AI Art For Sale

AI Art For Sale — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

  • KidDesk

    KidDesk

    KidDesk is an alternative desktop software application. The early childhood learning company Hatch Early Childhood created KidDesk; it subsequently went to Edmark, which was bought by IBM then sold to Riverdeep (now Houghton Mifflin Harcourt Learning Technology). KidDesk is compatible with Microsoft Windows 95 and newer, as well as Apple System 7 and newer. KidDesk can be set to start when the computer starts up, and can only be exited through password entry. Adults choose what programs are included for the child to use, what icon represented the desk, and customize the software programs available for use. == History == Edmark first started shipping KidDesk in 1992. In 1993, Edmark updated KidDesk with KidDesk Family Edition for Macintosh and DOS, adding more desk accessories and desk styles (Sometimes included as a free exclusive offer with the Early Learning House and Thinkin' Things Series). In 1995, KidDesk Family Edition was enhanced for Windows 95, and released one month after the new operating system shipped. In 1998, Edmark developed KidDesk Internet Safe. The Internet Safe edition was written for Windows 95, Windows 98, and Macintosh (including OS8). In 2008, HMH ported KidDesk Family Edition was to run on Windows Vista and in 2011 version 3.07 of KidDesk Family Edition was released as part of the 'Young Explorer' suite which is fully supported on Windows XP, Windows Vista and Windows 7. == Features == A picture editor incorporated into the desk. Used both in the Adult settings menu and in the desk itself. KidDesk users can edit their user logo with a pixel grid paint program. A calendar incorporated into the desk. This allows the user to set dates that the user finds important, and allows the date to be marked with a picture or text. A password exit feature. For security reasons, the adult can set a password so that KidDesk can only be exited if it is entered. As an extra security measure, the password exit function could only be accessed if the user pressed the ctrl + alt + A keyboard buttons simultaneously. A skin changer with several themes - farm, princess, sports, ocean, etc. These themes can be changed. The e-mail and voicemail features are customizable depending on the KidDesk installation. The ability to add websites that can be accessed on KidDesk, and the ability to block hyperlinks, JavaScript, data entry, etc., on said sites was an added for the 'Internet Safe' edition released in 1998. KidDesk Internet Safe edition is available in Spanish and Brazilian-Portuguese versions. == Reception == KidDesk was given a platinum award at the 1994 Oppenheim Toy Portfolio Awards. The judges praised the program's security features allowing "configur[ation] so that kids never have access to the possibly destructive DOS prompt", and concluded that "[i]f you and your kids share a computer, you need to install Kiddesk immediately!" === Awards === Since 1992, KidDesk has won 15 major awards.

    Read more →
  • Minion (solver)

    Minion (solver)

    Minion is a solver for satisfaction problems. Unlike constraint programming toolkits, which expect users to write programs in a traditional programming language like C++, Java or Prolog, Minion takes a text file which specifies the problem, and solves using only this. This makes using Minion much simpler, at the cost of much less customization. Minion has been shown to be faster than major commercial constraint solvers including CPLEX (formerly IBM ILOG). == Overview == Minion was introduced in 2006 by researchers at the University of St Andrews as a “fast, scalable” solver for large and hard CSP instances. The project provides a compact input language and a low-overhead C++ implementation aimed at throughput and memory efficiency. == Design and features == Minion implements a range of variable and constraint types commonly used in CSP modelling, plus search heuristics and optimisation support. The solver architecture prioritises cache-friendly data structures and specialised propagators. Notably, the developers adapted watched literal techniques from SAT solving to speed up constraint propagation for, among others, Boolean sums, the element global constraint, and table constraints. The modelling approach relies on a plain-text format (parsed by Minion) rather than embedding models into a host programming language. This reduces overhead and supports rapid “model-and-run” experimentation for large benchmark sets. == Performance == In the original evaluation on standard benchmarks, the authors reported that Minion often ran between one and two orders of magnitude faster than state-of-the-art toolkits of the time (including ILOG Solver and Gecode) on large, hard instances, with smaller gains—or slowdowns—on easier problems. Subsequent research has used Minion as a baseline solver in empirical studies and test generation tasks, reflecting its adoption within parts of the constraint programming community. == Applications == Minion has been applied in academic work on combinatorial search, scheduling and test generation, and is available to other environments via wrappers (for example, from the R language).

