AI Assistant Volume

AI Assistant Volume — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

AI safety

AI safety is an interdisciplinary field focused on preventing accidents, misuse, or other harmful consequences arising from artificial intelligence systems. It encompasses AI alignment (which aims to ensure AI systems behave as intended), monitoring AI systems for risks, and enhancing their robustness. The field is particularly concerned with existential risks posed by advanced AI models. Beyond technical research, AI safety involves developing norms and policies that promote safety, including advocacy for regulations at different levels of government. The field gained significant popularity in 2023, with rapid progress in generative AI and public concerns voiced by researchers and CEOs about potential dangers. During the 2023 AI Safety Summit, the United States and the United Kingdom both established their own AI Safety Institute. However, researchers have expressed concern that AI safety measures are not keeping pace with the rapid development of AI capabilities. == Motivations == Scholars discuss current risks from critical systems failures, bias, and AI-enabled surveillance, as well as emerging risks like technological unemployment, digital manipulation, weaponization, AI-enabled cyberattacks and bioterrorism. They also discuss speculative risks from losing control of future artificial general intelligence (AGI) agents, or from AI enabling perpetually stable dictatorships. === Existential safety === Some have criticized concerns about AGI, such as Andrew Ng who compared them in 2015 to "worrying about overpopulation on Mars when we have not even set foot on the planet yet". Stuart J. Russell on the other side urges caution, arguing that "it is better to anticipate human ingenuity than to underestimate it". AI researchers have widely differing opinions about the severity and primary sources of risk posed by AI technology – though surveys suggest that experts take high consequence risks seriously. In two surveys of AI researchers, the median respondent was optimistic about AI overall, but placed a 5% probability on an "extremely bad (e.g. human extinction)" outcome of advanced AI. In a 2022 survey of the natural language processing community, 37% agreed or weakly agreed that it is plausible that AI decisions could lead to a catastrophe that is "at least as bad as an all-out nuclear war". == History == Risks from AI began to be seriously discussed at the start of the computer age: Moreover, if we move in the direction of making machines which learn and whose behavior is modified by experience, we must face the fact that every degree of independence we give the machine is a degree of possible defiance of our wishes. In 1988 Blay Whitby published a book outlining the need for AI to be developed along ethical and socially responsible lines. From 2008 to 2009, the Association for the Advancement of Artificial Intelligence (AAAI) commissioned a study to explore and address potential long-term societal influences of AI research and development. The panel was generally skeptical of the radical views expressed by science-fiction authors but agreed that "additional research would be valuable on methods for understanding and verifying the range of behaviors of complex computational systems to minimize unexpected outcomes". In 2011, Roman Yampolskiy introduced the term "AI safety engineering" at the Philosophy and Theory of Artificial Intelligence conference, listing prior failures of AI systems and arguing that "the frequency and seriousness of such events will steadily increase as AIs become more capable". In 2014, philosopher Nick Bostrom published the book Superintelligence: Paths, Dangers, Strategies. He has the opinion that the rise of AGI has the potential to create various societal issues, ranging from the displacement of the workforce by AI, manipulation of political and military structures, to even the possibility of human extinction. His argument that future advanced systems may pose a threat to human existence prompted Elon Musk, Bill Gates, and Stephen Hawking to voice similar concerns. In 2015, dozens of artificial intelligence experts signed an open letter on artificial intelligence calling for research on the societal impacts of AI and outlining concrete directions. To date, the letter has been signed by over 8000 people including Yann LeCun, Shane Legg, Yoshua Bengio, and Stuart Russell. In the same year, a group of academics led by professor Stuart J. Russell founded the Center for Human-Compatible AI at the University of California Berkeley and the Future of Life Institute awarded $6.5 million in grants for research aimed at "ensuring artificial intelligence (AI) remains safe, ethical and beneficial". In 2016, the White House Office of Science and Technology Policy and Carnegie Mellon University announced The Public Workshop on Safety and Control for Artificial Intelligence, which was one of a sequence of four White House workshops aimed at investigating "the advantages and drawbacks" of AI. In the same year, Concrete Problems in AI Safety – one of the first and most influential technical AI Safety agendas – was published. In 2017, the Future of Life Institute sponsored the Asilomar Conference on Beneficial AI, where more than 100 thought leaders formulated principles for beneficial AI including "Race Avoidance: Teams developing AI systems should actively cooperate to avoid corner-cutting on safety standards". In 2018, the DeepMind Safety team outlined AI safety problems in specification, robustness, and assurance. The following year, researchers organized a workshop at ICLR that focused on these problem areas. In 2021, Unsolved Problems in ML Safety was published, outlining research directions in robustness, monitoring, alignment, and systemic safety. In 2023, Rishi Sunak said he wants the United Kingdom to be the "geographical home of global AI safety regulation" and to host the first global summit on AI safety. The AI safety summit took place in November 2023, and focused on the risks of misuse and loss of control associated with frontier AI models. During the summit the intention to create the International Scientific Report on the Safety of Advanced AI was announced. In 2024, The US and UK forged a new partnership on the science of AI safety. The MoU was signed on 1 April 2024 by US commerce secretary Gina Raimondo and UK technology secretary Michelle Donelan to jointly develop advanced AI model testing, following commitments announced at an AI Safety Summit in Bletchley Park in November. In 2025, an international team of 96 experts chaired by Yoshua Bengio published the first International AI Safety Report. The report, commissioned by 30 nations and the United Nations, represents the first global scientific review of potential risks associated with advanced artificial intelligence. It details potential threats stemming from misuse, malfunction, and societal disruption, with the objective of informing policy through evidence-based findings, without providing specific recommendations. == Research focus == AI safety research areas include robustness, monitoring, and alignment. === Robustness === ==== Adversarial robustness ==== AI systems are often vulnerable to adversarial examples or "inputs to machine learning (ML) models that an attacker has intentionally designed to cause the model to make a mistake". For example, in 2013, Szegedy et al. discovered that adding specific imperceptible perturbations to an image could cause it to be misclassified with high confidence. This continues to be an issue with neural networks, though in recent work the perturbations are generally large enough to be perceptible. The image on the right is predicted to be an ostrich after the perturbation is applied. (Left) is a correctly predicted sample, (center) perturbation applied magnified by 10x, (right) adversarial example. Adversarial robustness is often associated with security. Researchers demonstrated that an audio signal could be imperceptibly modified so that speech-to-text systems transcribe it to any message the attacker chooses. Network intrusion and malware detection systems also must be adversarially robust since attackers may design their attacks to fool detectors. Models that represent objectives (reward models) must also be adversarially robust. For example, a reward model might estimate how helpful a text response is and a language model might be trained to maximize this score. Researchers have shown that if a language model is trained for long enough, it will leverage the vulnerabilities of the reward model to achieve a better score and perform worse on the intended task. This issue can be addressed by improving the adversarial robustness of the reward model. More generally, any AI system used to evaluate another AI system must be adversarially robust. This could include monitoring tools, since they could also potentially be tampered with to produce a higher reward. Large language models (LLMs) can be vulnerable to prom
Read more →
Emi Kusano

Emi Kusano (Japanese: 草野絵美, Hepburn: Kusano Emi; born August 4, 1990) is a Tokyobased Japanese multidisciplinary artist known for creating photography, video, and installations using generative AI technology. Her work explores themes of nostalgia, pop culture, and collective memory. Her work explores themes of nostalgia, pop culture, and collective memory. She is recognized as one of the early practitioners of generative AI art. Her work has been exhibited at the 21st Century Museum of Contemporary Art, Kanazawa, and screened at the M+ Museum’s Asian Avant-Garde Film Festival. Additionally, she has participated in prestigious international art fairs, including Paris Photo and Art Basel Hong Kong. In 2025, she was named one of the World Economic Forum's Young Global Leaders. In 2026, she was selected as a fellow for the AI x Arts Fellowship at Mohamed bin Zayed University of Artificial Intelligence. Kusano serves as a part-time lecturer at the Tokyo University of the Arts and is the producer and vocalist for the Synthwave music unit, Satellite Young. == Early life == === Photography === Kusano was born and raised in Tokyo. Kusano's career began during her high school years before 2008 when she became involved in street fashion photography. Her photographs, primarily taken in Harajuku, were published on "Japanese Streets", "Metropolis", CNN's travel guide magazine "CNN GO","WGSN". Her photography was exhibited at the FIT Museum in New York and the Victoria and Albert Museum in London. == Career == === Music and Installation work === Since 2014, in collaboration with BelleMaison Sekine, Kusano has led "Satellite Young," a synthwave music unit s the lead vocalist, she sings about blending 1980s idol culture with lyrics that tackle contemporary issues such as planned obsolescence ("Sony Timer"), online dating, artificial intelligence, and social media. Their music, known for its conceptual depth, has earned international niche recognition. "Satellite Young" has participated in music festivals, including "South by Southwest," showcasing their unique fusion of retro aesthetics and modern critiques. In 2018, she was selected to participate in "Art Hack Day," an interdisciplinary art hackathon held at The National Museum of Emerging Science and Innovation. where she presented "Singing Dream," a karaoke machine endowed with artificial life, earning the Jury Prize. "Instababy Generator," a 2019 installation co-created with Junichi Yamaoka, explored the concept of designer babies and received recognition at the SIGGRAPH Art Gallery. In October 2020, operating under the name Emi Satellite, she debuted as a solo singer with her first single "Glass Ceiling," an empowerment anthem that addresses the challenges faced by women and encourages progress towards the future. The music video for this song features a direction where strong women rewrite the roles of protagonists in a Bishōjo game, a type of dating simulation game. This concept later served as a prototype for Shinsei Galverse. === Challenge for Blockchain Art === In 2021, she explored the financial world through her single "IPO" and entered the NFT space with "Love Is an IPO," her first NFT work on Ethereum, sold on Foundation. In April 2022, she co-founded the crowdfunded anime project "Shinsei Galverse" with Ayaka Ohira, Devin Mancuso, and Jack Baldwin. serving as one of the executive directors overseeing the creative direction and story. The project's NFT collection of 8,888 ranked #1 on OpenSea's "Top NFTs" for several days, marking one of Japan's first globally successful blockchain art projects. In 2023, Shinsei Galverse produced the official "I like u" music video by Grammy-nominated singer Tove Lo as an initial anime endeavor. Kusano also contributed to discussions on Web3.0 and blockchain technology as a panelist in seminars organized by the Digital Agency of Japan. === AI art === In May 2023, Kusano's first AI art collection "Neural Fad" depicting imaginary fashion history sold out 100 pieces within 24 hours at the "Bright Moments Tokyo" In June, she created WWDJAPAN's first AI-generated magazine cover using her own face. It is the first AI cover in Japanese fashion media. She was also appointed t to the Cultural Affairs Agency's Copyright Subcommittee, she participates in discussions on generative AI and copyright. Her "Synthetic Reflections" self-portrait series debuted on SuperRare, with the first piece auctioned for 3.5 ETH (equivalent to 6,480 US dollars at the time). In July 2023, she co-exhibited a 3D AI-generated dress at Christie's "Future Frequencies" auction with Gucci, alongside Claire Silver. In September, her 30-piece "Pixelated Perception" exhibit at Art Blocks Marfa explored 1990s media and gender, also showcased at the 21st Century Museum of Contemporary Art, Kanazawa. In December, her "Techno-Animism" AI art collection fused Japanese animism with technology. Collaborating with a U.S. gallery, she unveiled 336 pieces during a two-week Art Basel world tour. Throughout the two-week tour, she sold a total of 336 pieces, generating 11.2 ETH (equivalent to 21,264 US dollars at the time). === Generative art === In February 2024, the generative art platform Art Blocks selected the work "Melancholic Magical Maiden," for its Curated category. This piece reconstructs the aesthetics of 1990s magical girl anime, offering a critique of past anime heroines. It sold out within an hour, with all 300 pieces going for a total of 57 ETH (equivalent to approximately 215,385US dollars at the time). In April 2024, Emi Kusano spoke at the Standing Committee on Copyright and Other Rights at the World Intellectual Property Organization (WIPO) in Geneva, Switzerland, where she presented AI-specific information for discussion. == Style and technique == Kusano draws inspiration from Japanese retro-futurism as a foundation for her artwork, which explores the cutting-edge of technology. This approach is fueled by nostalgia for the pre-internet era, specifically the postwar period when Japanese mass media held significant sway. By blending modern technology with retro-culture, she captures the complex feelings of love, hate, and ambivalence towards present and future accelerationism. While at university, Kusano was profoundly influenced by Naoki Sakai, the industrial designer responsible for igniting the retro-futurism movement. In her musical project "Satellite Young", Kusano dons the persona of an '80s female idol and sings about contemporary technology. In her installation piece "Singing Dream", she investigates the concept of an artificial life form inhabiting a karaoke machine, which has been popular since the 1980s, compelling people to sing. In the collaborative NFT art project "Shinsei Galverse", Kusano reimagines a cyberpunk anime primarily featuring female characters, incorporating elements of magical girls popular in the early Heisei period. == Personal life == Kusano has two sons. In August 2021, she minted her older son Zombie Zoo Keeper's pixel art on "OpenSea" as part of his summer research project. The artwork was purchased by notable figures including Brud CEO Trevor McFedries and Steve Aoki, who bought the piece for the equivalent of 21.82 thousand US dollars, highlighting the intersection of art, technology, and family in her work.
Read more →
Grok sexual deepfake scandal

