ClubDJPro (often referred to as ClubDJ) is a DJ console and video mixing tool developed by Cube Software Solutions Inc. software. It was released in June 2005. == User interface == ClubDJPro has a GUI that was designed to allow aesthetic revisions via Skins. The skin engine that ClubDJPro uses allows for the ability to expand the software to take up the entire screen. As of 4.4.3.3 there are 3 user changeable skins included in the program which are changeable in the preferences tab. They are called 'AquaLung', 'Eleanor', and 'Grabber'. == Editions == ClubDJPro is available in two different editions, with separate features depending upon their target consumer group. DJ Edition - Can play audio files only. VJ Edition - Contains all of the features of the DJ Edition, in addition to support for video, karaoke, and visualizations. == Supported MIDI Controllers == Supported since version 2.0: Hercules Console Hercules Console MK2 Hercules Control MP3 PCDJ DAC-2 Controller == History == The initial "final release" of ClubDJPro was released on June 24, 2005. On June 26, 2009, the 4th iteration of the ClubDJPro software was released. The development of the software and website appears to have halted. As of March 2018 the website continues to show a new version "Coming Spring 2016".
Weight initialization
In deep learning, weight initialization or parameter initialization describes the initial step in creating a neural network. A neural network contains trainable parameters that are modified during training: weight initialization is the pre-training step of assigning initial values to these parameters. The choice of weight initialization method affects the speed of convergence, the scale of neural activation within the network, the scale of gradient signals during backpropagation, and the quality of the final model. Proper initialization is necessary for avoiding issues such as vanishing and exploding gradients and activation function saturation. Note that even though this article is titled "weight initialization", both weights and biases are used in a neural network as trainable parameters, so this article describes how both of these are initialized. Similarly, trainable parameters in convolutional neural networks (CNNs) are called kernels and biases, and this article also describes these. == Constant initialization == We discuss the main methods of initialization in the context of a multilayer perceptron (MLP). Specific strategies for initializing other network architectures are discussed in later sections. For an MLP, there are only two kinds of trainable parameters, called weights and biases. Each layer l {\displaystyle l} contains a weight matrix W ( l ) ∈ R n l − 1 × n l {\displaystyle W^{(l)}\in \mathbb {R} ^{n_{l-1}\times n_{l}}} and a bias vector b ( l ) ∈ R n l {\displaystyle b^{(l)}\in \mathbb {R} ^{n_{l}}} , where n l {\displaystyle n_{l}} is the number of neurons in that layer. A weight initialization method is an algorithm for setting the initial values for W ( l ) , b ( l ) {\displaystyle W^{(l)},b^{(l)}} for each layer l {\displaystyle l} . The simplest form is zero initialization: W ( l ) = 0 , b ( l ) = 0 {\displaystyle W^{(l)}=0,b^{(l)}=0} Zero initialization is usually used for initializing biases, but it is not used for initializing weights, as it leads to symmetry in the network, causing all neurons to learn the same features. In this page, we assume b = 0 {\displaystyle b=0} unless otherwise stated. Recurrent neural networks typically use activation functions with bounded range, such as sigmoid and tanh, since unbounded activation may cause exploding values. (Le, Jaitly, Hinton, 2015) suggested initializing weights in the recurrent parts of the network to identity and zero bias, similar to the idea of residual connections and LSTM with no forget gate. In most cases, the biases are initialized to zero, though some situations can use a nonzero initialization. For example, in multiplicative units, such as the forget gate of LSTM, the bias can be initialized to 1 to allow good gradient signal through the gate. For neurons with ReLU activation, one can initialize the bias to a small positive value like 0.1, so that the gradient is likely nonzero at initialization, avoiding the dying ReLU problem. == Random initialization == Random initialization means sampling the weights from a normal distribution or a uniform distribution, usually independently. === LeCun initialization === LeCun initialization, popularized in (LeCun et al., 1998), is designed to preserve the variance of neural activations during the forward pass. It samples each entry in W ( l ) {\displaystyle W^{(l)}} independently from a distribution with mean 0 and variance 1 / n l − 1 {\displaystyle 