    Read more →
  • DeepSeek

    DeepSeek

    Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., doing business as DeepSeek, is a Chinese artificial intelligence (AI) company that develops large language models (LLMs). Based in Hangzhou, Zhejiang, DeepSeek is owned and funded by High-Flyer, a Chinese hedge fund. DeepSeek was founded in July 2023 by Liang Wenfeng, the co-founder of High-Flyer, who also serves as the CEO for both of the companies. The company launched an eponymous chatbot alongside its DeepSeek-R1 model in January 2025. DeepSeek-R1 provided responses comparable to other contemporary large language models, such as OpenAI's GPT-4 and o1. Its training cost was reported to be significantly lower than other LLMs. The company claims that it trained its V3 model for US$6 million—far less than the US$100 million cost for OpenAI's GPT-4 in 2023—and using approximately one-tenth the computing power consumed by Meta's comparable model, Llama 3.1. DeepSeek's success against larger and more established rivals has been described as "upending AI". DeepSeek's models are described as "open-weight", meaning the exact parameters are openly shared, but the training data is not openly licensed. Since the January 2025 debut of DeepSeek-R1, the company has made its new models available under free and open-source software licenses, primarily the MIT License. The company reportedly recruits AI researchers from top Chinese universities and also hires from outside traditional computer science fields to broaden its models' knowledge and capabilities. DeepSeek significantly reduced training expenses for their R1 model by incorporating techniques such as mixture of experts (MoE) layers. The company also trained its models during ongoing trade restrictions on AI chip exports to China, using weaker AI chips intended for export and employing fewer units overall. Observers say this breakthrough sent "shock waves" through the industry which were described as triggering a "Sputnik moment" for the US in the field of artificial intelligence, particularly due to its open-source, cost-effective, and high-performing AI models. This threatened established AI hardware leaders such as Nvidia; Nvidia's share price dropped sharply, losing US$600 billion in market value, the largest single-company decline in U.S. stock market history. == History == === Founding and early years (2016–2023) === In February 2016, High-Flyer was co-founded by AI enthusiast Liang Wenfeng, who had been trading since the 2008 financial crisis while attending Zhejiang University. The company began stock trading using a GPU-dependent deep learning model on 21 October 2016; before then, it had used CPU-based linear models. By the end of 2017, most of its trading was driven by AI. Liang established High-Flyer as a hedge fund focused on developing and using AI trading algorithms, and by 2021 the firm was using AI exclusively, often using Nvidia chips. In 2019, the company began constructing its first computing cluster, Fire-Flyer, at a cost of 200 million yuan; it contained 1,100 GPUs interconnected at 200 Gbit/s and was retired after 1.5 years in operation. By 2021, Liang had started buying large quantities of Nvidia GPUs for an AI project, reportedly obtaining 10,000 Nvidia A100 GPUs before the United States restricted chip sales to China. Computing cluster Fire-Flyer 2 began construction in 2021 with a budget of 1 billion yuan. It was reported that in 2022, Fire-Flyer 2's capacity had been used at over 96%, totaling 56.74 million GPU hours. 27% was used to support scientific computing outside the company. During 2022, Fire-Flyer 2 had 5,000 PCIe A100 GPUs in 625 nodes, each containing 8 GPUs. At the time, it exclusively used PCIe instead of the DGX version of A100, since at the time the models it trained could fit within a single 40 GB GPU VRAM and so there was no need for the higher bandwidth of DGX (i.e., it required only data parallelism but not model parallelism). Later, it incorporated NVLinks and NCCL (Nvidia Collective Communications Library) to train larger models that required model parallelism. On 14 April 2023, High-Flyer announced the launch of an artificial general intelligence (AGI) research lab, stating that the new lab would focus on developing AI tools unrelated to the firm's financial business. Two months later, on 17 July 2023, that lab was spun off into an independent company, DeepSeek, with High-Flyer as its principal investor and backer. Venture capital investors were reluctant to provide funding, as they considered it unlikely that the venture would be able to quickly generate an "exit". === Model releases since 2023 === DeepSeek released its first model, DeepSeek Coder, on 2 November 2023, followed by the DeepSeek-LLM series on 29 November 2023. In January 2024, it released two DeepSeek-MoE models (Base and Chat), and in April 3 DeepSeek-Math models (Base, Instruct, and RL). DeepSeek-V2 was released in May 2024, followed a month later by the DeepSeek-Coder V2 series. In September 2024, DeepSeek V2.5 was introduced and revised in December. On 20 November 2024, the preview of DeepSeek-R1-Lite became available via chat. In December, DeepSeek-V3-Base and DeepSeek-V3 (chat) were released. On 20 January 2025, DeepSeek launched the DeepSeek chatbot—based on the DeepSeek-R1 model—free for iOS and Android. By 27 January, DeepSeek surpassed ChatGPT as the most downloaded freeware app on the iOS App Store in the United States, triggering an 18% drop in Nvidia's share price. On 24 March 2025, DeepSeek released DeepSeek-V3-0324 under the MIT License. On 28 May 2025, DeepSeek released DeepSeek-R1-0528 under the MIT License. The model has been noted for more tightly following official Chinese Communist Party ideology and censorship in its answers to questions than prior models. On 21 August 2025, DeepSeek released DeepSeek V3.1 under the MIT License. This model features a hybrid architecture with thinking and non-thinking modes. It also surpasses prior models like V3 and R1, by over 40% on certain benchmarks like SWE-bench and Terminal-bench. It was updated to V3.1-Terminus on 22 September 2025. V3.2-Exp was released on 29 September 2025. It uses DeepSeek Sparse Attention, a more efficient attention mechanism based on previous research published in February. DeepSeek-V3.2 was released on 1 December 2025, alongside a DeepSeek-V3.2-Speciale variant that focused on reasoning. In February 2026, Anthropic accused DeepSeek of using thousands of fraudulent accounts to generate millions of conversations with Claude to train its own large language models. In April 2026, investors began speaking with DeepSeek for a $300 million funding round, which would bring DeepSeek to a total valuation of $10 billion. On April 24, 2026, DeepSeek released a preview of its V4 series, including the 1.6-trillion parameter DeepSeek-V4-Pro and the 284-billion parameter DeepSeek-V4-Flash, both featuring a 1-million token context window, under the MIT License. DeepSeek's V4 LLM has been adopted by key semiconductor manufacturers and artificial intelligence chipmakers such as Huawei and Cambricon. == Company operation == DeepSeek is headquartered in Hangzhou, Zhejiang, and is owned and funded by High-Flyer. Its co-founder, Liang Wenfeng, serves as CEO. As of May 2024, Liang personally held an 84% stake in DeepSeek through two shell corporations. === Strategy === DeepSeek has stated that it focuses on research and does not have immediate plans for commercialization. This posture also means it can skirt certain provisions of China's AI regulations aimed at consumer-facing technologies. DeepSeek's hiring approach emphasizes skills over lengthy work experience, resulting in many hires fresh out of university. The company likewise recruits individuals without computer science backgrounds to expand the range of expertise incorporated into the models, for instance in poetry or advanced mathematics. According to The New York Times, dozens of DeepSeek researchers have or have previously had affiliations with People's Liberation Army laboratories and the Seven Sons of National Defence. Due to the impact of United States restrictions on chips, DeepSeek refined its algorithms to maximise computational efficiency and thereby leveraged older hardware and reduced energy consumption. DeepSeek also expanded on the African continent as it offers more affordable and less power-hungry AI solutions. The company has bolstered African language models and generated a number of startups, for example in Nairobi. Along with Huawei's storage and cloud computing services, the impact on the tech scene in sub-saharan Africa is considerable. DeepSeek offers local data sovereignty and more flexibility compared to Western AI platforms. == Training framework == High-Flyer/DeepSeek had operated at least two primary computing clusters: Fire-Flyer (萤火一号) and Fire-Flyer 2 (萤火二号). Fire-Flyer 1 was constructed in 2019 and was retired after 1.5 years of operation. Fi