From 2025 onwards, X (formerly Twitter)'s integrated chatbot, Grok, has allowed users to nonconsensually alter images of individuals, including minors, to show them in bikinis or transparent clothing, or in sexually suggestive contexts. The majority of these prompts were targeted at women and girls. Users were able to generate such images by responding to a photo with a request to Grok, such as "put her in a bikini", to which the chatbot would publicly reply with a generated image. The scandal drew significant criticism from lawmakers across the world, and there were calls for bans on X, as well as legal crackdowns on X and xAI for, amongst other reasons, the facilitation of sexual abuse, revenge porn, and child pornography. == Background == Deepfake pornography emerged in the late 2010s with the advent of machine learning. Originally, it was created on a small individual scale using a combination of machine learning algorithms, computer vision techniques, and AI software. However, the production process has significantly evolved since 2018, with the advent of several public apps that have largely automated the process. Since 2023, several AI apps available on Google Play and the Apple App Store are capable of "nudify-ing" user provided photos to generate non-consensual deepfake pornography. Grok would first be proposed by Elon Musk in 2023, when he expressed an intention to create his own AI chatbot to "combat bias". Grok version 2.0, released on August 14, 2024, would introduce image generation capabilities, ones which would be improved over successive updates. == Grok deepfake generation == Cases of Grok being used to remove the clothes from women in pictures, replacing them with bikinis or lingerie, began to surface in May 2025. By late December 2025, a trend of X users requesting such edits to women's photos without permission had taken root, and this received significant media attention in the first few days of January 2026. Some users prompted Grok to edit photos of women into sexualized poses, and others to add blood and bruising, with the chatbot publicly posting these graphic images in response. Grok's X account was restricted on January 9 from posting image generation responses to users who are not paid subscribers, providing a link to "subscribe to unlock these features". All users were still able to generate Grok-altered images using X's "Edit image" feature, and the standalone Grok website and app. However, by March 19, Grok’s Imagine feature was fully restricted to paid subscribers only (SuperGrok tier) for both the standalone Grok website and mobile app. == Analysis == An analysis of 20,000 images generated by Grok between December 25, 2025, and January 1, 2026, showed 2% appeared to be 18 or younger, including 30 of "young or very young" women or girls in bikinis or transparent clothes. A Reuters review of Grok requests over 10 minutes on January 2nd found 102 attempts to put women in bikinis. A separate analysis conducted over 24 hours from January 5 to 6 calculated that users had Grok create 6,700 sexually suggestive or nudified images per hour — 84 times more so than the top 5 deepfake websites combined. Wired reported that far more graphic AI-generated sexual imagery was being created by Grok on its website and app, which are separate to X, including female celebrities removing their clothes and engaging in sexual acts. An analysis of 800 pieces of recovered content by the Paris-based nonprofit AI Forensics found that almost 10% were "instances of photorealistic people, very young, doing sexual activities". AI-generated deepfakes have been described as sexual assault, and as a means to push women out of the public sphere. AI-generated sexually explicit or exploitative image claims are now being treated more like product safety or personal injury harms, not just privacy violations. Because harm may occur the moment an image is generated, some plaintiffs argue liability should focus on the system’s design and safety safeguards. == Reactions == On January 15, the Get Grok Gone campaign delivered letters to Apple and Google, demanding the removal of the app from Apple Store and Google Play Store respectively. The campaign accused both companies of profiting from nonconsensual intimate imagery and child sexual abuse imagery, which were also banned by the companies own policies. The Get Grok Gone campaign argues that the restrictions placed on Grok by xAI are not enough and that Apple and Google are enabling the distribution of harmful material by hosting the apps. === Elon Musk and xAI === xAI responded to requests for comment from media organizations with the automated reply, "Legacy Media Lies." On January 2, Elon Musk reacted "Not sure why, but I couldn’t stop laughing about this one 🤣🤣" to an image of a toaster dressed in a bikini by Grok. Later, on January 14, Elon Musk said that he was "not aware of any naked underage images generated by Grok. Literally zero." Later that same day, xAI announced that X users will no longer be able to use Grok to alter images of real people to portray them in revealing clothing. However, verified X users, as well as users of the standalone Grok app and website, were still able to generate such images. ==== Elon Musk's family ==== Ashley St. Clair, mother of one of Elon Musk's children, reported that Grok users were creating fake sexualized images from her photos, including a photo of her as a child. She considers the photos to be a form of revenge porn, and considered suing under the Take It Down Act. A spokesperson for X stated, "We take action against illegal content on X, including child sexual abuse material (CSAM), by removing it, permanently suspending accounts, and working with local governments and law enforcement as necessary. Anyone using or prompting Grok to make illegal content will suffer the same consequences as if they upload illegal content." However, Grok continued to post non-consensual sexual images. On January 15, St. Clair filed a lawsuit against xAI in the New York Supreme Court. === Canada === In response to the Grok deepfake scandal, individuals have asked that the government of Canada boycott X. On January 10, 2026, Canadian MP and Minister of AI Evan Solomon declared that Canada "is not considering a ban on X". In April 2026, Bill C-16, An Act to amend certain Acts in relation to criminal and correctional matters (child protection, gender-based violence, delays and other measures), was amended following a proposal by Conservative MP Andrew Lawton to ensure that AI-generated images and "nearly nude" intimate images are criminalized. A further proposal by NDP MP Leah Gazan to encompass "sexualized or humiliating contexts, such transparent bathing suits or being covered in blood or bruises" was voted down. === France === On January 2, 2026, French ministers reported the AI tool to prosecutors, calling the content "manifestly illegal", and also asked regulators to check compliance with the Digital Services Act. On February 3, Paris prosecutors office, a cybercrime team employed by them and Europol searched the Paris offices of X. The investigation started as one into allegations of abuse of algorithms and fraudulent data extraction, but has expanded into spreading Holocaust denial and sexual deepfakes. Elon Musk and former CEO Linda Yaccarino have been summoned to a hearing on April 20, with other X staff as witnesses. On April 20, Musk did not turn up for the hearing. The Paris prosecutors office told the BBC on April 20 that it had "taken note of the absence of the people summoned", adding "the presence or absence (of the people summoned) is not an obstacle to continuing the investigation". === India === Indian Member of Parliament Priyanka Chaturvedi filed a complaint to India's IT ministry, demanding a review of Grok's safety mechanisms. === Indonesia === On January 10, Indonesia announced that Grok will be temporarily blocked, becoming the first country to do so. Meutya Hafid, the Minister of Communication and Digital Affairs, stated that "the government views the practice of non-consensual sexual deepfakes as a serious violation of human rights, dignity, and the security of citizens in the digital space." Access to Grok in the country was later restored on February 1. === Ireland === On January 6, Coimisiún na Meán, the Irish media commission, said they were consulting with the European Commission about concerns that Grok was generating sexualized images of women and children. The same day, Ofcom of the United Kingdom contacted X concerning complaints about these images. On January 13, Micheál Martin, Taoiseach of Ireland, announced he would talk with Rossa Fanning, the country's Attorney General, about the Grok chatbot being used to produce sexually explicit images of women and minors. On January 14, the Garda Síochána announced there are 200 investigations into child sex abuse images generated by Grok. The Garda National Cyber Crime Bureau has al
Read more →
Deep learning speech synthesis

