run TensorFlow Lite machine learning (ML) models on small client computing devices such as smartphones known as edge computing. === TensorFlow Lite === In May 2017, Google announced TensorFlow Lite as a software stack to support machine learning models for mobile and embedded devices, and in November 2017, provided the developer preview. In January 2019, the TensorFlow team released a developer preview of the mobile GPU inference engine with OpenGL ES 3.1 Compute Shaders on Android devices and Metal Compute Shaders on iOS devices. In May 2019, Google announced that their TensorFlow Lite Micro (also known as TensorFlow Lite for Microcontrollers) and ARM's uTensor would be merging. It was renamed as LiteRT in 2024. === TensorFlow 2.0 === As TensorFlow's market share among research papers was declining to the advantage of PyTorch, the TensorFlow Team announced a release of a new major version of the library in September 2019. TensorFlow 2.0 introduced many changes, the most significant being TensorFlow eager, which changed the automatic differentiation scheme from the static computational graph to the "Define-by-Run" scheme originally made popular by Chainer and later PyTorch. Other major changes included removal of old libraries, cross-compatibility between trained models on different versions of TensorFlow, and significant improvements to the performance on GPU. == Features == === AutoDifferentiation === AutoDifferentiation is the process of automatically calculating the gradient vector of a model with respect to each of its parameters. With this feature, TensorFlow can automatically compute the gradients for the parameters in a model, which is useful to algorithms such as backpropagation which require gradients to optimize performance. To do so, the framework must keep track of the order of operations done to the input Tensors in a model, and then compute the gradients with respect to the appropriate parameters. === Eager execution === TensorFlow includes an "eager execution" mode, which means that operations are evaluated immediately as opposed to being added to a computational graph which is executed later. Code executed eagerly can be examined step-by step-through a debugger, since data is augmented at each line of code rather than later in a computational graph. This execution paradigm is considered to be easier to debug because of its step by step transparency. === Distribute === In both eager and graph executions, TensorFlow provides an API for distributing computation across multiple devices with various distribution strategies. This distributed computing can often speed up the execution of training and evaluating of TensorFlow models and is a common practice in the field of AI. === Losses === To train and assess models, TensorFlow provides a set of loss functions (also known as cost functions). Some popular examples include mean squared error (MSE) and binary cross entropy (BCE). === Metrics === In order to assess the performance of machine learning models, TensorFlow gives API access to commonly used metrics. Examples include various accuracy metrics (binary, categorical, sparse categorical) along with other metrics such as Precision, Recall, and Intersection-over-Union (IoU). === TF.nn === TensorFlow.nn is a module for executing primitive neural network operations on models. Some of these operations include variations of convolutions (1/2/3D, Atrous, depthwise), activation functions (Softmax, RELU, GELU, Sigmoid, etc.) and their variations, and other operations (max-pooling, bias-add, etc.). === Optimizers === TensorFlow offers a set of optimizers for training neural networks, including ADAM, ADAGRAD, and Stochastic Gradient Descent (SGD). When training a model, different optimizers offer different modes of parameter tuning, often affecting a model's convergence and performance. == Usage and extensions == === TensorFlow === TensorFlow serves as a core platform and library for machine learning. TensorFlow's APIs use Keras to allow users to make their own machine-learning models. In addition to building and training their model, TensorFlow can also help load the data to train the model, and deploy it using TensorFlow Serving. TensorFlow provides a stable Python Application Program Interface (API), as well as APIs without backwards compatibility guarantee for JavaScript, C++, and Java. Third-party language binding packages are also available for C#, Haskell, Julia, MATLAB, Object Pascal, R, Scala, Rust, OCaml, and Crystal. Bindings that are now archived and unsupported include Go and Swift. === TensorFlow.js === TensorFlow also has a library for machine learning in JavaScript. Using the provided JavaScript APIs, TensorFlow.js allows users to use either Tensorflow.js models or converted models from TensorFlow or TFLite, retrain the given models, and run on the web. === LiteRT === LiteRT, formerly known as Te