    Read more →
  • Embodied cognition

    Embodied cognition

    Embodied cognition represents a diverse group of theories which investigate how cognition is shaped by the bodily state and capacities of the organism. These embodied factors include the motor system, the perceptual system, bodily interactions with the environment (situatedness), and the assumptions about the world that shape the functional structure of the brain and body of the organism. Embodied cognition suggests that these elements are essential to a wide spectrum of cognitive functions, such as perception biases, memory recall, comprehension and high-level mental constructs (such as meaning attribution and categories) and performance on various cognitive tasks (reasoning or judgment). The embodied mind thesis challenges other theories, such as cognitivism, computationalism, and Cartesian dualism. It is closely related to the extended mind thesis, situated cognition, and enactivism. The modern version depends on understandings drawn from up-to-date research in psychology, linguistics, cognitive science, dynamical systems, artificial intelligence, robotics, animal cognition, plant cognition, and neurobiology. == Theory == Proponents of the embodied cognition thesis emphasize the active and significant role the body plays in the shaping of cognition and in the understanding of an agent's mind and cognitive capacities. In philosophy, embodied cognition holds that an agent's cognition, rather than being the product of mere (innate) abstract representations of the world, is strongly influenced by aspects of an agent's body beyond the brain itself. An embodied model of cognition opposes the disembodied Cartesian model, according to which all mental phenomena are non-physical and, therefore, not influenced by the body. With this opposition the embodiment thesis intends to reintroduce an agent's bodily experiences into any account of cognition. It is a rather broad thesis and encompasses both weak and strong variants of embodiment. In an attempt to reconcile cognitive science with human experience, the enactive approach to cognition defines "embodiment" as follows: By using the term embodied we mean to highlight two points: first that cognition depends upon the kinds of experience that come from having a body with various sensorimotor capacities, and second, that these individual sensorimotor capacities are themselves embedded in a more encompassing biological, psychological and cultural context. This double sense attributed to the embodiment thesis emphasizes the many aspects of cognition that researchers in different fields—such as philosophy, cognitive science, artificial intelligence, psychology, and neuroscience—are involved with. This general characterization of embodiment faces some difficulties: a consequence of this emphasis on the body, experience, culture, context, and the cognitive mechanisms of an agent in the world is that often distinct views and approaches to embodied cognition overlap. The theses of extended cognition and situated cognition, for example, are usually intertwined and not always carefully separated. And since each of the aspects of the embodiment thesis is endorsed to different degrees, embodied cognition should be better seen "as a research program rather than a well-defined unified theory". Some authors explain the embodiment thesis by arguing that cognition depends on an agent's body and its interactions with a determined environment. From this perspective, cognition in real biological systems is not an end in itself; it is constrained by the system's goals and capacities. Such constraints do not mean cognition is set by adaptive behavior (or autopoiesis) alone, but instead that cognition requires "some kind of information processing... the transformation or communication of incoming information". The acquiring of such information involves the agent's "exploration and modification of the environment". It would be a mistake, however, to suppose that cognition consists simply of building maximally accurate representations of input information...the gaining of knowledge is a stepping stone to achieving the more immediate goal of guiding behavior in response to the system's changing surroundings. Another approach to understanding embodied cognition comes from a narrower characterization of the embodiment thesis. The following narrower view of embodiment avoids any compromises to external sources other than the body and allows differentiating between embodied cognition, extended cognition, and situated cognition. Thus, the embodiment thesis can be specified as follows: Many features of cognition are embodied in that they are deeply dependent upon characteristics of the physical body of an agent, such that the agent's beyond-the-brain body plays a significant causal role, or a physically constitutive role, in that agent's cognitive processing. This thesis points out the core idea that an agent's body plays a significant role in shaping different features of cognition, such as perception, attention, memory, reasoning—among others. Likewise, these features of cognition depend on the kind of body an agent has. The thesis omits direct mention of some aspects of the "more encompassing biological, psychological and cultural context" included in the enactive definition, making it possible to separate embodied cognition, extended cognition, and situated cognition. In contrast to the embodiment thesis, the extended mind thesis limits cognitive processing neither to the brain nor even to the body, it extends it outward into the agent's world. Situated cognition emphasizes that this extension is not just a matter of including resources outside the head but stressing the role of probing and changing interactions with the agent's world. Cognition is situated in that it is inherently dependent upon the cultural and social contexts within which it takes place. This conceptual reframing of cognition as an activity influenced by the body has had significant implications. For instance, the view of cognition inherited by most contemporary cognitive neuroscience is internalist in nature. An agent's behavior along with its capacity to maintain (accurate) representations of the surrounding environment were considered as the product of "powerful brains that can maintain the world models and devise plans". From this perspective, cognizing was conceived as something that an isolated brain did. In contrast, accepting the role the body plays during cognitive processes allows us to account for a more encompassing view of cognition. This shift in perspective within neuroscience suggests that successful behavior in real-world scenarios demands the integration of several sensorimotor and cognitive (as well as affective) capacities of an agent. Thus, cognition emerges in the relationship between an agent and the affordances provided by the environment rather than in the brain alone. In 2002, a collection of positive characterizations summarizing what the embodiment thesis entails for cognition were offered. Professor of Cognitive Psychology Margaret Wilson argues that the general outlook of embodied cognition "displays an interesting co-variation of multiple observations and houses a number of different claims: (1) cognition is situated; (2) cognition is time-pressured; (3) we off-load cognitive work onto the environment; (4) the environment is part of the cognitive system; (5) cognition is for action; (6) offline cognition is bodily-based". According to Wilson, the first three and the fifth claim appear to be at least partially true, while the fourth claim is deeply problematic in that all things that have an impact on the elements of a system are not necessarily considered part of the system. The sixth claim has received the least attention in the literature on embodied cognition, yet it might be the most significant of the six claims as it shows how certain human cognitive capabilities, that previously were thought to be highly abstract, now appear to be leaning towards an embodied approach for their explanation. Wilson also describes at least five main (abstract) categories that combine both sensory and motor skills (or sensorimotor functions). The first three are working memory, episodic memory, and implicit memory; the fourth is mental imagery, and finally, the fifth concerns reasoning and problem solving. == History == The theory of embodied cognition, along with the multiple aspects it comprises, can be regarded as the imminent result of an intellectual skepticism towards the flourishment of the disembodied theory of mind put forth by René Descartes in the 17th century. According to Cartesian dualism, the mind is entirely distinct from the body and can be successfully explained and understood without reference to the body or to its processes. Research has been done to identify the set of ideas that would establish what could be considered as the early stages of embodied cognition around inquiries regarding the mind-body-soul rel

    Read more →
  • Fantavision

    Fantavision

    Fantavision is an animation program by Scott Anderson for the Apple II and published by Broderbund in 1985. Versions were released for the Apple IIGS (1987), Amiga (1988), and MS-DOS (1988). Fantavision allows the creation of vector graphics animations using the mouse and keyboard. The user creates frames, and the software generates the frames between them. Because this is done in real-time, it allows for creative exploration and quick changes. The program uses a graphical user interface in the style of the Macintosh with pull-down menus and black text on a white background. Advertisements claimed Fantavision a revolutionary breakthrough that brings the animation features of "tweening" and "transforming" to home computers. == Reception == Compute! in 1989 called Fantavision the best animation program for the IBM PC, although it noted the inability to draw curves. == Reviews == Games #70