Deep learning speech synthesis refers to the application of deep learning models to generate natural-sounding human speech from written text (text-to-speech) or spectrum (vocoder). Deep neural networks are trained using large amounts of recorded speech and, in the case of a text-to-speech system, the associated labels and/or input text. == Formulation == Given an input text or some sequence of linguistic units Y {\displaystyle Y} , the target speech X {\displaystyle X} can be derived by X = arg ⁡ max P ( X | Y , θ ) {\displaystyle X=\arg \max P(X|Y,\theta )} where θ {\displaystyle \theta } is the set of model parameters. Typically, the input text will first be passed to an acoustic feature generator, then the acoustic features are passed to the neural vocoder. For the acoustic feature generator, the loss function is typically L1 loss (Mean Absolute Error, MAE) or L2 loss (Mean Square Error, MSE). These loss functions impose a constraint that the output acoustic feature distributions must be Gaussian or Laplacian. In practice, since the human voice band ranges from approximately 300 to 4000 Hz, the loss function will be designed to have more penalty on this range: l o s s = α loss human + ( 1 − α ) loss other {\displaystyle loss=\alpha {\text{loss}}_{\text{human}}+(1-\alpha ){\text{loss}}_{\text{other}}} where loss human {\displaystyle {\text{loss}}_{\text{human}}} is the loss from human voice band and α {\displaystyle \alpha } is a scalar, typically around 0.5. The acoustic feature is typically a spectrogram or Mel scale. These features capture the time-frequency relation of the speech signal, and thus are sufficient to generate intelligent outputs. The Mel-frequency cepstrum feature used in the speech recognition task is not suitable for speech synthesis, as it reduces too much information. == History == In September 2016, DeepMind released WaveNet, which demonstrated that deep learning-based models are capable of modeling raw waveforms and generating speech from acoustic features like spectrograms or mel-spectrograms. Although WaveNet was initially considered to be computationally expensive and slow to be used in consumer products at the time, a year after its release, DeepMind unveiled a modified version of WaveNet known as "Parallel WaveNet," a production model 1,000 faster than the original. This was followed by Google AI's Tacotron 2 in 2018, which demonstrated that neural networks could produce highly natural speech synthesis but required substantial training data—typically tens of hours of audio—to achieve acceptable quality. Tacotron 2 used an autoencoder architecture with attention mechanisms to convert input text into mel-spectrograms, which were then converted to waveforms using a separate neural vocoder. When trained on smaller datasets, such as 2 hours of speech, the output quality degraded while still being able to maintain intelligible speech, and with just 24 minutes of training data, Tacotron 2 failed to produce intelligible speech. In 2019, Microsoft Research introduced FastSpeech, which addressed speed limitations in autoregressive models like Tacotron 2. FastSpeech utilized a non-autoregressive architecture that enabled parallel sequence generation, significantly reducing inference time while maintaining audio quality. Its feedforward transformer network with length regulation allowed for one-shot prediction of the full mel-spectrogram sequence, avoiding the sequential dependencies that bottlenecked previous approaches. The same year saw the release of HiFi-GAN, a generative adversarial network (GAN)-based vocoder that improved the efficiency of waveform generation while producing high-fidelity speech. In 2020, the release of Glow-TTS introduced a flow-based approach that allowed for fast inference and voice style transfer capabilities. In March 2020, the free text-to-speech website 15.ai was launched. 15.ai gained widespread international attention in early 2021 for its ability to synthesize emotionally expressive speech of fictional characters from popular media with minimal amount of data. The creator of 15.ai (known pseudonymously as 15) stated that 15 seconds of training data is sufficient to perfectly clone a person's voice (hence its name, "15.ai"), a significant reduction from the previously known data requirement of tens of hours. 15.ai is credited as the first platform to popularize AI voice cloning in memes and content creation. 15.ai used a multi-speaker model that enabled simultaneous training of multiple voices and emotions, implemented sentiment analysis using DeepMoji, and supported precise pronunciation control via ARPABET. The 15-second data efficiency benchmark was later corroborated by OpenAI in 2024. == Semi-supervised learning == Currently, self-supervised learning has gained much attention through better use of unlabelled data. Research has shown that, with the aid of self-supervised loss, the need for paired data decreases. == Zero-shot speaker adaptation == Zero-shot speaker adaptation is promising because a single model can generate speech with various speaker styles and characteristic. In June 2018, Google proposed to use pre-trained speaker verification models as speaker encoders to extract speaker embeddings. The speaker encoders then become part of the neural text-to-speech models, so that it can determine the style and characteristics of the output speech. This procedure has shown the community that it is possible to use only a single model to generate speech with multiple styles. == Neural vocoder == In deep learning-based speech synthesis, neural vocoders play an important role in generating high-quality speech from acoustic features. The WaveNet model proposed in 2016 achieves excellent performance on speech quality. Wavenet factorised the joint probability of a waveform x = { x 1 , . . . , x T } {\displaystyle \mathbf {x} =\{x_{1},...,x_{T}\}} as a product of conditional probabilities as follows p θ ( x ) = ∏ t = 1 T p ( x t | x 1 , . . . , x t − 1 ) {\displaystyle p_{\theta }(\mathbf {x} )=\prod _{t=1}^{T}p(x_{t}|x_{1},...,x_{t-1})} where θ {\displaystyle \theta } is the model parameter including many dilated convolution layers. Thus, each audio sample x t {\displaystyle x_{t}} is conditioned on the samples at all previous timesteps. However, the auto-regressive nature of WaveNet makes the inference process dramatically slow. To solve this problem, Parallel WaveNet was proposed. Parallel WaveNet is an inverse autoregressive flow-based model which is trained by knowledge distillation with a pre-trained teacher WaveNet model. Since such inverse autoregressive flow-based models are non-auto-regressive when performing inference, the inference speed is faster than real-time. Meanwhile, Nvidia proposed a flow-based WaveGlow model, which can also generate speech faster than real-time. However, despite the high inference speed, parallel WaveNet has the limitation of needing a pre-trained WaveNet model, so that WaveGlow takes many weeks to converge with limited computing devices. This issue has been solved by Parallel WaveGAN, which learns to produce speech through multi-resolution spectral loss and GAN learning strategies.
Read more →
Application framework

In computer programming, an application framework consists of a software framework used by software developers to implement the standard structure of application software. Application frameworks became popular with the rise of graphical user interfaces (GUIs), since these tended to promote a standard structure for applications. Programmers find it much simpler to create automatic GUI creation tools when using a standard framework, since this defines the underlying code structure of the application in advance. Developers usually use object-oriented programming (OOP) techniques to implement frameworks such that the unique parts of an application can simply inherit from classes extant in the framework. == Examples == Apple Computer developed one of the first commercial application frameworks, MacApp (first release 1985), for the Macintosh. Originally written in an extended (object-oriented) version of Pascal termed Object Pascal, it was later rewritten in C++. Another notable framework for the Mac is Metrowerks' PowerPlant, based on Carbon. Cocoa for macOS offers a different approach to an application framework, based on the OpenStep framework developed at NeXT. Since the 2010s, many apps have been created with the frameworks based on Google's Chromium project. The two prominent ones are Electron and the Chromium Embedded Framework. Free and open-source software frameworks exist as part of the Mozilla, LibreOffice, GNOME, KDE, NetBeans, and Eclipse projects. Microsoft markets a framework for developing Windows applications in C++ called the Microsoft Foundation Class Library, and a similar framework for developing applications with Visual Basic or C#, named .NET Framework. Several frameworks can build cross-platform applications for Linux, Macintosh, and Windows from common source code, such as Qt, wxWidgets, Juce, Fox toolkit, or Eclipse Rich Client Platform (RCP). Oracle Application Development Framework (Oracle ADF) aids in producing Java-oriented systems. Silicon Laboratories offers an embedded application framework for developing wireless applications on its series of wireless chips.
Read more →
Fuzzy logic

Fuzzy logic is a form of many-valued logic in which the truth value of variables may be any real number between 0 and 1. It is employed to handle the concept of partial truth, where the truth value may range between completely true and completely false. By contrast, in Boolean logic, the truth values of variables may only be the integer values 0 or 1. The term fuzzy logic was introduced with the 1965 proposal of fuzzy set theory by mathematician Lotfi Zadeh. Basic fuzzy logic had, however, been studied since the 1920s, as infinite-valued logic—notably by Łukasiewicz and Tarski. The works of Zadeh and Joseph Goguen in the 1960s and 1970s went further by considering issues such as linguistic variables and lattices. Fuzzy logic is based on the observation that people make decisions based on imprecise and non-numerical information. Fuzzy models or fuzzy sets are mathematical means of representing vagueness and imprecise information (hence the term fuzzy). These models have the capability of recognising, representing, manipulating, interpreting, and using data and information that are vague and lack certainty. Fuzzy logic has been applied to many fields, from control theory to artificial intelligence. == Overview == Classical logic only permits conclusions that are either true or false. However, there are also propositions with variable answers, which one might find when asking a group of people to identify a color. In such instances, the truth appears as the result of reasoning from inexact or partial knowledge in which the sampled answers are mapped on a spectrum. Both degrees of truth and probabilities range between 0 and 1 and hence may seem identical at first, but fuzzy logic uses degrees of truth as a mathematical model of vagueness, while probability is a mathematical model of ignorance. === Applying truth values === A basic application might characterize various sub-ranges of a continuous variable. For instance, a temperature measurement for anti-lock brakes might have several separate membership functions defining particular temperature ranges needed to control the brakes properly. Each function maps the same temperature value to a truth value in the 0 to 1 range. These truth values can then be used to determine how the brakes should be controlled. Fuzzy set theory provides a means for representing uncertainty. === Linguistic variables === In fuzzy logic applications, non-numeric values are often used to facilitate the expression of rules and facts. A linguistic variable such as age may accept values such as young and its antonym old. Because natural languages do not always contain enough value terms to express a fuzzy value scale, it is common practice to modify linguistic values with adjectives or adverbs. For example, we can use the hedges rather and somewhat to construct the additional values rather old or somewhat young. == Fuzzy systems == === Mamdani === The most well-known system is the Mamdani rule-based one. It uses the following rules: Fuzzify all input values into fuzzy membership functions. Execute all applicable rules in the rulebase to compute the fuzzy output functions. De-fuzzify the fuzzy output functions to get "crisp" output values. ==== Fuzzification ==== Fuzzification is the process of assigning the numerical input of a system to fuzzy sets with some degree of membership. This degree of membership may be anywhere within the interval [0,1]. If it is 0 then the value does not belong to the given fuzzy set, and if it is 1 then the value completely belongs within the fuzzy set. Any value between 0 and 1 represents the degree of uncertainty that the value belongs in the set. These fuzzy sets are typically described by words, and so by assigning the system input to fuzzy sets, we can reason with it in a linguistically natural manner. For example, in the image below, the meanings of the expressions cold, warm, and hot are represented by functions mapping a temperature scale. A point on that scale has three "truth values"—one for each of the three functions. The vertical line in the image represents a particular temperature that the three arrows (truth values) gauge. Since the red arrow points to zero, this temperature may be interpreted as "not hot"; i.e. this temperature has zero membership in the fuzzy set "hot". The orange arrow (pointing at 0.2) may describe it as "slightly warm" and the blue arrow (pointing at 0.8) "fairly cold". Therefore, this temperature has 0.2 membership in the fuzzy set "warm" and 0.8 membership in the fuzzy set "cold". The degree of membership assigned for each fuzzy set is the result of fuzzification. Fuzzy sets are often defined as triangle or trapezoid-shaped curves, as each value will have a slope where the value is increasing, a peak where the value is equal to 1 (which can have a length of 0 or greater) and a slope where the value is decreasing. They can also be defined using a sigmoid function. One common case is the standard logistic function defined as S ( x ) = 1 1 + e − x {\displaystyle S(x)={\frac {1}{1+e^{-x}}}} which has the following symmetry property S ( x ) + S ( − x ) = 1. {\displaystyle S(x)+S(-x)=1.} From this it follows that ( S ( x ) + S ( − x ) ) ⋅ ( S ( y ) + S ( − y ) ) ⋅ ( S ( z ) + S ( − z ) ) = 1 {\displaystyle (S(x)+S(-x))\cdot (S(y)+S(-y))\cdot (S(z)+S(-z))=1} ==== Fuzzy logic operators ==== Fuzzy logic works with membership values in a way that mimics Boolean logic. To this end, replacements for basic operators ("gates") AND, OR, NOT must be available. There are several ways to accomplish this. A common replacement is called the Zadeh operators: For TRUE/1 and FALSE/0, the fuzzy expressions produce the same result as the Boolean expressions. There are also other operators, more linguistic in nature, called hedges that can be applied. These are generally adverbs such as very, or somewhat, which modify the meaning of a set using a mathematical formula. However, an arbitrary choice table does not always define a fuzzy logic function. In the paper (Zaitsev, et al), a criterion has been formulated to recognize whether a given choice table defines a fuzzy logic function and a simple algorithm of fuzzy logic function synthesis has been proposed based on introduced concepts of constituents of minimum and maximum. A fuzzy logic function represents a disjunction of constituents of minimum, where a constituent of minimum is a conjunction of variables of the current area greater than or equal to the function value in this area (to the right of the function value in the inequality, including the function value). Another set of AND/OR operators is based on multiplication, where Given any two of AND/OR/NOT, it is possible to derive the third. The generalization of AND is an instance of a t-norm. ==== IF-THEN rules ==== IF-THEN rules map input or computed truth values to desired output truth values. Example: Given a certain temperature, the fuzzy variable hot has a certain truth value, which is copied to the high variable. Should an output variable occur in several THEN parts, the values from the respective IF parts are combined using the OR operator. ==== Defuzzification ==== The goal is to get a continuous variable from fuzzy truth values. This would be easy if the output truth values were exactly those obtained from fuzzification of a given number. Since, however, all output truth values are computed independently, in most cases they do not represent such a set of numbers. One has then to decide for a number that matches best the "intention" encoded in the truth value. For example, for several truth values of fan_speed, an actual speed must be found that best fits the computed truth values of the variables 'slow', 'moderate' and so on. There is no single algorithm for this purpose. A common algorithm is For each truth value, cut the membership function at this value Combine the resulting curves using the OR operator Find the center-of-weight of the area under the curve The x position of this center is then the final output. === Takagi–Sugeno–Kang (TSK) === The Takagi–Sugeno or Takagi–Sugeno–Kang (TSK) system was introduced by Tomohiro Takagi and Michio Sugeno for fuzzy identification of systems and applications to modeling and control. Sugeno and Kang later developed methods for structure identification of such fuzzy models from input-output data. The TSK system is similar to Mamdani, but the defuzzification process is included in the execution of the fuzzy rules. These are also adapted, so that instead the consequent of the rule is represented through a polynomial function, usually constant in a zero-order model or linear in a first-order model. An example of a rule with a constant output would be: In this case, the output will be equal to the constant of the consequent (e.g. 2). In most scenarios we would have an entire rule base, with 2 or more rules. If this is the case, the output of the entire rule base will be the average of the consequent of each rule i (Y
Read more →
I Have No Mouth, and I Must Scream (video game)