AlphaFold
AlphaFold is an artificial intelligence (AI) program developed by DeepMind, a subsidiary of Alphabet, which performs predictions of protein structure. It is designed using deep learning techniques. AlphaFold 1 (2018) placed first in the overall rankings of the 13th Critical Assessment of Structure Prediction (CASP) in December 2018. It was particularly successful at predicting the most accurate structures for targets rated as most difficult by the competition organizers, where no existing template structures were available from proteins with partially similar sequences. AlphaFold 2 (2020) repeated this placement in the CASP14 competition in November 2020. It achieved a level of accuracy much higher than any other entry. It scored above 90 on CASP's global distance test (GDT) for approximately two-thirds of the proteins, a test measuring the similarity between a computationally predicted structure and the experimentally determined structure, where 100 represents a complete match. The inclusion of metagenomic data has improved the quality of the prediction of multiple sequence alignments. One of the biggest sources of the training data was the custom-built Big Fantastic Database of 65,983,866 protein families, represented as multiple sequence alignments and Hidden Markov models, covering 2,204,359,010 protein sequences from reference databases, metagenomes, and metatranscriptomes. AlphaFold 2's results at CASP14 were described as "astounding" and "transformational". However, some researchers noted that the accuracy was insufficient for a third of its predictions, and that it did not reveal the underlying mechanism or rules of protein folding for the protein folding problem, which remains unsolved. Despite this, the technical achievement was widely recognized. On 15 July 2021, the AlphaFold 2 paper was published in Nature as an advance access publication alongside open source software and a searchable database of species proteomes. As of November 2025, the paper had been cited nearly 43,000 times. AlphaFold 3 was announced on 8 May 2024. It can predict the structure of complexes created by proteins with DNA, RNA, various ligands, and ions. The new prediction method shows a minimum 50% improvement in accuracy for protein interactions with other molecules compared to existing methods. Demis Hassabis and John Jumper shared one half of the 2024 Nobel Prize in Chemistry, awarded "for protein structure prediction," while the other half went to David Baker "for computational protein design." Hassabis and Jumper had previously won the Breakthrough Prize in Life Sciences and the Albert Lasker Award for Basic Medical Research in 2023 for their leadership of the AlphaFold project. == Background == Proteins consist of chains of amino acids which spontaneously fold to form the three dimensional (3-D) structures of the proteins. The 3-D structure is crucial to understanding the biological function of the protein. Protein structures can be determined experimentally through techniques such as X-ray crystallography, cryo-electron microscopy and nuclear magnetic resonance (NMR), which are all expensive and time-consuming. Such efforts, using the experimental methods, have identified the structures of about 170,000 proteins over the last 60 years, while there are over 200 million known proteins across all life forms. Over the years, researchers have applied numerous computational methods to predict the 3D structures of proteins from their amino acid sequences, accuracy of such methods in best possible scenario is close to experimental techniques (NMR) by the use of homology modeling based on molecular evolution. CASP, which was launched in 1994 to challenge the scientific community to produce their best protein structure predictions, found that GDT scores of only about 40 out of 100 can be achieved for the most difficult proteins