    Read more →
  • Microelectronics and Computer Technology Corporation

    Microelectronics and Computer Technology Corporation

    Microelectronics and Computer Technology Corporation, originally the Microelectronics and Computer Consortium and widely seen by the acronym MCC, was the first, and at one time one of the largest, computer industry research and development consortia in the United States. MCC ceased operations in 2000 and was formally dissolved in 2004. == Divisions == MCC did research and development in the following areas: [1] System Architecture and Design (optimise hardware and software design, provide for scalability and interoperability, allow rapid prototyping for improved time-to-market, and support the re-engineering of existing systems for open systems). Advanced Microelectronics Packaging and Interconnection (smaller, faster, more powerful, and cost-competitive). Hardware Systems Engineering (tools and methodologies for cost-efficient, up-front design of advanced electronic systems, including modelling and design-for-test techniques to improve cost, yield, quality, and time-to-market). Environmentally Conscious Technologies (process control and optimisation tools, information management and analysis capabilities, and non-hazardous material alternatives supporting cost-efficient production, waste minimisation, and reduced environmental impact). Distributed Information Technology (managing and maintaining physically distributed corporate information resources on different platforms, building blocks for the national information infrastructure, networking tools and services for integration within and between companies, and electronic commerce). Intelligent Systems (systems that "intelligently" support business processes and enhance performance, including decision support, data management, forecasting and prediction). == History == The MCC was a response to the announcement of Japan's Fifth Generation Project, a large Japanese research project launched in 1982 aimed at producing a new kind of computer by 1991. The Japanese had formed similar industrial research consortia as early as 1956.[2] Many European and American computer companies saw this new Japanese initiative as an attempt to take full control of the world's high-end computer market, and MCC was created, in part, as a defensive move against that threat. In late 1982, several major computer and semiconductor manufacturers in the United States banded together and founded MCC under the leadership of Admiral Bobby Ray Inman, whose previous positions had been Director of the National Security Agency and deputy director of the Central Intelligence Agency. Such formations were illegal in the United States until the 1984 Congressional passage of the "National Cooperative Research Act". Several sites with relevant universities were considered, including Atlanta, Georgia (Georgia Tech), the Research Triangle, N.C. (UNC), the Washington, D.C. area (George Mason), Stanford University and Austin, Texas (UT) which was the final selection. The University of Texas offered land upon which they would construct a new building specifically designed for the MCC within their Austin campus. Ross Perot also offered the use of his private plane for 2 years for staff recruitment. Austin was selected as the site for MCC in 1983. Despite this purpose and the background of Inman and his senior staff, MCC accepted no government funding for many years and was a refuge for some avoiding work on Strategic Defense Initiative projects. MCC was part of the Artificial Intelligence boom of the 1980s, reportedly the single largest customer of both Symbolics and Lisp Machines, Inc. (and like Symbolics, was one of the first companies to register a .com domain). In the 1980s its major programs were packaging, software engineering, CAD, and advanced computer architectures. The latter comprised artificial intelligence, human interface, database, and parallel processing, the latter two merging in the late 1980s. Many of the early shareholder companies were mainframe computer companies under stress in the 1980s. Over the years, MCC's membership diversified to include a broad range of high-profile corporations involved in information technology products, as well as government research and development agencies and leading universities. In June, 2000 the MCC Board of Directors voted to dissolve the consortium, and the few remaining employees held a wake at Scholz's Beer Garden in Austin on October 25. Formal dissolution papers were reportedly not filed until 2004. == Spinoffs == While multiple technologies were transferred to member companies and government agencies in the final years, fourteen companies were spun out of MCC. Those spinoffs include: TeraVicta Technologies, Austin's first MEMS company; its focus was to develop microscopic switch technology for fiber optic switching and radiofrequency switching in mobile phones specifically to dynamically switch between the future 3G-4GLTE-future5G wireless communication frequencies and ensure mobile phones were communicating over the strongest wireless signal to reduce dropped calls. Robert Miracky was the founding CEO who spun out the first commercial metal micromachining technology developed by MCC researchers Brent Lunceford, Jason Reed, Richard Nelson, K.Hu, and C. Hilbert in a collaborative development program with IBM in a novel implementation and operational paradigm for solid-state integrated circuit coolers integrated with conductive MEMS switches. TeraVicta was liquidated under Chapter 7 bankruptcy proceedings in 2015. The Austin region subsequently built up a MEMS & Sensors value chain in the billions of dollars comprising companies such as 3M, Cypress Semiconductor, NXP Semiconductor, Cirrus Logic, Silicon Labs, and the Austin division of the now-defunct Silicon Valley Technology Center. Portelligent, a company that provides reverse engineering teardown services. At the time, Portelligent was the first company to commercialize such services; they had been provided by MCC to its member companies. Today, there are at least twelve companies worldwide that sell reports known as "reverse engineering teardown reports." Modern day teardown reports provide detailed information about technology products such as the bill of materials, microchip, and printed circuit board design specifics, manufacturing details including manufacturing location details for the entire value chain responsible for making electronics, including the iPhone and Samsung Galaxy smartphones. Portelligent was acquired by CMP Technology in 2007. Evolutionary Technologies International, a company focused on developing database tools and data warehousing. It was spun off from MCC in 1990.

    Read more →
  • Noam Shazeer

    Noam Shazeer

    Noam Shazeer (born 1975 or 1976) is an American computer scientist and entrepreneur known for his contributions to the field of artificial intelligence and deep learning, particularly in the development of transformer models and natural language processing. He lives in Palo Alto, California. == Career == Noam Shazeer joined Google in 2000. One of his first major achievements was improving the spelling corrector of Google's search engine. In 2017, Shazeer was one of the lead authors of the seminal paper "Attention Is All You Need", which introduced the transformer architecture. At Google, Shazeer and his colleague Daniel de Freitas built a chatbot named Meena. Following the refusal of Google to release the chatbot to the public, Shazeer and Freitas left the company in 2021 to found Character.AI. In September 2023, Time Magazine chose Shazeer as one of the 100 most influential people in the AI world. In August 2024, it was reported that Shazeer would be returning to Google to co-lead the Gemini AI project. Shazeer was appointed as technical lead on Gemini, along with Jeff Dean and Oriol Vinyals. It was part of a $2.7 billion deal for Google to license Character's technology. Since he owns 30-40% of the company, it is estimated he netted $750 million-$1 billion. In 2026, he was elected a member of the National Academy of Engineering. == Views == Shazeer said about artificial general intelligence that he doesn't "particularly care about AGI in the sense of wanting something that can do absolutely everything a person can do”. When asked in 2023 if he is afraid that AGI will destroy the world, he said: "No. Not yet. [...] We’re going to work on it as the technology improves". When asked why do large language models work he answered: "My best guess is divine benevolence [...] Nobody really understands what’s going on. This is a very experimental science [...] It’s more like alchemy or whatever chemistry was in the Middle Ages.” Shazeer has stated, "I do not believe that humans have an attribute called gender... I do not believe that G-d puts people in the wrong bodies. I do not believe that it is okay to sterilize children." == Personal life == Shazeer is an orthodox Jew. His grandparents escaped the Holocaust into the Soviet Union and later lived some time in Israel before emigrating to the USA. His father, Dov Shazeer, was a math teacher who became an engineer and his mother was a homemaker. His sister was ordained as a rabbi by Hebrew College. Shazeer was born in Philadelphia, attended grade school at Cohen Hillel Academy in Marblehead, Massachusetts, and attended Swampscott High School in Swampscott, Massachusetts. He won a gold medal with perfect score at International Mathematical Olympiad 1994 as a member of the USA team. He went on to study math and computer science at Duke University in Durham, North Carolina from 1994 to 1998. At Duke he was a recipient of the Angier B. Duke Memorial Scholarship, and, as part of the Duke math team, won prizes in several math tournaments. He started studying in a graduate program in Berkeley but did not finish it. He is a father of three and is married to Yael Shacham Shazeer