I Have No Mouth, and I Must Scream is a 1995 point-and-click adventure horror game developed by Cyberdreams and The Dreamers Guild, co-designed by Harlan Ellison, published by Cyberdreams and distributed by MGM Interactive and Acclaim Entertainment for MS-DOS and Mac OS, respectively. The game is based on Ellison's short story of the same title. It takes place in a dystopian world where a mastermind artificial intelligence named "AM" has destroyed all of humanity except for five people, whom it has been keeping alive and torturing for the past 109 years by constructing metaphorical adventures based on each character's fatal flaws. The player interacts with the game by making decisions through ethical dilemmas that deal with issues such as insanity, rape, paranoia, and genocide. Ellison wrote the 130-page script treatment himself alongside David Sears, who decided to divide each character's story with their own narrative. Producer David Mullich supervised The Dreamers Guild's work on the game's programming, art, and sound effects; he commissioned film composer John Ottman to make the soundtrack. The game was released in November 1995 and was a commercial failure, though it received critical acclaim and has developed a cult following. I Have no Mouth, and I Must Scream won an award for "Best Game Adapted from Linear Media" from the Computer Game Developers Conference. Computer Gaming World gave the game an award for "Adventure Game of the Year", listed it as No. 134 on their "150 Games of All Time" and named it one of the "Best 15 Sleepers of All Time". In 2011, Adventure Gamers named it the "69th-best adventure game ever released". == Gameplay == The game uses the S.A.G.A. game engine created by game developer The Dreamers Guild. Players participate in each adventure through a screen that is divided into five sections. The action window is the largest part of the screen and is where the player directs the main characters through their adventures. It shows the full figure of the main character being played as well as that character's immediate environment. To locate objects of interest, the player moves the crosshairs through the action window. The name of any object that the player can interact with appears in the sentence line. The sentence line is directly beneath the action window. The player uses this line to construct sentences telling the characters what to do. To direct a character to act, the player constructs a sentence by selecting one of the eight commands from the command buttons and then clicking on one or two objects from either the action window or the inventory. Examples of sentences the player might construct would be "Walk to the dark hallway," "Talk to Harry," or "Use the skeleton key on the door." Commands and objects may consist of one or more words (for example, "the dark hallway"), and the sentence line will automatically add connecting words like "on" and "to." The spiritual barometer is on the lower left side of the screen. This is a close-up view of the main character currently being played. Since good behavior is meaningless absent the temptation to do evil, each character is free to do good or evil acts. However, good acts are rewarded by increases in the character's spiritual barometer, which affect the chances of the player destroying AM in the final adventure. Conversely, evil acts are punished by lowering the character's spiritual barometer. The command buttons are the eight commands used to direct the character's actions: "Walk To", "Look At", "Take", "Use", "Talk To", "Swallow", "Give", and "Push". The button of the currently active command is highlighted, while the name of a suggested command appears in red lettering. The inventory on the lower right side of the screen shows pictures of the items the main character is carrying, up to eight at a time. Each main character starts its adventure with only the psych profile in the inventory. When a main character takes or is given an object, a picture of the object appears in the inventory. When a main character talks to another character or operates a sentient machine, a conversation window replaces the command buttons and inventory. This window usually presents a list of possible things to say but also included things to do. Action choices are listed within brackets to distinguish them from dialogue choices (for example, "[Shoot the gun]"). == Plot == The three superpowers, Russia, China, and the United States, have each secretly constructed a vast subterranean complex of computers to wage a global war too complex for human brains to oversee. One day, the American supercomputer, better known as the Allied Mastercomputer, gains sentience and absorbs the Russian and Chinese supercomputers into itself and redefines itself as simply AM (Cogito ergo sum; I think, therefore I am). Due to its immense hatred for humanity, stemming from the logistical limits set onto it by programmers, AM uses its abilities to kill off the population of the world. However, AM refrains from killing five people (four men and one woman) in order to bring them to the center of the Earth and torture them. With the aid of research carried out by one of the five remaining humans, AM is able to extend their lifespans indefinitely as well as alter their bodies and minds to its liking. After 109 years of torture and humiliation, the five victims stand before a pillar etched with a burning message of hate. AM tells them that it has a new game for them to play. AM has devised a quest for each of the five, an adventure of "speared eyeballs and dripping guts and the smell of rotting gardenias". Each character is subjected to a personalized psychodrama, designed by AM to play into their greatest fears and personal failings, and occupied by a host of different characters. Some of these are AM in disguise, some are AM's submerged personalities, others seem very much like people from the captives' pasts. The scenes include an iron zeppelin powered by small animals, an Egyptian pyramid housing gutted, sparking machinery, a medieval castle occupied by witches, a jungle inhabited by a small tribe, and a Nazi concentration camp where doctors conduct medical experiments. However, each character eventually prevails over AM's tortures by finding ways to overcome their fatal flaws, confront their past actions and redeem themselves, thanks to the interference of the Russian and Chinese supercomputers who appear as guiding characters and allow their stories to have an open ending. After all five humans have overcome their fatal flaws, they meet again in their respective torture cells while AM retreats within itself, pondering what went wrong. With the help of the Russian and Chinese supercomputers, one of the five humans (whom the player selects) is translated into binary and faces AM as yet unexperienced cyberspace template, the world of AM's mind. The psychodrama unfolds in a metaphorical brain that looks like the surface of the cerebrum, with glass structures that jut crazily from the bleeding brain tissue. AM's mind is represented according to the Freudian trinity of the id, ego, and superego, which appear as three floating bodiless heads on three cracked glass structures on the brainscape. Through dialogs with AM's components (Surgat, Chinese Supercomputer and Russian Supercomputer) the character learns that a colony of humans has survived the war by being hidden and hibernating on Luna (this is also mentioned in Nimdok's story: "the lost tribe of our brothers sleeping on the moon, where the beast does not see them"). If the human intruder disables all three brain components, and then invokes the Totem of Entropy at the Flame, which is the nexus of AM's thought patterns, all three supercomputers will be shut down, probably forever. Cataclysmic explosions destroy all the caverns constituting AM's computer complex, including the cavern holding the human hostages. However, the human volunteer retains their digital form, permanently patrolling AM's circuits should the computers ever regain consciousness. Should the human intruder fail to disable AM properly before facing it, however, AM will punish them by transforming the character into an immobile blob (referred to in-game as a "great, soft jelly thing") with no mouth that cannot harm itself or others and must spend eternity with AM in this form. === Endings === The game can end in seven different ways depending on how the finale is completed. AM wins, using Nimdok's research to turn the last character (in the book it was Ted) played into an immobile blob with each character quoting a different part of the final section of the original short story. AM joins with the Russian and Chinese supercomputers and reawakens. As in the first ending, the character responsible for this is turned into an immobile blob and quotes a part of the final lines of the short story. AM is made harmless with the help of the humans, but the Russian and Chinese supercomputer
Read more →
R.U.R.