by 2016. AlphaFold started competing in the 2018 CASP using an artificial intelligence (AI) deep learning technique. == Algorithm == DeepMind is known to have trained the program on over 170,000 protein structures from the Protein Data Bank, a public repository of protein sequences and structures. The program uses a form of attention network, a deep learning technique that focuses on having the AI identify parts of a larger problem, then piece it together to obtain the overall solution. The overall training was conducted on processing power between 100 and 200 GPUs. === AlphaFold 1 (2018) === AlphaFold 1 (2018) was built on work developed by various teams in the 2010s, work that looked at the large databases of related protein sequences now available from many different organisms (most without known 3D structures), to try to find changes at different residues (peptides) that appeared to be correlated, even though the residues were not consecutive in the main chain. Such correlations suggest that the residues may be close to each other physically, even though not close in the sequence, allowing a contact map to be estimated. Building on recent work prior to 2018, AlphaFold 1 extended this by estimating a probability distribution for the distances between residues, effectively transforming the contact map into a distance map. It also used more advanced learning methods than previously to develop the inference. The code was not made publicly available, except to run on sequences of proteins in the 2018 CASP competition. === AlphaFold 2 (2020) === The 2020 version of the program (AlphaFold 2, 2020) is significantly different from the original version that won CASP 13 in 2018, according to the team at DeepMind. AlphaFold 1 used a number of separately trained modules to produce a guide potential, which was then combined with a physics-based energy potential. AlphaFold 2 replaced this with a system of interconnected sub-networks, forming a single, differentiable, end-to-end model based on pattern recognition. This model was trained in an integrated manner. After the neural network's prediction converges, a final refinement step applies local physical constraints using energy minimization based on the AMBER force field. This step only slightly adjusts the predicted structure. A key part of the 2020 system are two modules, believed to be based on a transformer design, which are used to progressively refine a vector of information for each relationship (or "edge" in graph-theory terminology) between an amino acid residue of the protein and another amino acid residue (these relationships are represented by the array shown in green); and between each amino acid position and each different sequences in the input sequence alignment (these relationships are represented by the array shown in red). Internally these refinement transformations contain layers that have the effect of bringing relevant data together and filtering out irrelevant data (the "attention mechanism") for these relationships, in a context-dependent way, learned from training data. These transformations are iterated, the updated information output by one step becoming the input of the next, with the sharpened residue/residue information feeding into the update of the residue/sequence information, and then the improved residue/sequence information feeding into the update of the residue/residue information. As the iteration progresses, according to one report, the "attention algorithm ... mimics the way a person might assemble a jigsaw puzzle: first connecting pieces in small clumps—in this case clusters of amino acids—and then searching for ways to join the clumps in a larger whole." The output of these iterations then informs the final structure prediction module, which also uses transformers, and is itself then iterated. In an example presented by DeepMind, the structure prediction module achieved a correct topology for the target protein on its first iteration, scored as having a GDT_TS of 78, but with a large number (90%) of stereochemical violations – i.e. unphysical bond angles or lengths. With subsequent iterations the number of stereochemical violations fell. By the third iteration the GDT_TS of the prediction was approaching 90, and by the eighth iteration the number of stereochemical violations was approaching zero. The training data was originally restricted