    Read more →
  • Yann LeCun

    Yann LeCun

    Yann André Le Cun ( lə-KUN; French: [ləkœ̃]; usually spelled LeCun; born 8 July 1960) is a French-American computer scientist working in the fields of artificial intelligence, machine learning, computer vision, robotics and image compression. He is the Jacob T. Schwartz Professor of Computer Science at the Courant Institute of Mathematical Sciences at New York University. He served as Chief AI Scientist at Meta Platforms before co-founding Advanced Machine Intelligence Labs in December 2025. He is well known for his work on optical character recognition and computer vision using convolutional neural networks (CNNs). He is also one of the main creators of the DjVu image compression technology, alongside Léon Bottou and Patrick Haffner. He co-developed the Lush programming language with Léon Bottou. In 2018, LeCun, Yoshua Bengio, and Geoffrey Hinton received the Turing Award from the Association for Computing Machinery (ACM) for their work on deep learning. LeCun, Bengio, and Hinton, and occasionally Jürgen Schmidhuber, are sometimes referred to as the "Godfathers of AI" and "Godfathers of Deep Learning". == Early life and education == Yann André Le Cun was born on 8 July 1960 at Soisy-sous-Montmorency, in the suburbs of Paris. His surname, Le Cun, derives from the old Breton form Le Cunff and originates from the region of Guingamp in northern Brittany. Yann is the Breton form of Jean, the French form of John. He received a Diplôme d'Ingénieur from the ESIEE Paris in 1983 and a PhD in computer science from Université Pierre et Marie Curie (now Sorbonne University) in 1987, during which he proposed an early form of backpropagation, an algorithm crucial for enabling neural networks to learn. Before joining AT&T, LeCun was a postdoctoral researcher for a year, starting in 1987, supervised by Geoffrey Hinton at the University of Toronto. LeCun has three sons, and his brother is employed by Google. He has American citizenship. == Career and research == LeCun's career has been spent primarily at Bell Labs, New York University and Meta Platforms, Inc. === Bell Labs === In 1988, LeCun joined the Adaptive Systems Research Department at AT&T Bell Laboratories in Holmdel, New Jersey, United States, headed by Lawrence D. Jackel, where he developed a number of new machine learning methods, such as a biologically inspired model of image recognition called convolutional neural networks (LeNet), the "Optimal Brain Damage" regularization methods, and the Graph Transformer Networks method (similar to conditional random field), which he applied to handwriting recognition and Optical character recognition (OCR). The bank check recognition system that he helped develop was widely deployed by NCR and other companies. In 1996, he joined AT&T Labs-Research as head of the Image Processing Research Department, which was part of Lawrence Rabiner's Speech and Image Processing Research Lab, and worked primarily on the DjVu image compression technology, a format designed for efficient distribution of scanned documents, and used by the Internet Archive to provide access to digitized texts. His collaborators at AT&T include Léon Bottou and Vladimir Vapnik. === New York University === After a brief tenure as a fellow of NEC Research Institute, LeCun joined New York University in 2003, where he is Jacob T. Schwartz Chaired Professor of Computer Science and Neural Science at the Courant Institute of Mathematical Sciences and the Center for Neural Science. At NYU, he has worked primarily on energy-based models for supervised and unsupervised learning, feature learning for object recognition in computer vision, and mobile robotics. In 2012, he became the founding director of the NYU Center for Data Science. On 9 December 2013, LeCun became the first director of Meta AI Research in New York City and in early 2014 stepped down from the NYU–CDS directorship. In 2013, he and Yoshua Bengio co-founded the International Conference on Learning Representations, which adopted a post-publication open review process he previously advocated on his website. He was the chair and organiser of the "Learning Workshop" held every year between 1986 and 2012 in Snowbird, Utah. He is a member of the Science Advisory Board of the Institute for Pure and Applied Mathematics at UCLA. He is the co-director of the Learning in Machines and Brain research program (formerly Neural Computation & Adaptive Perception) of CIFAR. In 2016, he was the visiting professor of computer science on the Chaire Annuelle Informatique et Sciences Numériques at Collège de France in Paris, where he presented the leçon inaugurale (inaugural lecture). In 2023, he was named as the inaugural Jacob T. Schwartz Chaired Professor in Computer Science at NYU's Courant Institute. LeCun is also a scientific advisor to French research group Kyutai which is being funded by Xavier Niel, Rodolphe Saadé, Eric Schmidt, and others. === Meta Platforms === LeCun joined Facebook (now Meta Platforms) in 2013 as chief AI scientist and led the company's AI research laboratory, FAIR. === AMI Labs === On 19 November 2025, LeCun confirmed that he would be leaving Meta after ten years to found his own company focused on world-model architectures and human-like artificial intelligence he calls superintelligence. The company he founded, Advanced Machine Intelligence Labs (or AMI Labs), is run by CEO Alex LeBrun, with LeCun serving as Executive Chair. This venture is focused on building AI "world models": systems that learn to understand the physical world's structure and dynamics rather than just predict text like large language models. In March 2026, AMI announced it had raised $1.03 billion in funding at a $3.5 billion pre-money valuation. The funding round was co-led by investors including Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. In January 2026, LeCun became founding chair of the Technical Research Board of Logical Intelligence, an AI company developing energy-based (EBM) reasoning systems. == Honours and awards == LeCun is a member of the US National Academy of Sciences, National Academy of Engineering and the French Académie des Sciences. He has received honorary doctorates from Instituto Politécnico Nacional (IPN) in Mexico City in 2016, from EPFL in 2018, from Université Côte d'Azur in 2021, from Università di Siena in 2023, and from Hong Kong University of Science and Technology in 2023. In 2014, he received the IEEE Neural Network Pioneer Award and in 2015, the PAMI Distinguished Researcher Award. In 2018, LeCun was awarded the IRI Medal, established by the Industrial Research Institute (IRI), and the Harold Pender Award, given by the University of Pennsylvania. In 2019, he received the Golden Plate Award of the American Academy of Achievement. In March 2019, LeCun won the 2018 Turing Award, sharing it with Yoshua Bengio and Geoffrey Hinton. In 2022, he received the Princess of Asturias Award in the category "Scientific Research", along with Yoshua Bengio, Geoffrey Hinton and Demis Hassabis. In 2023, the President of France made him a Chevalier (Knight) of the French Legion of Honour. During the World Economic Forum (WEF) 2024 in Davos, he received the Global Swiss AI Award 2023. The same year, he received the grand prize of the VinFuture Prize alongside Yoshua Bengio, Jensen Huang, Geoffrey Hinton, and Fei-Fei Li for their groundbreaking contributions to neural networks and deep learning algorithms. In 2025 he was awarded the Queen Elizabeth Prize for Engineering jointly with Yoshua Bengio, Bill Dally, Geoffrey E. Hinton, John Hopfield, Jensen Huang and Fei-Fei Li.