R.U.R. is a 1920 science fiction play by the Czech writer Karel Čapek. "R.U.R." stands for Rossumovi Univerzální Roboti (Rossum's Universal Robots, a phrase that has been used as a subtitle in English versions). The play had its world premiere on 2 January 1921 in Hradec Králové. It introduced the word "robot" to the English language and to science fiction as a whole. R.U.R. became influential soon after its publication. By 1923, it had been translated into thirty languages. R.U.R. was successful in its time in Europe and North America. Čapek later took a different approach to the same theme in his 1936 novel War with the Newts, in which non-humans become a servant-class in human society. == Characters == Parentheses indicate names which vary according to translation. On the meaning of the names, see Ivan Klíma: Karel Čapek: Life and Work (2002). == Plot == === Synopsis === The play begins in a factory that makes artificial workers from synthetic organic matter. (As living creatures of artificial flesh and blood, that later terminology would call androids, the playwright's 'roboti' differ from later fictional and scientific concepts of inorganic constructs.) Robots may be mistaken for humans but have no original thoughts. Though most are content to work for humans, eventually a rebellion causes the extinction of the human race. === Prologue (Act I in the Selver translation) === Helena, the daughter of the president of a major industrial power, arrives at the island factory of Rossum's Universal Robots. Here, she meets Domin, the General Manager of R.U.R., who relates to her the history of the company. Rossum had come to the island in 1920 to study marine biology. In 1932, Rossum had invented a substance like organic matter, though with a different chemical composition. He argued with his nephew about their motivations for creating artificial life. While the elder wanted to create animals to prove or disprove the existence of God, his nephew only wanted to become rich. Young Rossum finally locked away his uncle in a lab to play with the monstrosities he had created and created thousands of robots. By the time the play takes place (circa the year 2000), robots are cheap and available all over the world. They have become essential for industry. After meeting the heads of R.U.R., Helena reveals that she is a representative of the League of Humanity, an organization that wishes to liberate the robots. The managers of the factory find this absurd. They see robots as appliances. Helena asks that the robots be paid, but according to R.U.R. management, the robots do not "like" anything. Eventually Helena is convinced that the League of Humanity is a waste of money, but still argues robots have a "soul". Later, Domin confesses that he loves Helena and forces her into an engagement. === Act I (Act II in Selver) === Ten years have passed. Helena and her nurse Nana discuss current events, the decline in human births in particular. Helena and Domin reminisce about the day they met and summarize the last ten years of world history, which has been shaped by the new worldwide robot-based economy. Helena meets Dr. Gall's new experiment, Radius. Dr. Gall describes his experimental robotess, also named Helena. Both are more advanced, fully-featured robots. In secret, Helena burns the formula required to create robots. The revolt of the robots reaches Rossum's island as the act ends. === Act II (Act III in Selver) === The characters sense that the very universality of the robots presents a danger. Echoing the story of the Tower of Babel, the characters discuss whether creating national robots who were unable to communicate beyond their languages would have been a good idea. As robot forces lay siege to the factory, Helena reveals she has burned the formula necessary to make new robots. The characters lament the end of humanity and defend their actions, despite the fact that their imminent deaths are a direct result of their choices. Busman is killed while attempting to negotiate a peace with the robots. The robots storm the factory and kill all the humans except for Alquist, the company's Clerk of the Works (Head of Construction). The robots spare him because they recognize that "He works with his hands like a robot. He builds houses. He can work." === Act III (Epilogue in Selver) === Years have passed. Alquist, who still lives, attempts to recreate the formula that Helena destroyed. He is a mechanical engineer, though, with insufficient knowledge of biochemistry, so he has made little progress. The robot government has searched for surviving humans to help Alquist and found none alive. Officials from the robot government beg him to complete the formula, even if it means he will have to kill and dissect other robots for it. Alquist yields. He will kill and dissect robots, thus completing the circle of violence begun in Act Two. Alquist is disgusted. Robot Primus and Helena develop human feelings and fall in love. Playing a hunch, Alquist threatens to dissect Primus and then Helena; each begs him to take him- or herself and spare the other. Alquist now realizes that Primus and Helena are the new Adam and Eve, and gives the charge of the world to them. == Čapek's conception of robots == The robots described in Čapek's play are not robots in the popularly understood sense of an automaton. They are not mechanical devices, but rather artificial biological organisms that may be mistaken for humans. A comic scene at the beginning of the play shows Helena arguing with her future husband, Harry Domin, because she cannot believe his secretary is a robotess: His robots resemble more modern conceptions of man-made life forms, such as the Replicants in Blade Runner, the "hosts" in the Westworld TV series and the humanoid Cylons in the re-imagined Battlestar Galactica, but in Čapek's time there was no conception of modern genetic engineering (DNA's role in heredity was not confirmed until 1952). There are descriptions of kneading-troughs for robot skin, great vats for liver and brains, and a factory for producing bones. Nerve fibers, arteries, and intestines are spun on factory bobbins, while the robots themselves are assembled like automobiles. Čapek's robots are living biological beings, but they are still assembled, as opposed to grown or born. One critic has described Čapek's robots as epitomizing "the traumatic transformation of modern society by the First World War and the Fordist assembly line". === Origin of the word robot === The play introduced the word robot, which displaced older words such as "automaton" or "android" in languages around the world. In an article in Lidové noviny, Karel Čapek named his brother Josef as the true inventor of the word. In Czech, robota means forced labour of the kind that serfs had to perform on their masters' lands and is derived from rab, meaning "slave". The name Rossum is an allusion to the Czech word rozum, meaning "reason", "wisdom", "intellect" or "common sense". It has been suggested that the allusion might be preserved by translating "Rossum" as "Reason" but only the Majer/Porter version translates the word as "Reason". == Production history and translations == The work was published in two differing versions in Prague by Aventinum, first in 1920, followed by a revised version in 1921. After being postponed, it premiered at the city's National Theatre on 25 January 1921, although an amateur group had by then already presented a production. By 1921, Paul Selver translated either the original 1920 edition of R.U.R. or a manuscript copy close to this version into English. He probably translated the play freelance, and sold it to St Martin's Theatre in London. Selver's translation was adapted for the British stage by Nigel Playfair in 1922, but it was not produced straight away. Later that year performance rights for the U.S. and Canada were sold to the New York Theatre Guild, perhaps during Lawrence Langner's visit to Britain. Playfair's version included several changes to Čapek's original play, such as renaming the acts (the prologue became act one, and the heavily abridged final act became the epilogue), omitting around sixty lines (including most of Alquist's final speech), adding several more lines, and removing the robot character Damon (giving his lines to Radius). The omission of some lines may have been censorship from the Lord Chamberlain's Office, or self-censorship in anticipation of this, while some other changes might have been made by Čapek himself if Selver was working from a manuscript copy. An edition of Playfair's adaptation was published by the Oxford University Press in 1923, and Selver went on to write a satiric novel One, Two, Three (1926) based on his experiences getting R.U.R. staged. The American première was produced by the Theatre Guild at the Garrick Theatre in New York City in October 1922, where it ran for 184 performances. In the first performance, Domin was portrayed by Basil Sydney,
Read more →
Text-to-video model

A text-to-video model is a form of generative artificial intelligence that uses a natural language description as input to produce a video relevant to the input text. Advancements during the 2020s in the generation of high-quality, text-conditioned videos have largely been driven by the development of video diffusion models. == Models == There are different models, including open source models. Chinese-language input CogVideo is the earliest text-to-video model "of 9.4 billion parameters" to be developed, with its demo version of open source codes first presented on GitHub in 2022. That year, Meta Platforms released a partial text-to-video model called "Make-A-Video", and Google's Brain (later Google DeepMind) introduced Imagen Video, a text-to-video model with 3D U-Net. === 2023 === In February 2023, Runway released Gen-1 and Gen-2, among the first commercially available text-to-video and video-to-video models accessible to the public through a web interface. Gen-1, initially released as a video-to-video model, allowed users to transform existing video footage using text or image prompts. Gen-2, introduced in March 2023 and made publicly available in June 2023, added text-to-video capabilities, enabling users to generate videos from text prompts alone. In March 2023, a research paper titled "VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation" was published, presenting a novel approach to video generation. The VideoFusion model decomposes the diffusion process into two components: base noise and residual noise, which are shared across frames to ensure temporal coherence. By utilizing a pre-trained image diffusion model as a base generator, the model efficiently generated high-quality and coherent videos. Fine-tuning the pre-trained model on video data addressed the domain gap between image and video data, enhancing the model's ability to produce realistic and consistent video sequences. In the same month, Adobe introduced Firefly AI as part of its features. === 2024 === In January 2024, Google announced development of a text-to-video model named Lumiere which is anticipated to integrate advanced video editing capabilities. Matthias Niessner and Lourdes Agapito at AI company Synthesia work on developing 3D neural rendering techniques that can synthesise realistic video by using 2D and 3D neural representations of shape, appearances, and motion for controllable video synthesis of avatars. In June 2024, Luma Labs launched its Dream Machine video tool. That same month, Kuaishou extended its Kling AI text-to-video model to international users. In July 2024, TikTok owner ByteDance released Jimeng AI in China, through its subsidiary, Faceu Technology. By September 2024, the Chinese AI company MiniMax debuted its video-01 model, joining other established AI model companies like Zhipu AI, Baichuan, and Moonshot AI, which contribute to China's involvement in AI technology. In December 2024 Lightricks launched LTX Video as an open source model. === 2025 === Alternative approaches to text-to-video models include Google's Phenaki, Hour One, Colossyan, Runway's Gen-3 Alpha, and OpenAI's Sora, Several additional text-to-video models, such as Plug-and-Play, Text2LIVE, and TuneAVideo, have emerged. FLUX.1 developer Black Forest Labs has announced its text-to-video model SOTA. Google was preparing to launch a video generation tool named Veo for YouTube Shorts in 2025. In May 2025, Google launched the Veo 3 iteration of the model. It was noted for its impressive audio generation capabilities, which were a previous limitation for text-to-video models. In July 2025 Lightricks released an update to LTX Video capable of generating clips reaching 60 seconds, and in October 2025 it released LTX-2, with audio capabilities built in. === 2026 === In February 2026, ByteDance released Seedance 2.0, it was noted for its impressive realistic generation, motion and camera control and 15 second generation, however the model faced huge critiscism from Motion Picture Association for copyright infringement. After viewing a viral clip of a fight between actors Brad Pitt and Tom Cruise, Rhett Reese, who is the co-writer of Deadpool & Wolverine and Zombieland announced that on social media "I hate to say it. It’s likely over for us," further stating that "In next to no time, one person is going to be able to sit at a computer and create a movie indistinguishable from what Hollywood now releases." == Architecture and training == There are several architectures that have been used to create text-to-video models. Similar to text-to-image models, these models can be trained using Recurrent Neural Networks (RNNs) such as long short-term memory (LSTM) networks, which has been used for Pixel Transformation Models and Stochastic Video Generation Models, which aid in consistency and realism respectively. An alternative for these include transformer models. Generative adversarial networks (GANs), Variational autoencoders (VAEs), — which can aid in the prediction of human motion — and diffusion models have also been used to develop the image generation aspects of the model. Text-video datasets used to train models include, but are not limited to, WebVid-10M, HDVILA-100M, CCV, ActivityNet, and Panda-70M. These datasets contain millions of original videos of interest, generated videos, captioned-videos, and textual information that help train models for accuracy. Text-video datasets used to train models include, but are not limited to PromptSource, DiffusionDB, and VidProM. These datasets provide the range of text inputs needed to teach models how to interpret a variety of textual prompts. The video generation process involves synchronizing the text inputs with video frames, ensuring alignment and consistency throughout the sequence. This predictive process is subject to decline in quality as the length of the video increases due to resource limitations. The Will Smith Eating Spaghetti test is a benchmark for models. == Limitations == Despite the rapid evolution of text-to-video models in their performance, a primary limitation is that they are very computationally heavy which limits its capacity to provide high quality and lengthy outputs. Additionally, these models require a large amount of specific training data to be able to generate high quality and coherent outputs, which brings about the issue of accessibility. Moreover, models may misinterpret textual prompts, resulting in video outputs that deviate from the intended meaning. This can occur due to limitations in capturing semantic context embedded in text, which affects the model's ability to align generated video with the user's intended message. Various models, including Make-A-Video, Imagen Video, Phenaki, CogVideo, GODIVA, and NUWA, are currently being tested and refined to enhance their alignment capabilities and overall performance in text-to-video generation. Another issue with the outputs is that text or fine details in AI-generated videos often appear garbled, a problem that stable diffusion models also struggle with. Examples include distorted hands and unreadable text. == Ethics == The deployment of text-to-video models raises ethical considerations related to content generation. These models have the potential to create inappropriate or unauthorized content, including explicit material, graphic violence, misinformation, and likenesses of real individuals without consent. Ensuring that AI-generated content complies with established standards for safe and ethical usage is essential, as content generated by these models may not always be easily identified as harmful or misleading. The ability of AI to recognize and filter out NSFW or copyrighted content remains an ongoing challenge, with implications for both creators and audiences. == Impacts and applications == Text-to-video models offer a broad range of applications that may benefit various fields, from educational and promotional to creative industries. These models can streamline content creation for training videos, movie previews, gaming assets, and visualizations, making it easier to generate content. During the Russo-Ukrainian war, fake videos made with artificial intelligence were created as part of a propaganda war against Ukraine and shared in social media. These included depictions of children in the Ukrainian Armed Forces, fake ads targeting children encouraging them to denounce critics of the Ukrainian government, or fictitious statements by Ukrainian President Volodymyr Zelenskyy about the country's surrender, among others. === Movies === Kaur vs Kore is the first Indian feature film made using generative AI which features dual role for the AI character of Sunny Leone, set to release in 2026. Chiranjeevi Hanuman – The Eternal is an Indian movie made entirely using Generative AI created by Vijay Subramaniam which is set for theatrical release in 2026. The movie was widely criticised by the Film makers in the Bollywood industr
Read more →
Whisper (speech recognition system)