to single peptide chains. However, the October 2021 update, named AlphaFold-Multimer, included protein complexes in its training data. DeepMind stated this update succeeded about 70% of the time at accurately predicting protein-protein interactions. === AlphaFold 3 (2024) === Announced on 8 May 2024, AlphaFold 3 was co-developed by Google DeepMind and Isomorphic Labs, both subsidiaries of Alphabet. AlphaFold 3 is not limited to proteins, as it can also predict the structures of protein complexes with DNA, RNA, post-translational modifications and selected ligands and ions. AlphaFold 3 introduces the "Pairformer," a deep learning architecture inspired by the transformer, which is considered similar to, but si
Actor-critic algorithm
The actor-critic algorithm (AC) is a family of reinforcement learning (RL) algorithms that combine policy-based RL algorithms such as policy gradient methods, and value-based RL algorithms such as value iteration, Q-learning, SARSA, and TD learning. An AC algorithm consists of two main components: an "actor" that determines which actions to take according to a policy function, and a "critic" that evaluates those actions according to a value function. Some AC algorithms are on-policy, some are off-policy. Some apply to either continuous or discrete action spaces. Some work in both cases. == Overview == The actor-critic methods can be understood as an improvement over pure policy gradient methods like REINFORCE via introducing a baseline. === Actor === The actor uses a policy function π ( a | s ) {\displaystyle \pi (a|s)} , while the critic estimates either the value function V ( s ) {\displaystyle V(s)} , the action-value Q-function Q ( s , a ) , {\displaystyle Q(s,a),} the advantage function A ( s , a ) {\displaystyle A(s,a)} , or any combination thereof. The actor is a parameterized function π θ {\displaystyle \pi _{\theta }} , where θ {\displaystyle \theta } are the parameters of the actor. The actor takes as argument the state of the environment s {\displaystyle s} and produces a probability distribution π θ ( ⋅ | s ) {\displaystyle \pi _{\theta }(\cdot |s)} . If the action space is discrete, then ∑ a π θ ( a | s ) = 1 {\displaystyle \sum _{a}\pi _{\theta }(a|s)=1} . If the action space is continuous, then ∫ a π θ ( a | s ) d a = 1 {\displaystyle \int _{a}\pi _{\theta }(a|s)da=1} . The goal of policy optimization is to improve the actor. That is, to find some θ {\displaystyle \theta } that maximizes the expected episodic reward J ( θ ) {\displaystyle J(\theta )} : J ( θ ) = E π θ [ ∑ t = 0 T γ t r t ] {\displaystyle J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\gamma ^{t}r_{t}\right]} where γ {\displaystyle \gamma } is the discount factor, r t {\displaystyle r_{t}} is the reward at step t {\displaystyle t} , and T {\displaystyle T} is the time-horizon (which can be infinite). The goal of policy gradient method is to optimize J ( θ ) {\displaystyle J(\theta )} by gradient ascent on the policy gradient ∇ J ( θ ) {\displaystyle \nabla J(\theta )} . As detailed on the policy gradient method page, there are many unbiased estimators of the policy gradient: ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ j ≤ T ∇ θ ln π θ ( A j | S j ) ⋅ Ψ j | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{0\leq j\leq T}\nabla _{\theta }\ln \pi _{\theta }(A_{j}|S_{j})\cdot \Psi _{j}{\Big |}S_{0}=s_{0}\right]} where Ψ j {\textstyle \Psi _{j}} is a linear sum of the following: ∑ 0 ≤ i ≤ T ( γ i R i ) {\textstyle \sum _{0\leq i\leq T}(\gamma ^{i}R_{i})} . γ j ∑ j ≤ i ≤ T ( γ i − j R i ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})} : the REINFORCE algorithm. γ j ∑ j ≤ i ≤ T ( γ i − j R i ) − b ( S j ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})-b(S_{j})} : the REINFORCE with baseline algorithm. Here b {\displaystyle b} is an arbitrary function. γ j ( R j + γ V π θ ( S j + 1 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma V^{\pi _{\theta }}(S_{j+1})-V^{\pi _{\theta }}(S_{j})\right)} : TD(1) learning. γ j Q π θ ( S j , A j ) {\textstyle \gamma ^{j}Q^{\pi _{\theta }}(S_{j},A_{j})} . γ j A π θ ( S j , A j ) {\textstyle \gamma ^{j}A^{\pi _{\theta }}(S_{j},A_{j})} : Advantage Actor-Critic (A2C). γ j ( R j + γ R j + 1 + γ 2 V π θ ( S j + 2 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma R_{j+1}+\gamma ^{2}V^{\pi _{\theta }}(S_{j+2})-V^{\pi _{\theta }}(S_{j})\right)} : TD(2) learning. γ j ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(n) learning. γ j ∑ n = 1 ∞ λ n − 1 1 − λ ⋅ ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\sum _{n=1}^{\infty }{\frac {\lambda ^{n-1}}{1-\lambda }}\cdot \left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(λ) learning, also known as GAE (generalized advantage estimate). This is obtained by an exponentially decaying sum of the TD(n) learning terms. === Critic === In the unbiased estimators given above, certain functions such as V π θ , Q π θ , A π θ {\displaystyle V^{\pi _{\theta }},Q^{\pi _{\theta }},A^{\pi _{\theta }}} appear. These are approximated by the critic. Since these functions all depend on the actor, the critic must learn alongside the actor. The critic is learned by value-based RL algorithms. For example, if the critic is estimating the state-value function V π θ ( s ) {\displaystyle V^{\pi _{\theta }}(s)} , then it can be learned by any value function approximation method. Let the critic be a function approximator V ϕ ( s ) {\displaystyle V_{\phi }(s)} with parameters ϕ {\displaystyle \phi } . The simplest example is TD(1) learning, which trains the critic to minimize the TD(1) error: δ i = R i + γ V ϕ ( S i + 1 ) − V ϕ ( S i ) {\displaystyle \delta _{i}=R_{i}+\gamma V_{\phi }(S_{i+1})-V_{\phi }(S_{i})} The critic parameters are updated by gradient descent on the squared TD error: ϕ ← ϕ − α ∇ ϕ ( δ i ) 2 = ϕ + α δ i ∇ ϕ V ϕ ( S i ) {\displaystyle \phi \leftarrow \phi -\alpha \nabla _{\phi }(\delta _{i})^{2}=\phi +\alpha \delta _{i}\nabla _{\phi }V_{\phi }(S_{i})} where α {\displaystyle \alpha } is the learning rate. Note that the gradient is taken with respect to the ϕ {\displaystyle \phi } in V ϕ ( S i ) {\displaystyle V_{\phi }(S_{i})} only, since the ϕ {\displaystyle \phi } in γ V ϕ ( S i + 1 ) {\displaystyle \gamma V_{\phi }(S_{i+1})} constitutes a moving target, and the gradient is not taken with respect to that. This is a common source of error in implementations that use automatic differentiation, and requires "stopping the gradient" at that point. Similarly, if the critic is estimating the action-value function Q π θ {\displaystyle Q^{\pi _{\theta }}} , then it can be learned by Q-learning or SARSA. In SARSA, the critic maintains an estimate of the Q-function, parameterized by ϕ {\displaystyle \phi } , denoted as Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} . The temporal difference error is then calculated as δ i = R i + γ Q θ ( S i + 1 , A i + 1 ) − Q θ ( S i , A i ) {\displaystyle \delta _{i}=R_{i}+\gamma Q_{\theta }(S_{i+1},A_{i+1})-Q_{\theta }(S_{i},A_{i})} . The critic is then updated by θ ← θ + α δ i ∇ θ Q θ ( S i , A i ) {\displaystyle \theta \leftarrow \theta +\alpha \delta _{i}\nabla _{\theta }Q_{\theta }(S_{i},A_{i})} The advantage critic can be trained by training both a Q-function Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} and a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then let A ϕ ( s , a ) = Q ϕ ( s , a ) − V ϕ ( s ) {\displaystyle A_{\phi }(s,a)=Q_{\phi }(s,a)-V_{\phi }(s)} . Although, it is more common to train just a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then estimate the advantage by A ϕ ( S i , A i ) ≈ ∑ j ∈ 0 : n − 1 γ j R i + j + γ n V ϕ ( S i + n ) − V ϕ ( S i ) {\displaystyle A_{\phi }(S_{i},A_{i})\approx \sum _{j\in 0:n-1}\gamma ^{j}R_{i+j}+\gamma ^{n}V_{\phi }(S_{i+n})-V_{\phi }(S_{i})} Here, n {\displaystyle n} is a positive integer. The higher n {\displaystyle n} is, the more lower is the bias in the advantage estimation, but at the price of higher variance. The Generalized Advantage Estimation (GAE) introduces a hyperparameter λ {\displaystyle \lambda } that smoothly interpolates between Monte Carlo returns ( λ = 1 {\displaystyle \lambda =1} , high variance, no bias) and 1-step TD learning ( λ = 0 {\displaystyle \lambda =0} , low variance, high bias). This hyperparameter can be adjusted to pick the optimal bias-variance trade-off in advantage estimation. It uses an exponentially decaying average of n-step returns with λ {\displaystyle \lambda } being the decay strength. == Variants == Asynchronous Advantage Actor-Critic (A3C): Parallel and asynchronous version of A2C. Soft Actor-Critic (SAC): Incorporates entropy maximization for improved exploration. Deep Deterministic Policy Gradient (DDPG): Specialized for continuous action spaces.