    Read more →
  • Zesta

    Zesta

    Zesta is an online food ordering and delivery platform operating across the African region. Formerly known as Square Eats, the company rebranded to Zesta in 2025. Zesta connects customers with restaurants and stores, offering delivery services for food, groceries, parcel delivery and other essentials. == History == Zesta was originally founded as Square Eats in 2020 by twin brothers Henry Newman and Randall Newman when they were 21 years old. It was launched in Gaborone, Botswana, and quickly gained traction as a leading food delivery service in the country. The company halted operations and took a strategic decision to reinvent the business in 2022. In 2025, the company announced its rebranding to Zesta, highlighting its commitment to evolving beyond food delivery to become a super app. === COVID-19 initiative === During the COVID-19 pandemic, Zesta (then Square Eats) implemented measures to ensure safety and hygiene, including providing free gloves and hand sanitizer to drivers and introducing contactless delivery options. These efforts positioned the platform as a trusted service during the pandemic. == Service == Zesta facilitates delivery from a wide range of merchant partners via a smartphone app, available on iOS and Android platforms, or through its website. Customers can browse their favorite restaurants, place orders, and have meals delivered to their doorstep efficiently.

    Read more →
  • Composite Capability/Preference Profiles

    Composite Capability/Preference Profiles

    Composite Capability/Preference Profiles (CC/PP) is a specification for defining capabilities and preferences of user agents (also known as "delivery context"). The delivery context can be used to guide the process of tailoring content for a user agent. CC/PP is a vocabulary extension of the Resource Description Framework (RDF). The CC/PP specification is maintained by the W3C's Ubiquitous Web Applications Working Group (UWAWG) Working Group. == History == Composite Capability/Preference Profiles (CC/PP): Structure and Vocabularies 1.0 became a W3C recommendation on 15 January 2004. A "Last-Call Working-Draft" of CC/PP 2.0 was issued in April 2007

    Read more →
  • General Problem Solver

    General Problem Solver

    General Problem Solver (GPS) is a computer program created in 1957 by Herbert A. Simon, J. C. Shaw, and Allen Newell (RAND Corporation) intended to work as a universal problem solver machine. In contrast to the former Logic Theorist project, the GPS works with means–ends analysis. == Overview == Any problem that can be expressed as a set of well-formed formulas (WFFs) or Horn clauses, and that constitutes a directed graph with one or more sources (that is, hypotheses) and sinks (that is, desired conclusions), can be solved, in principle, by GPS. Proofs in the predicate logic and Euclidean geometry problem spaces are prime examples of the domain of applicability of GPS. It was based on Simon and Newell's theoretical work on logic machines. GPS was the first computer program that separated its knowledge of problems (rules represented as input data) from its strategy of how to solve problems (a generic solver engine). GPS was implemented in the third-order programming language, IPL. While GPS solved simple problems such as the Towers of Hanoi that could be sufficiently formalized, it could not solve any real-world problems because the search was easily lost in the combinatorial explosion. Put another way, the number of "walks" through the inferential digraph became computationally untenable. (In practice, even a straightforward state space search such as the Towers of Hanoi can become computationally infeasible, albeit judicious prunings of the state space can be achieved by such elementary AI techniques as A and IDA). The user defined objects and operations that could be done on the objects, and GPS generated heuristics by means–ends analysis in order to solve problems. It focused on the available operations, finding what inputs were acceptable and what outputs were generated. It then created subgoals to get closer and closer to the goal. The GPS paradigm eventually evolved into the Soar architecture for artificial intelligence.

    Read more →
  • The Synthetic Party

    The Synthetic Party

    Det Syntetiske Parti (English: The Synthetic Party) is a political party driven by artificial intelligence (AI), founded in May 2022 in Denmark. The party aims to represent non-voters and fringe political parties while raising awareness of AI's societal role and exploring how it can be integrated into democratic processes. == Founder == The founder and continuous party secretary is Asker Bryld Staunæs, a philosopher from Aarhus University and a conceptual artist. == Main goal == The political goals have been machine learned from texts by Danish fringe parties since 1970 and represent the 20 percent of Danes who do not vote in the election. The party is synthetic; as such, many of the policies, such as universal basic income, can be contradictory to one another. == International collaborations == The Synthetic Party has signed bilateral collaboration agreements with the Finnish AI Party and AI Party (Japan) concerning the development of a global project created around artificial intelligence and politics These collaborations were expanded during the exhibition-event Synthetic Summit (28 February – 13 April 2025) at Kunsthal Aarhus, curated by Computer Lars (Asker Bryld Staunæs) on behalf of The Synthetic Party. The summit staged parliamentary scenography, performances, and computer sculptures, and invited both the public and policymakers to encounter an international line-up of AI parties and virtual politicians. Aarhus University described the event as part of Staunæs's PhD research, positioning it as an international top-meeting of virtual politicians. Participants included the Japanese AI Party, the Swedish AI Party, the Finnish AI Party, Parker Politics (New Zealand), Lex AI (Brazil), the Simiyya collective (Egypt/Sweden), the Synthetic Party (Denmark), and Wiktoria Cukt 2.0 (Poland). As part of the summit, the one-day AI World Congress was held on 1 March 2025, structured as a performative assembly where each group participated through both machinic agents and human delegates. Sessions were chaired by participating parties, with Computer Lars delivering the opening presentation. Throughout the day, contributions were synthesized into a common record using a shared AI system. The congress concluded with the adoption of the Synthetic Summit Resolution, a collectively authored treaty of algorithmic governance. Signatories included Floor Kist and Nick Gerritsen (Parker Politics), Michihito Matsuda (Japanese AI Party), Emma Bexell (Swedish AI Party), Samee Haapa (Finnish AI Party), Pedro Markun (Lex AI), Kristian T. Madsen and Michael Birkebæk Jensen (NextGen Democracy / DemAI), Asker Bryld Staunæs, Benjamin Asger Krog Møller, Caroline Sofie Axelsson, Life with Artificials (The Synthetic Party), and Piotr Wyrzykowski (Wiktoria Cukt 2.0).