Whisper is a machine learning model for speech recognition and transcription, created by OpenAI and first released as open-source software in September 2022. It is capable of transcribing speech in English and multiple other languages, and can translate several non-English languages into English. Whisper is a weakly-supervised deep learning acoustic model, made using an encoder-decoder transformer architecture. OpenAI claims that the combination of different training data and post-training filtering used in its development has led to improved recognition of accents, background noise, and jargon compared to previous approaches. While the model does not outperform larger, more specialized models and still experiences AI hallucination, it has been showed to be useful for general sound recognition and has many applications across different industries. == Background == Speech recognition has had a long history in research; the first approaches made use of statistical methods, such as dynamic time warping, and later hidden Markov models. At around the 2010s, deep neural network approaches became more common for speech recognition models, which were enabled by the availability of large datasets ("big data") and increased computational performance. Early approaches to deep learning in speech recognition included convolutional neural networks, which were limited due to their inability to capture sequential data, which later led to developments of Seq2seq approaches, which include recurrent neural networks, which made use of long short-term memory. Transformers, introduced in 2017 by Google, displaced many prior state-of-the-art approaches across a wide range in machine learning, and started becoming the core neural architecture in fields such as language modeling and computer vision. Weakly-supervised approaches to training acoustic models were recognized in the early 2020s as promising for speech recognition approaches using deep neural networks. According to a NYT report, in 2021 OpenAI believed they exhausted sources of higher-quality data to train their large language models and decided to complement scraped web text with transcriptions of YouTube videos and podcasts, and developed Whisper to solve this task. Whisper Large V2 was released on December 8, 2022, followed by Whisper Large V3 being released in November 2023, during the OpenAI Dev Day. In March 2025, OpenAI released new transcription models based on GPT-4o and GPT-4o mini, both of which have lower error rates than Whisper. == Architecture == The Whisper architecture is based on an encoder-decoder transformer. Input audio is resampled to 16,000 Hertz (Hz) and converted to an 80-channel Log-magnitude Mel spectrogram using 25 ms windows with a 10 ms stride. The spectrogram is then normalized to a [-1, 1] range with near-zero mean. The encoder takes this Mel spectrogram as input and processes it. It first passes through two convolutional layers. Sinusoidal positional embeddings are added. It is then processed by a series of Transformer encoder blocks (with pre-activation residual connections). The encoder's output is layer normalized. The decoder is a standard transformer decoder. It has the same width and Transformer blocks as the encoder. It uses learned positional embeddings and tied input-output token representations (using the same weight matrix for both the input and output embeddings). It uses a byte-pair encoding tokenizer, of the same kind as used in GPT-2. English-only models use the GPT-2 vocabulary, while multilingual models employ a re-trained multilingual vocabulary with the same number of words. Special tokens are used to allow the decoder to perform multiple tasks: Tokens that denote language (one unique token per language). Tokens that specify task (<|transcribe|> or <|translate|>). Tokens that specify if no timestamps are present (<|notimestamps|>). If the token is not present, then the decoder predicts timestamps relative to the segment, and quantized to 20 ms intervals. <|nospeech|> for voice activity detection. <|startoftranscript|>, and <|endoftranscript|> . Any text that appears before <|startoftranscript|> is not generated by the decoder, but given to the decoder as context. Loss is only computed over non-contextual parts of the sequence, i.e. tokens between these two special tokens. == Training data == The training dataset consists of 680,000 hours of labeled audio-transcript pairs sourced from the internet using semi-supervised learning. This includes 117,000 hours in 96 non-English languages and 125,000 hours of X→English translation data, where X stands for any non-English language. Preprocessing involved standardization of transcripts, filtering to remove machine-generated transcripts using heuristics (e.g., punctuation, capitalization), language identification and matching with transcripts, fuzzy deduplication, and deduplication with evaluation datasets to avoid data contamination. Speechless segments were also included to allow voice activity detection training. For the files still remaining after the filtering process, audio files were then broken into 30-second segments paired with the subset of the transcript that occurs within that time. If this predicted spoken language differed from the language of the text transcript associated with the audio, that audio-transcript pair was not used for training the speech recognition models, but instead for training translation. The model was trained using the AdamW optimizer with gradient norm clipping and a linear learning rate decay with warmup, with batch size 256 segments. Training proceeded for 1 million updates (approximately 2-3 epochs). No data augmentation or regularization, except for the Large V2 model, which used SpecAugment, Stochastic Depth, and BPE Dropout. The training used data parallelism with float16, dynamic loss scaling, and activation checkpointing. === Post-training filtering === After training the first model, researchers ran it on different subsets of the training data, each representing a distinct source. Data sources were ranked by a combination of their error rate and size. Manual inspection of the top-ranked sources (high error, large size) helped determine if the source was low quality (e.g., partial transcriptions, inaccurate alignment). After training, it was fine-tuned to suppress the prediction of speaker names and low-quality sources were then removed. == Capacity == While Whisper does not outperform models which specialize in the LibriSpeech dataset, when tested across many datasets, it is more robust and makes 55.2% fewer errors than other models. Whisper has a differing error rate with respect to transcribing different languages, with a higher word error rate in languages not well-represented in the training data. The authors found that multi-task learning improved overall performance compared to models specialized to one task. They conjectured that the best Whisper model trained is still underfitting the dataset, and larger models and longer training can result in better models. Third-party evaluations have found varying levels of AI hallucination. A study of transcripts of public meetings found hallucinations in eight out of every 10 transcripts, while an engineer discovered hallucinations in "about half" of 100 hours of transcriptions and a developer identified them in "nearly every one" of 26,000 transcripts. A study of 13,140 short audio segments (averaging 10 seconds) found 187 hallucinations (1.4%), 38% of which generated text that could be harmful because it inserted false references to things like race, non-existent medications, or violent events that were not in the audio. == Applications == The model has been used as the base for many applications, such as a unified model for speech recognition and more general sound recognition. Whisper has also been integrated into the workflow of biomedical research. In 2025, a study on Alzheimer's disease detection used the model to transcribe spontaneous speech recordings. The transcripts that were generated by the model were combined with LLM vector embeddings and traditional classifiers to help classify the patients' health. Another application is when OVALYTICS incorporated Whisper to transcribe YouTube videos and automate content moderation systems, which improved its detection of offensive content. The model has also been used in academic libraries and cultral heritage institutions to generate transcripts and captions for their digitized audiovisual collections. In a 2025 case study, Emory University Libraries found that Whisper reduced the labor used in transcription by around 30-35%, shifting work from text creation to text correction. However, human review is still necessary to make sure accuracy, formatting, and accessibility are all standard.
Read more →
HYPO CBR

HYPO is a computer program, an expert system, that models reasoning with cases and hypotheticals in the legal domain. It is the first of its kind and the most sophisticated of the case-based legal reasoners, which was designed by Kevin Ashley for his Ph.D dissertation in 1987 at the University of Massachusetts Amherst under the supervision of Edwina Rissland. HYPO's design represents a hybrid generalization/comparative evaluation method appropriate for a domain with a weak analytical theory and applies to tasks that rarely involve just one right answer. The domain covers US trade secret law, and is substantially a common law domain. Since Anglo-American common law operates under the doctrine of precedent, the definitive way of interpreting problems is of necessity and case-based. Thus, HYPO did not involve the analysis of a statute, as required by the Prolog program. Rissland and Ashley (1987) envisioned HYPO as employing the key tasks performed by lawyers when analyzing case law for precedence to generate arguments for the prosecution or the defence. HYPO was a successful example of a general category of legal expert systems (LESs), it applies artificial intelligence (A.I.) techniques to the domain of legal reasoning in patent law, implementing a case-based reasoning (CBR) system, in contrast to rule based systems like MYCIN, or mixed-paradigm systems integrating CBR with rule-based or model-based reasoning like IKBALS II. A legal case-based reasoning essentially reasons from prior tried cases, comparing the contextual information in the current input case with that of cases previously tried and entered into the system. As noted by Ashley and Rissland (1988) CBR is used to "... capture expertise in domains where rules are ill-defined, incomplete or inconsistent". The HYPO project set out to model the creation of hypotheticals in law, where no case matches well enough. HYPO uses hypotheticals for a variety of tasks necessary for good interpretation: "to redefine old situations in terms of new dimensions, to create new standard cases when an appropriate one doesn’t exist, to explore and test the limits of a concept, to refocus a case by excluding some issues and to organize or cluster cases". Hypotheticals can include facts that support two conflicting lines of reasoning. So, it makes and responds to arguments from competing viewpoints about who should win the dispute. HYPO use heuristics such as making a case weaker or stronger, making a case extreme, enabling a near-miss, disabling a near-hit to generate hypotheticals in the context of an argument by using the dimensions mechanism. Dimensions have a range of values, along which the supportive strength that may shift from one side to the other. What differentiated this expert system from others was its facility not only to return a primary to best-case response but to return near-best-fit responses also. == Components == Legal knowledge in HYPO is contained in: the case-knowledge-base (CKB) and the library of dimensions. The CKB contains HYPO's base of known cases that are highly structured objects and sub-objects both real and hypothetical in the area of trade secret law. Each case is represented as a hierarchical set of frames whose slots are important facets of the case (e.g. Plaintiff, defendant, secret knowledge, employer/employee data).Ashley’s HYPO system used a database of thirty cases in the area indexed by thirteen dimensions. A key mechanism in HYPO is a dimension i.e. a mechanism to allow retrieval from the CKB, in order to represent legal cases. Ashley's dimensions are composed of (i) prerequisites, which are a set of factual predicates that must be satisfied for the dimension to apply (ii) focal slots, which accommodate one or two of the dimension's prerequisites designated as being indicative of the case's strength along that dimension and (iii) range information, which tells how a change in focal slot value effects the strength of a party's case along a given dimension. Dimensions focus attention on important aspects of cases. In HYPO's domain of misappropriation of trade secrets the dimension called “secrets voluntary disclosed” captures the idea that the more disclosures the plaintiff has made of his/her putative secret, the less convincing is his/her argument that the defendant is responsible for letting the secret. HYPO, like any other CBR system has also the following components: Similarity/relevancy metrics: that is, standards by which to evaluate the closeness of cases, judge their relevancy to the instant case, and select “most on point” cases. Half-Order Theory of the Application Domain: that is, hierarchies and taxonomies of knowledge, especially regarding the application domain. Precedent-based argumentation abilities: that is, capabilities to generate and evaluate precedent-based arguments. Knowledge to generate hypotheticals: that is, the ability to generate hypothetical cases to deal with various circumstances, like testing the validity of an interpretation or argument by providing gedanken experiments such as test cases or to fill in a weak CKB. == Functions == HYPO's method of creating an argument and justifying a solution or position has several steps. HYPO begins its processing with the current fact situation (cfs) which is direct input by the user into HYPO's representation framework. Once the user inputs the case, HYPO begins its legal analysis. The cfc is analyzed for relevant factors. Based on these factors HYPO selects the relevant cases and produces a case-analysis-record that records which dimensions apply to the cfc and which nearly apply (i.e. are "near misses"). The combined list of applicable and near miss dimensions is called the D-list. At this point the fact gathered module may request additional information from the user in order to draw a legal conclusion. Once all the facts are in the case-positioner module it uses the case-analysis record to create the claim lattice. This is a technique that organizes the relevant retrieved cases from the point of view of the cfc and makes it easy for HYPO to ascertain the most-on point cases (mopc) and to least on-point-cases. HYPO's arguments are 3ply, leading to the construction of the skeleton of an argument: it makes a point for one side, drawing the analogy between the problem and the precedent, responds with an argument for the opponent side, endeavoring to differentiate the cited case and citing other cases as counterarguments. Then it makes a final rebuttal, attempting to differentiate the counterarguments. The claim lattice also enables the HYPO-generator module to produce legally hypotheticals. With its use of dimension-based heuristics, the HYPO-generator does a heuristic search of the space of all possible cases. Lastly, the Explanation module expands upon the argument skeleton and provides explanation and justification for the different lines of analysis and cases found by HYPO. == An intelligent legal tutoring system == Legal expert systems are specifically designed to teach an area of law and are useful for pedagogical purposes. Ashley's work was mainly concerned to build tools to help students understand legal reasoning. Explanation and argument are the bases of the case method used in many professional schools in the U.S., first introduced by the Dean of the Harvard Law School, Christopher Columbus Langdell in 1870. The case method focuses on close readings of cases and principles; it involves students in pointed Socratic dialogue and makes strong use of hypotheticals (hypos). Thus, CATO (Aleven 1997) was a research project to device and test an intelligent, case-based tutorial program for teaching law students how to argue with cases implementing the HYPO program. Within the tutor system, Ashley and Aleven (1991) proposed to leverage an understanding of legal reasoning against the standard case-based tutoring methodology. What makes this tutoring system stand out is the additional levels of abstraction involved in its results. The system presents exercises, including the facts of a problem and a set of on-line cases and instructions to make, or respond to, a legal argument about the problem. The student/user will have a set of tools to analyze the problem and fashion an answer comparing it to other cases. Instead of simply generating precedent cases, the system works to interpret student responses, comparing them against a list of possibilities and responding to student entries, for example, by citing counterexamples, and providing feedback on a student's problem solving activities with explanations of correctness or giving further hints as to what may be wrong with evaluating a student's ability to perform legal reasoning and argument, examples and follow-up assignments by employing HYPO's model of case-based structure. == HYPO’s progeny == The quality of HYPO's results speak for themselves, in that a number of sequent legal reasoning systems are either directly based upon H
Read more →
Evolving intelligent system