Common Crawl
The Common Crawl Foundation (Common Crawl) is a nonprofit 501(c)(3) organization that crawls the web and freely provides its archives and datasets to the public. Common Crawl was founded by Gil Elbaz. The data had mostly been primarily used by researchers and some startups until the 2020s, when AI companies started training large language models using the data. In November 2025, an investigation by The Atlantic revealed that Common Crawl misled publishers when it claimed it respected paywalls in its scraping and it was not honoring requests from publishers to have their content removed from its databases. == History == Common Crawl was founded in 2007 in San Francisco. It began publishing its crawls in 2011. By 2013, sites like TinEye were building their products off of Common Crawl. The crawl reduces the reliance of companies and researchers on Google, which has the biggest dataset. Common Crawl was designed to have more and fresher data that was more efficient to analyze and utilize than the Wayback Machine created by the Internet Archive. By 2015, 1.8 billion webpages were on the Common Crawl, which started by crawling a list of URLs donated by the search engine Blekko. They use Amazon Web Services, which provides some of its services for free, allowing computing costs to average $2-4000/month. The Common Crawl website listed 30 studies based on Common Crawl data. Before 2023, Common Crawl was not very well known outside of academic researchers who utilize the data. Common Crawl received its first requests to redact information in 2023 and increasingly started seeing its crawler, CCBot, blocked. In 2023, it began receiving significant financial support from AI companies, including Anthropic and OpenAI, each of which donated $250,000. It was also used to train Google DeepMind's large language model Gemini. By April 2023, Common Crawl was capturing 3.1 billion webpages, with an estimated 5% of pages before 2021 containing hate speech or slurs. As of 2024, Common Crawl had been cited in more than 10,000 academic studies. By 2024, The Pile and Common Crawl had been the two main training datasets being used to train AI models. In November 2025, an investigation by technology journalist Alex Reisner for The Atlantic revealed that Common Crawl misled publishers when it claimed it respected paywalls in its scraping and when it said that it was honoring requests from publishers to have their content removed from its databases. It included misleading results in the public search function on its website that showed no entries for websites that had requested their archives be removed, when in fact those sites were still included in its scrapes used by AI companies. As of 2025, Reisner found that CCBot was the most widely-blocked bot by the top 1000 websites. A 2026 article in LWN.net discussed an advantage to services like Common Crawl being that it can limit the scraping costs to websites by allowing companies and researchers to download the data from Common Crawl instead of scraping it themselves. In April 2026, Common Crawl experimentally began to distribute its data through Hugging Face Storage Bucket, in addition to its standard storage on Amazon S3. == Organization == Peter Norvig and Joi Ito have served on the advisory board. Rich Skrenta is the executive director. It has received funding almost exclusively from the Elbaz Family Foundation Trust until 2023 when it started receiving donations from the AI industry. == Refined versions == A number of organizations take raw Common Crawl data and refine it into datasets that exclude edgy content or are otherwise higher-quality for their purposes, such as FineWeb, DCLM and C4. === Colossal Clean Crawled Corpus === Google version of the Common Crawl is called the Colossal Clean Crawled Corpus, or C4 for short. It was constructed for the training of the T5 language model series in 2019. As of 2023, there were some concerns over copyrighted content in the C4 as well as racist content. A 2024 study found that 45% of content was explicitly restricted by websites' terms of service to be used for purposes like AI training by for-profit companies.