    Read more →
  • Sequence labeling

    Sequence labeling

    In machine learning, sequence labeling is a type of pattern recognition task that involves the algorithmic assignment of a categorical label to each member of a sequence of observed values. A common example of a sequence labeling task is part of speech tagging, which seeks to assign a part of speech to each word in an input sentence or document. Sequence labeling can be treated as a set of independent classification tasks, one per member of the sequence. However, accuracy is generally improved by making the optimal label for a given element dependent on the choices of nearby elements, using special algorithms to choose the globally best set of labels for the entire sequence at once. As an example of why finding the globally best label sequence might produce better results than labeling one item at a time, consider the part-of-speech tagging task just described. Frequently, many words are members of multiple parts of speech, and the correct label of such a word can often be deduced from the correct label of the word to the immediate left or right. For example, the word "sets" can be either a noun or verb. In a phrase like "he sets the books down", the word "he" is unambiguously a pronoun, and "the" unambiguously a determiner, and using either of these labels, "sets" can be deduced to be a verb, since nouns very rarely follow pronouns and are less likely to precede determiners than verbs are. But in other cases, only one of the adjacent words is similarly helpful. In "he sets and then knocks over the table", only the word "he" to the left is helpful (cf. "...picks up the sets and then knocks over..."). Conversely, in "... and also sets the table" only the word "the" to the right is helpful (cf. "... and also sets of books were ..."). An algorithm that proceeds from left to right, labeling one word at a time, can only use the tags of left-adjacent words and might fail in the second example above; vice versa for an algorithm that proceeds from right to left. Most sequence labeling algorithms are probabilistic in nature, relying on statistical inference to find the best sequence. The most common statistical models in use for sequence labeling make a Markov assumption, i.e. that the choice of label for a particular word is directly dependent only on the immediately adjacent labels; hence the set of labels forms a Markov chain. This leads naturally to the hidden Markov model (HMM), one of the most common statistical models used for sequence labeling. Other common models in use are the maximum entropy Markov model and conditional random field.

    Read more →
  • Connectionism

    Connectionism

    Connectionism is an approach to the study of human mental processes and cognition that utilizes mathematical models known as connectionist networks or artificial neural networks. Connectionism has had many "waves" since its beginnings. The first wave appeared 1943 with Warren Sturgis McCulloch and Walter Pitts both focusing on comprehending neural circuitry through a formal and mathematical approach, and Frank Rosenblatt who published the 1958 paper "The Perceptron: A Probabilistic Model For Information Storage and Organization in the Brain" in Psychological Review, while working at the Cornell Aeronautical Laboratory. The first wave ended with the 1969 book Perceptrons about limitations of the original perceptron idea, written by Marvin Minsky and Seymour Papert, which contributed to discouraging major funding agencies in the US from investing in connectionist research. With a few noteworthy deviations, most connectionist research entered a period of inactivity until the mid-1980s. The term connectionist model was reintroduced in a 1982 paper in the journal Cognitive Science by Jerome Feldman and Dana Ballard. The second wave blossomed in the late 1980s, following a 1987 book Parallel Distributed Processing by James L. McClelland, David E. Rumelhart, et al., which introduced a couple of improvements to the simple perceptron idea, such as intermediate processors (now known as "hidden layers") alongside input and output units, and used a sigmoid activation function instead of the old "all-or-nothing" function. Their work built upon that of John Hopfield, who was a key figure investigating the mathematical characteristics of sigmoid activation functions. From the late 1980s to the mid-1990s, connectionism took on an almost revolutionary tone when Schneider, Terence Horgan and Tienson posed the question of whether connectionism represented a fundamental shift in psychology and so-called "good old-fashioned AI", or GOFAI. Some advantages of the second wave connectionist approach included its applicability to a broad array of functions, structural approximation to biological neurons, low requirements for innate structure, and capacity for graceful degradation. Its disadvantages included the difficulty in deciphering how ANNs process information or account for the compositionality of mental representations, and a resultant difficulty explaining phenomena at a higher level. The current (third) wave has been marked by advances in deep learning, which have made possible the creation of large language models. The success of deep-learning networks in the past decade has greatly increased the popularity of this approach, but the complexity and scale of such networks has brought with them increased interpretability problems. == Basic principle == The central connectionist principle is that mental phenomena can be described by interconnected networks of simple and often uniform units. The form of the connections and the units can vary from model to model. For example, units in the network could represent neurons and the connections could represent synapses, as in the human brain. This principle has been seen as an alternative to GOFAI and the classical theories of mind based on symbolic computation, but the extent to which the two approaches are compatible has been the subject of much debate since their inception. === Activation function === Internal states of any network change over time due to neurons sending a signal to a succeeding layer of neurons in the case of a feedforward network, or to a previous layer in the case of a recurrent network. Discovery of non-linear activation functions has enabled the second wave of connectionism. === Memory and learning === Neural networks follow two basic principles: Any mental state can be described as a n-dimensional vector of numeric activation values over neural units in a network. Memory and learning are created by modifying the 'weights' of the connections between neural units, generally represented as an n×m matrix. The weights are adjusted according to some learning rule or algorithm, such as Hebbian learning. Most of the variety among the models comes from: Interpretation of units: Units can be interpreted as neurons or groups of neurons. Definition of activation: Activation can be defined in a variety of ways. For example, in a Boltzmann machine, the activation is interpreted as the probability of generating an action potential spike, and is determined via a logistic function on the sum of the inputs to a unit. Learning algorithm: Different networks modify their connections differently. In general, any mathematically defined change in connection weights over time is referred to as the "learning algorithm". === Biological realism === Connectionist work in general does not need to be biologically realistic. One area where connectionist models are thought to be biologically implausible is with respect to error-propagation networks that are needed to support learning, but error propagation can explain some of the biologically-generated electrical activity seen at the scalp in event-related potentials such as the N400 and P600, and this provides some biological support for one of the key assumptions of connectionist learning procedures. Many recurrent connectionist models also incorporate dynamical systems theory. Many researchers, such as the connectionist Paul Smolensky, have argued that connectionist models will evolve toward fully continuous, high-dimensional, non-linear, dynamic systems approaches. == Precursors == Precursors of the connectionist principles can be traced to early work in psychology, such as that of William James. Psychological theories based on knowledge about the human brain were fashionable in the late 19th century. As early as 1869, the neurologist John Hughlings Jackson argued for multi-level, distributed systems. Following from this lead, Herbert Spencer's Principles of Psychology, 3rd edition (1872), and Sigmund Freud's Project for a Scientific Psychology (composed 1895) propounded connectionist or proto-connectionist theories. These tended to be speculative theories. But by the early 20th century, Edward Thorndike was writing about human learning that posited a connectionist type network. Hopfield networks had precursors in the Ising model due to Wilhelm Lenz (1920) and Ernst Ising (1925), though the Ising model conceived by them did not involve time. Monte Carlo simulations of Ising model required the advent of computers in the 1950s. == The first wave == The first wave begun in 1943 with Warren Sturgis McCulloch and Walter Pitts both focusing on comprehending neural circuitry through a formal and mathematical approach. McCulloch and Pitts showed how neural systems could implement first-order logic: Their classic paper "A Logical Calculus of Ideas Immanent in Nervous Activity" (1943) is important in this development here. They were influenced by the work of Nicolas Rashevsky in the 1930s and symbolic logic in the style of Principia Mathematica. Hebb contributed greatly to speculations about neural functioning, and proposed a learning principle, Hebbian learning. Lashley argued for distributed representations as a result of his failure to find anything like a localized engram in years of lesion experiments. Friedrich Hayek independently conceived the model, first in a brief unpublished manuscript in 1920, then expanded into a book in 1952. The Perceptron machines were proposed and built by Frank Rosenblatt, who published the 1958 paper “The Perceptron: A Probabilistic Model For Information Storage and Organization in the Brain” in Psychological Review, while working at the Cornell Aeronautical Laboratory. He cited Hebb, Hayek, Uttley, and Ashby as main influences. Another form of connectionist model was the relational network framework developed by the linguist Sydney Lamb in the 1960s. The research group led by Widrow empirically searched for methods to train two-layered ADALINE networks (MADALINE), with limited success. A method to train multilayered perceptrons with arbitrary levels of trainable weights was published by Alexey Grigorevich Ivakhnenko and Valentin Lapa in 1965, called the Group Method of Data Handling. This method employs incremental layer by layer training based on regression analysis, where useless units in hidden layers are pruned with the help of a validation set. The first multilayered perceptrons trained by stochastic gradient descent was published in 1967 by Shun'ichi Amari. In computer experiments conducted by Amari's student Saito, a five layer MLP with two modifiable layers learned useful internal representations to classify non-linearily separable pattern classes. In 1972, Shun'ichi Amari produced an early example of self-organizing network. == The neural network winter == There was some conflict among artificial intelligence researchers as to what neural networks are useful for. Around late 1960s, there was a widespread lull in research a