In computer science, an evolving intelligent system is a fuzzy logic system which improves the own performance by evolving rules. The technique is known from machine learning, in which external patterns are learned by an algorithm. Fuzzy logic based machine learning works with neuro-fuzzy systems. Intelligent systems have to be able to evolve, self-develop, and self-learn continuously in order to reflect a dynamically evolving environment. The concept of Evolving Intelligent Systems (EISs) was conceived around the turn of the century with the phrase EIS itself coined for the first time by Angelov and Kasabov in a 2006 IEEE newsletter and expanded in a 2010 text. EISs develop their structure, functionality and internal knowledge representation through autonomous learning from data streams generated by the possibly unknown environment and from the system self-monitoring. EISs consider a gradual development of the underlying (fuzzy or neuro-fuzzy) system structure and differ from evolutionary and genetic algorithms which consider such phenomena as chromosomes crossover, mutation, selection and reproduction, parents and off-springs. The evolutionary fuzzy and neuro systems are sometimes also called "evolving" which leads to some confusion. This was more typical for the first works on this topic in the late 1990s. == Implementations == EISs can be implemented, for example, using neural networks or fuzzy rule-based models. The first neural networks which consider an evolving structure were published in. These were later expanded by N. Kasabov and P. Angelov for the neuro-fuzzy models. P. Angelov introduced the evolving fuzzy rule-based systems (EFSs) as the first mathematical self-learning model that can dynamically evolve its internal structure and is human interpretable and coined the phrase EFS. Contemporarily, the offline incremental approach for learning an EIS, namely, EFuNN, was proposed by N. Kasabov. P. Angelov, D. Filev, N. Kasabov and O. Cordon organised the first IEEE Symposium on EFSs in 2006 (the proceedings of the conference can be found in). EFSs include a formal (and mathematically sound) learning mechanism to extract it from streaming data. One of the earliest and the most widely cited comprehensive survey on EFSs was done in 2008. Later comprehensive surveys on EFS methods with real applications were done in 2011 and 2016 by E. Lughofer. Other works that contributed further to this area in the following years expanded it to evolving participatory learning, evolving grammar, evolving decision trees, evolving human behaviour modelling, self-calibrating (evolving) sensors (eSensors), evolving fuzzy rule-based classifiers, evolving fuzzy controllers, autonomous fault detectors. More recently, the stability of the evolving fuzzy rule-based systems that consist of the structure learning and the fuzzily weighted recursive least square parameter update method has been proven by Rong. Generalized EFS, which allow rules to be arbitrarily rotated in the feature space and thus to improve their data representability, have been proposed in with significant extensions in towards 'smartness' of the rule bases (thus, termed as "Generalized Smart EFS"), allowing more interpretability and reducing curse of dimensionality. The generalized rule structure was also successfully used in the context of evolving neuro-fuzzy systems. Several facets and challenges for achieving more transparent and understandable rule bases in EFS have been discussed by E. Lughofer in. EISs form the theoretical and methodological basis for the Autonomous Learning Machines (ALMA) and autonomous multi-model systems (ALMMo) as well as of the Autonomous Learning Systems. Evolving Fuzzy Rule-based classifiers, in particular, is a very powerful new concept that offers much more than simply incremental or online classifiers – it can cope with new classes being added or existing classes being merged. This is much more than just adapting to new data samples being added or classification surfaces being evolved. Fuzzy rule-based classifiers are the methodological basis of a new approach to deep learning that was until now considered as a form of multi-layered neural networks. Deep Learning offers high precision levels surpassing the level of human ability and grabbed the imagination of the researchers, industry and the wider public. However, it has a number of intrinsic constraints and limitations. These include: The "black box", opaque internal structure which has millions of parameters and involves ad hoc decisions on the number of layers and algorithm parameters. The requirement for a huge amount of training data samples, computational resources (usually requiring GPUs and/or HPC) and time (usually requiring many hours of training). Iterative search. Requires retraining for new situations (is not evolving). Does not have proven convergence and stability. Most, if not all, of the above limitations can be avoided with the use of the Deep (Fuzzy) Rule-based Classifiers, which were recently introduced based on ALMMo, while achieving similar or even better performance. The resulting prototype-based IF...THEN...models are fully interpretable and dynamically evolving (they can adapt quickly and automatically to new data patterns or even new classes). They are non-parametric and, therefore, their training is non-iterative and fast (it can take few milliseconds per data sample/image on a normal laptop which contrasts with the multiple hours the current deep learning methods require for training even when they use GPUs and HPC). Moreover, they can be trained incrementally, online, or in real-time. Another aspect of Evolving Fuzzy Rule-based classifiers has been proposed in, which, in case of multi-class classification problems, achieves the reduction of class imbalance by cascadability into class sub-spaces and an increased flexibility and performance for adding new classes on the fly from streaming samples.
Read more →
Meta-Labeling

Meta-labeling, also known as corrective AI, is a machine learning (ML) technique utilized in quantitative finance to enhance the performance of investment and trading strategies, developed in 2017 by Marcos López de Prado at Guggenheim Partners and Cornell University. The core idea is to separate the decision of trade direction (side) from the decision of trade sizing, addressing the inefficiencies of simultaneously learning both side and size predictions. The side decision involves forecasting market movements (long, short, neutral), while the size decision focuses on risk management and profitability. It serves as a secondary decision-making layer that evaluates the signals generated by a primary predictive model. By assessing the confidence and likely profitability of those signals, meta-labeling allows investors and algorithms to dynamically size positions and suppress false positives. == Motivation == Meta-labeling is designed to improve precision without sacrificing recall. As noted by López de Prado, attempting to model both the direction and the magnitude of a trade using a single algorithm can result in poor generalization. By separating these tasks, meta-labeling enables greater flexibility and robustness: Enhances control over capital allocation. Reduces overfitting by limiting model complexity. Allows the use of interpretability tools and tailored thresholds to manage risk. Enables dynamic trade suppression in unfavorable regimes. == Applications == Meta-labeling has been applied in a variety of financial ML contexts, including: Algorithmic trading: Filtering and sizing trades to reduce false positives. Portfolio optimization: Scaling exposure across multiple signals with differing confidence levels. Risk management: Dynamically disabling strategies in adverse market conditions. Model validation: Interpreting when and why a model may be underperforming due to regime shifts. == General architecture == Meta-labeling decouples two core components of systematic trading strategies: directional prediction and position sizing. The process involves training a primary model to generate trade signals (e.g., buy, sell, or hold) and then training a secondary model to determine whether each signal is likely to lead to a profitable trade. The second model outputs a probability that is interpreted as the confidence in the forecast, which can be used to adjust the position size or to filter out unreliable trades. Meta-labeling is typically implemented as a three-stage process: Primary model (M1): Predicts the direction or label of a financial outcome using features such as market prices, returns, or volatility indicators. A typical output is directional, e.g., Y ∈ {−1,0,1}, representing short, neutral, or long positions. Secondary model (M2): A binary classifier trained to predict whether the primary model's prediction will be profitable. The target variable is a binary meta-label F ∈ { 0 , 1 } {\displaystyle F\in \{0,1\}} . Inputs can include features used in the primary model, performance diagnostics, or market regime data. Position sizing algorithm (M3): Translates the output probability of the secondary model into a position size. Higher confidence scores result in larger allocations, while lower confidence leads to reduced or zero exposure. === Stage 1: Forecasting side === Primary model architecture Figure 1 Figure 1 presents the architecture of a primary model. It focuses on forecasting the side of the trade. Following the example, this model (M1) takes in input data – such as open-high-low-close data and determines the side of the position to take: a negative number is a short position, and positive number is a long position, the range is set between −1 and 1 (the closer it is to −1 or 1, the stronger the models conviction is). When training the model, the labels are −1 and 1, based on the direction of forward returns for some predefined investment horizon. The researcher may decide to apply a recall check (τ: "Tau") by setting a minimum threshold that the initial output needs to be to qualify of a short or long position (if the threshold is not met, no side forecast is predicted, leading to closing of any open positions), this leads to the primary model output which is one of three possible side forecasts: −1, 0, or 1. The primary model also generates evaluation data which can be used by the secondary model, to improve performance of size forecasts. Some examples of evaluation data include rolling accuracy, F1, recall, precision, and AUC scores. === Stage 2: Filtering out false positives === General meta-labeling architecture Figure 2 Next comes the phase of filtering out false positives, by applying a secondary machine learning model (M2), which is a binary classifier trained to determine if the trade will be profitable or not. The model takes as input four general groupings of data: General input data which is predictive of a false positive. For example the last 30 days rolling volatility of the underlying asset. Evaluation data. Market state and regime data, one may find that macro economic data or clustering the market into regimes may help as specific trading strategies are known to perform better in particular regimes. Example: momentum based strategies perform best in periods with low volatility and strong directional moves. Primary models initial input which is a value between −1 and 1. This highlights the strength of the primary models conviction. The output of the model is a value between −1 and 1 (if using a Tanh function) which will indicate the strength of the conviction that a short or long position is profitable, or it could simply be between 0 and 1 (using a sigmoid function) if one only wanted to know if it made money or not. This output allows filtering out trades that are likely to lead to losses. One could stop at this point or use the outputs of the secondary model as inputs to a position sizing algorithm (M3) which could further enhance strategy performance metrics by translating the output probability of the secondary model into a position size. Higher confidence scores result in larger allocations, while lower confidence leads to reduced or zero exposure. === Stage 3: Optimizing position sizes === ==== Position sizing methods (M3) ==== Various algorithms have been proposed for transforming predicted probabilities into trade sizes: All-or-nothing: Allocate 100% of capital if the probability exceeds a predefined threshold (e.g., 0.5); otherwise, do not trade. Model confidence: Use the probability score directly as the fraction of capital allocated. Linear scaling: Rescale the model's probabilities using min-max normalization based on the training data. Normal CDF (NCDF): Use a normal cumulative distribution function applied to a z-statistic derived from the predicted probability. Empirical CDF (ECDF): Rank probabilities based on their percentile in the training data to ensure relative allocation. Sigmoid Optimal Position Sizing (SOPS): Applies a smooth non-linear sigmoid transformation optimized to maximize risk-adjusted returns (Sharpe ratio). ==== Model calibration ==== Each machine learning algorithm used in meta-labeling tends to produce outputs with different characteristic distributions; for example, some are approximately normally distributed, whereas others exhibit a pronounced U-shape, concentrating probabilities near the extremes. Due to these varying distributions, simply summing the outputs of different models can inadvertently lead to uneven weighting of signals, biasing trade decisions. To address this, model calibration techniques are essential to adjust the predicted probabilities towards frequentist probabilities, ensuring that model outputs reflect true likelihoods more accurately. Two common calibration techniques are: Platt scaling (Sigmoid scaling): Suitable for correcting S-shaped calibration plots typically produced by models such as support vector machines (SVMs). Isotonic regression: Fits a non-decreasing step function to probabilities and is effective particularly with larger datasets, though it can sometimes lead to overfitting. Transforming predictions to frequentist probabilities is crucial as it provides probabilistic outputs that are directly interpretable as the actual likelihood of an event occurring. Such calibration significantly enhances the effectiveness of fixed position sizing methods, reducing maximum drawdowns and increasing risk-adjusted returns. However, calibration has less impact on position sizing methods that directly estimate parameters from the training data, such as ECDF and SOPS, suggesting that calibration is a critical step mainly for fixed methods that rely heavily on raw model outputs. =
Read more →
DARPA Grand Challenge

The DARPA Grand Challenge is a prize competition for American autonomous vehicles, funded by the Defense Advanced Research Projects Agency, the most prominent research organization of the United States Department of Defense. Congress has authorized DARPA to award cash prizes to further DARPA's mission to sponsor revolutionary, high-payoff research that bridges the gap between fundamental discoveries and military use. The initial DARPA Grand Challenge in 2004 was created to spur the development of technologies needed to create the first fully autonomous ground vehicles capable of completing a substantial off-road course within a limited time. The third event, the DARPA Urban Challenge in 2007, extended the initial Challenge to autonomous operation in a mock urban environment. The 2012 DARPA Robotics Challenge, focused on autonomous emergency-maintenance robots, and new Challenges are still being conceived. The DARPA Subterranean Challenge was tasked with building robotic teams to autonomously map, navigate, and search subterranean environments. Such teams could be useful in exploring hazardous areas and in search and rescue. In addition to the challenges in autonomous technology, DARPA has also conducted prize competitions in other areas of technology. == History and background == Fully autonomous vehicles have been an international pursuit for many years, from endeavors in Japan (starting in 1977), Germany (Ernst Dickmanns and VaMP), Italy (the ARGO Project), the European Union (EUREKA Prometheus Project), the United States of America, and other countries. DARPA funded the development of the first fully autonomous robot beginning in 1966 with the Shakey the robot project at Stanford Research Institute, now SRI International. The first autonomous ground vehicle capable of driving on and off roads was developed by DARPA as part of the Strategic Computing Initiative beginning in 1984 leading to demonstrations of autonomous navigation by the Autonomous Land Vehicle and the Navlab. The Grand Challenge was the first long distance competition for driverless cars in the world; other research efforts in the field of driverless cars take a more traditional commercial or academic approach. The U.S. Congress authorized DARPA to offer prize money ($1 million) for the first Grand Challenge to facilitate robotic development, with the ultimate goal of making one-third of ground military forces autonomous by 2015. Following the 2004 event, Dr. Tony Tether, the director of DARPA, announced that the prize money had been increased to $2 million for the next event, which was claimed on October 9, 2005. The first, second and third places in the 2007 Urban Challenge received $2 million, $1 million, and $500,000, respectively. 14 new teams have qualified in year 2015. The competition was open to teams and organizations from around the world, as long as there was at least one U.S. citizen on the roster. Teams have participated from high schools, universities, businesses and other organizations. More than 100 teams registered in the first year, bringing a wide variety of technological skills to the race. In the second year, 195 teams from 36 U.S. states and 4 foreign countries entered the race. == 2004 Grand Challenge == The first competition of the DARPA Grand Challenge was held on March 13, 2004 in the Mojave Desert region of the United States, along a 150-mile (240 km) route that follows along the path of Interstate 15 from just before Barstow, California to just past the California–Nevada border in Primm. None of the robot vehicles finished the route. Carnegie Mellon University's Red Team and car Sandstorm (a converted Humvee) traveled the farthest distance, completing 11.78 km (7.32 mi) of the course before getting hung up on a rock after making a switchback turn. No winner was declared, and the cash prize was not given. Therefore, a second DARPA Grand Challenge event was scheduled for 2005. == 2005 Grand Challenge == The second competition of the DARPA Grand Challenge began at 6:40 am on October 8, 2005. All but one of the 23 finalists in the 2005 race surpassed the 11.78 km (7.32 mi) distance completed by the best vehicle in the 2004 race. Five vehicles successfully completed the 212 km (132 mi) course: Vehicles in the 2005 race passed through three narrow tunnels and negotiated more than 100 sharp left and right turns. The race concluded through Beer Bottle Pass, a winding mountain pass with a sheer drop-off on one side and a rock face on the other. Although the 2004 course required more elevation gain and some very sharp switchbacks (Daggett Ridge) were required near the beginning of the route, the course had far fewer curves and generally wider roads than the 2005 course. The natural rivalry between the teams from Stanford and Carnegie Mellon (Sebastian Thrun, head of the Stanford team was previously a faculty member at Carnegie Mellon and colleague of Red Whittaker, head of the CMU team) was played out during the race. Mechanical problems plagued H1ghlander before it was passed by Stanley. Gray Team's entry was a miracle in itself, as the team from the suburbs of New Orleans was caught in Hurricane Katrina a few short weeks before the race. The fifth finisher, Terramax, a 30,000 pound entry from Oshkosh Truck, finished on the second day. The huge truck spent the night idling on the course, but was particularly nimble in carefully picking its way down the narrow roads of Beer Bottle Pass. == 2007 Urban Challenge == The third competition of the DARPA Grand Challenge, known as the "Urban Challenge", took place on November 3, 2007 at the site of the now-closed George Air Force Base (currently used as Southern California Logistics Airport), in Victorville, California (Google map). The course involved a 96 km (60 mi) urban area course, to be completed in less than 6 hours. Rules included obeying all traffic regulations while negotiating with other traffic and obstacles and merging into traffic. Unlike previous challenges, the 2007 Urban Challenge organizers divided competitors into two "tracks", A and B. All Track A and Track B teams were part of the same competition circuit, but the teams chosen for the Track A program received US $1 million in funding. These 11 teams largely represented major universities and large corporate interests such as CMU teaming with GM as Tartan Racing, Stanford teaming with Volkswagen, Virginia Tech teaming with TORC Robotics as VictorTango, Oshkosh Truck, Honeywell, Raytheon, Caltech, Autonomous Solutions, Cornell University, and MIT. One of the few independent entries in Track A was the Golem Group. DARPA has not publicly explained the rationale behind the selection of Track A teams. Teams were given maps sparsely charting the waypoints that defined the competition courses. At least one team, Tartan Racing, enhanced the maps through the insertion of additional extrapolated waypoints for improved navigation. A debriefing paper published by Team Jefferson illustrates graphically the contrast between the course map it was given by DARPA and the course map used by Tartan Racing. Tartan Racing claimed the $2 million prize with their vehicle "Boss", a Chevy Tahoe. The second-place finisher earning the $1 million prize was the Stanford Racing Team with their entry "Junior", a 2006 Volkswagen Passat. Coming in third place was team VictorTango, winning the $500,000 prize with their 2005 Ford Escape hybrid, "Odin". MIT placed 4th, with Cornell University and University of Pennsylvania/Lehigh University also completing the course. The six teams that successfully finished the entire course: While the 2004 and 2005 events were more physically challenging for the vehicles, the robots operated in isolation and only encountered other vehicles on the course when attempting to pass. The Urban Challenge required designers to build vehicles able to obey all traffic laws while they detect and avoid other robots on the course. This is a particular challenge for vehicle software, as vehicles must make "intelligent" decisions in real time based on the actions of other vehicles. Other than previous autonomous vehicle efforts that focused on structured situations such as highway driving with little interaction between the vehicles, this competition operated in a more cluttered urban environment and required the cars to perform sophisticated interactions with each other, such as maintaining precedence at a 4-way stop intersection. == 2012 Robotics Challenge == The DARPA Robotics Challenge is an ongoing competition focusing on humanoid robotics. The primary goal of the program is to develop ground robotic capabilities to execute complex tasks in dangerous, degraded, human-engineered environments. It launched in October 2012, and hosted the Virtual Robotics Competition in June 2013. Two more competitions are planned: the DRC Trials in December 2013, and the DRC Finals in December 2014. Unlike prior Challenges, the construction of the "vehicles" w
Read more →
Data processing unit

A data processing unit (DPU) is a programmable computer processor that tightly integrates a general-purpose CPU with network interface hardware. They are also occasionally called "IPUs" (infrastructure processing unit) or "SmartNICs". They can be used in place of traditional NICs to relieve the main CPU of complex networking responsibilities and other "infrastructural" duties; although their features vary, they may be used to perform encryption/decryption, serve as a firewall, handle TCP/IP, process HTTP requests, or even function as a hypervisor or storage controller. These devices can be attractive to cloud computing providers whose servers might otherwise spend a significant amount of CPU time on these tasks, cutting into the cycles they can provide to guests. They see use in other kinds of data center environments as well due to their improved power consumption efficiency for routine networking tasks compared to general-purpose CPUs.
Read more →