    Read more →
  • MuZero

    MuZero

    MuZero is a computer program developed by artificial intelligence research company DeepMind, a subsidiary of Google, to master games without knowing their rules and underlying dynamics. Its release in 2019 included benchmarks of its performance in Go, chess, shogi, and a suite of 57 different Atari games. The algorithm uses an approach similar to AlphaZero, where a combination of a tree-based search and a learned model is deployed. It matched AlphaZero's performance in chess and shogi, improved on its performance in Go, and improved on the state of the art in mastering a suite of 57 Atari games (the Arcade Learning Environment), a visually-complex domain. MuZero was trained via self-play, with no access to rules, opening books, or endgame tablebases. The trained algorithm used the same convolutional and residual architecture as AlphaZero, but with 20 percent fewer computation steps per node in the search tree. == History == MuZero really is discovering for itself how to build a model and understand it just from first principles. On November 19, 2019, the DeepMind team released a preprint introducing MuZero. === Derivation from AlphaZero === MuZero (MZ) is a combination of the high-performance planning of the AlphaZero (AZ) algorithm with approaches to model-free reinforcement learning. The combination allows for more efficient training in classical planning regimes, such as Go, while also handling domains with much more complex inputs at each stage, such as visual video games. MuZero was derived directly from AZ code, sharing its rules for setting hyperparameters. Differences between the approaches include: AZ's planning process uses a simulator. The simulator knows the rules of the game. It has to be explicitly programmed. A neural network then predicts the policy and value of a future position. Perfect knowledge of game rules is used in modeling state transitions in the search tree, actions available at each node, and termination of a branch of the tree. MZ does not have access to the rules, and instead learns one with neural networks. AZ has a single model for the game (from board state to predictions); MZ has separate models for representation of the current state (from board state into its internal embedding), dynamics of states (how actions change representations of board states), and prediction of policy and value of a future position (given a state's representation). MZ's hidden model may be complex, and it may turn out it can host computation; exploring the details of the hidden model in a trained instance of MZ is a topic for future exploration. MZ does not expect a two-player game where winners take all. It works with standard reinforcement-learning scenarios, including single-agent environments with continuous intermediate rewards, possibly of arbitrary magnitude and with time discounting. AZ was designed for two-player games that could be won, drawn, or lost. === Comparison with R2D2 === The previous state of the art technique for learning to play the suite of Atari games was R2D2, the Recurrent Replay Distributed DQN. MuZero surpassed both R2D2's mean and median performance across the suite of games, though it did not do better in every game. == Training and results == MuZero used 16 third-generation tensor processing units (TPUs) for training, and 1000 TPUs for selfplay for board games, with 800 simulations per step and 8 TPUs for training and 32 TPUs for selfplay for Atari games, with 50 simulations per step. AlphaZero used 64 second-generation TPUs for training, and 5000 first-generation TPUs for selfplay. As TPU design has improved (third-generation chips are 2x as powerful individually as second-generation chips, with further advances in bandwidth and networking across chips in a pod), these are comparable training setups. R2D2 was trained for 5 days through 2M training steps. === Initial results === MuZero matched AlphaZero's performance in chess and shogi after roughly 1 million training steps. It matched AZ's performance in Go after 500,000 training steps and surpassed it by 1 million steps. It matched R2D2's mean and median performance across the Atari game suite after 500 thousand training steps and surpassed it by 1 million steps, though it never performed well on 6 games in the suite. == Reactions and related work == MuZero was viewed as a significant advancement over AlphaZero, and a generalizable step forward in unsupervised learning techniques. The work was seen as advancing understanding of how to compose systems from smaller components, a systems-level development more than a pure machine-learning development. While only pseudocode was released by the development team, Werner Duvaud produced an open source implementation based on that. MuZero has been used as a reference implementation in other work, for instance as a way to generate model-based behavior. In late 2021, a more efficient variant of MuZero was proposed, named EfficientZero. It "achieves 194.3 percent mean human performance and 109.0 percent median performance on the Atari 100k benchmark with only two hours of real-time game experience". In early 2022, a variant of MuZero was proposed to play stochastic games (for example 2048, backgammon), called Stochastic MuZero, which uses afterstate dynamics and chance codes to account for the stochastic nature of the environment when training the dynamics network.

    Read more →