1/n_{l-1}} . For example, if the distribution is a continuous uniform distribution, then the distribution is U ( ± 3 / n l − 1 ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {3/n_{l-1}}})} . === Glorot initialization === Glorot initialization (or Xavier initialization) was proposed by Xavier Glorot and Yoshua Bengio. It was designed as a compromise between two goals: to preserve activation variance during the forward pass and to preserve gradient variance during the backward pass. For uniform initialization, it samples each entry in W ( l ) {\displaystyle W^{(l)}} independently and identically from U ( ± 6 / ( n l + 1 + n l − 1 ) ) {\displaystyle {\mathcal {U}}(\pm {\sqrt {6/(n_{l+1}+n_{l-1})}})} . In the context, n l − 1 {\displaystyle n_{l-1}} is also called the "fan-in", and n l + 1 {\displaystyle n_{l+1}} the "fan-out". When the fan-in and fan-out are equal, then Glorot initialization is the same as LeCun initialization. === He initialization === As Glorot initialization performs poorly for ReLU activation, He initialization (or Kaiming initialization) was proposed by Kaiming He et al. for networks with ReLU activation. It samples each entry in W ( l ) {\displaystyle W^{(l)}} from N ( 0 , 2 / n l − 1 ) {\displaystyle {\mathcal {N}}(0,2/n_{l-1})} . === Orthogonal initialization === (Saxe et al. 2013) proposed orthogonal initialization: initializing weight matrices as uniformly random (according to the Haar measure) semi-orthogonal matrices, multiplied by a factor that depends on the activation function of the layer. It was designed so that if one initializes a deep linear network this way, then its training time until convergence is independent of depth. Sampling a uniformly random semi-orthogonal matrix can be done by initializing X {\displaystyle X} by IID sampling its entries from a standard normal distribution, then calculate ( X X ⊤ ) − 1 / 2 X {\displaystyle \left(XX^{\top }\right)^{-1/2}X} or its transpose, depending on whether X {\displaystyle X} is tall or wide. For CNN kernels with odd widths and heights, orthogonal initialization is done this way: initialize the central point by a semi-orthogonal matrix, and fill the other entries with zero. As an illustration, a kernel K {\displaystyle K} of shape 3 × 3 × c × c ′ {\displaystyle 3\times 3\times c\times c'} is initialized by filling K [ 2 , 2 , : , : ] {\displaystyle K[2,2,:,:]} with the entries of a random semi-orthogonal matrix of shape c × c ′ {\displaystyle c\times c'} , and the other entries with zero. (Balduzzi et al., 2017) used it with stride 1 and zero-padding. This is sometimes called the Orthogonal Delta initialization. Related to this approach, unitary initialization proposes to parameterize the weight matrices to be unitary matrices, with the result that at initialization they are random unitary matrices (and throughout training, they remain unitary). This is found to improve long-sequence modelling in LSTM. Orthogonal initialization has been generalized to layer-sequential unit-variance (LSUV) initialization. It is a data-dependent initialization method, and can be used in convolutional neural networks. It first initializes weights of each convolution or fully connected layer with orthonormal matrices. Then, proceeding from the first to the last layer, it runs a forward pass on a random minibatch, and divides the layer's weights by the standard deviation of its output, so that its output has variance approximately 1. === Fixup initialization === In 2015, the introduction of residual connections allowed very deep neural networks to be trained, much deeper than the ~20 layers of the previous state of the art (such as the VGG-19). Residual connections gave rise to their own weight initialization problems and strategies. These are sometimes called "normalization-free" methods, since using residual connection could stabilize the training of a deep neural network so much that normalizations become unnecessary. Fixup initialization is designed specifically for networks with residual connections and without batch normalization, as follows: Initialize the classification layer and the last layer of each residual branch to 0. Initialize every other layer using a standard method (such as He initialization), and scale only the weight layers inside residual branches by L − 1 2 m − 2 {\displaystyle L^{-{\frac {1}{2m-2}}}} . Add a scalar multiplier (initialized at 1) in every branch and a scalar bias (initialized at 0) before each convolution, linear, and element-wise activation layer. Similarly, T-Fixup initialization is designed for Transformers without layer normalization. === Others === Instead of initializing all weights with random values on the order of O ( 1 / n ) {\displaystyle O(1/{\sqrt {n}})} , sparse initialization initialized only a small subset of the weights with larger random values, and the other weights zero, so that the total variance is still on the order of O ( 1 ) {\displaystyle O(1)} . Random walk initialization was designed for MLP so that during backpropagation, the L2 norm of gradient at each layer performs an unbiased random walk as one moves from the last layer to the first. Looks linear initialization was designed to allow the neural network to behave like a deep linear network at initialization, since W R e L U ( x ) − W R e L U ( − x ) = W x {\displaystyle W\;\mathrm {ReLU} (x)-W\;\mathrm {ReLU} (-x)=Wx} . It initializes a matrix W {\displaystyle W} of shape R n 2 × m {\displaystyle \mathbb {R} ^{{\frac {n}{2}}\times m}} by any method, such as orthogonal initialization, t
Trustworthy AI
Trustworthy AI refers to artificial intelligence systems that are designed to have transparent reasoning, are explainable (XAI), accountable, robust, fair and honest, respectful of data privacy, and steerable or alignable with human goals. == Terminology == Recent work in AI ethics distinguishes trustworthiness and trustability as two different conditions relevant to trustworthy AI. Trustworthiness is concerned with whether an AI system or the institutions deploying it merit trust by being reliable, fair, and accountable. Trustability, on the other hand, is the prior question of whether a given entity is even the kind of thing to which interpersonal trust can coherently apply as opposed to mere instrumental reliance. Some philosophers argue that current AI systems are best understood as tools that are not genuine targets of interpersonal trust. They argue that trust should be directed toward the human and institutional arrangements that govern the systems' design, deployment, and oversight. This stance supports interpreting "trustworthy AI" as trustworthy governance and use of AI rather than trust in the artifacts themselves. Transparency in AI involves making the processes and decisions of such systems understandable to users and stakeholders. Accountability ensures that there are protocols for addressing adverse outcomes or biases that may arise, with designated responsibilities for oversight and remediation. Robustness and security aim to ensure that AI systems perform reliably under various conditions and are safeguarded against malicious attacks. Harmlessness can be achieved by refusal training: training the models to avoid problematic requests, and by adding filters to detect and prevent discussion on biased, unethical, or dangerous outputs. There is research on how to train AI so that it aligns with human goals. == Techniques and ITU standardization == Trustworthy AI creation is a goal of AI governance and policymaking. To achieve transparency and data privacy, several privacy-enhancing technologies (PETs) can be used. These include: Homomorphic encryption for computing with encrypted data without ever decrypting it. Federated learning and secure multi-party computation (MPC) for distributing the model training without sharing information between the learning centers and computing servers. Differential privacy for exposing statistical data while guaranteeing that no private information is exposed. Zero-knowledge proof - providing proven validity for statements without disclosing any extra information. A work programme for achieving Trustworthy AI was set up by the International Telecommunication Union, an agency of the United Nations, initiated under its AI for Good programme. Its origin lies with the ITU-WHO Focus Group on Artificial Intelligence for Health, where a strong need for both privacy and analytics created demand for a standard in these technologies. In 2020, AI for Good moved online, and the TrustworthyAI seminar series was established to initiate discussions on these topics. This eventually led to standardization activities. === Multi-party computation === Secure multi-party computation (MPC) is being standardized under "Question 5" (the incubator) of ITU-T Study Group 17. === Homomorphic encryption === Homomorphic encryption allows for computing on encrypted data, where the outcomes or result is still encrypted and unknown to those performing the computation, but can be deciphered by the original encryptor. It is often developed with the goal of enabling use in jurisdictions different from the data creation (under, for instance, GDPR). ITU has been collaborating since the early stage of the HomomorphicEncryption.org standardization meetings, which has developed a standard on homomorphic encryption. The fifth homomorphic encryption meeting was hosted at ITU HQ in Geneva. === Federated learning === Zero-sum masks as used by federated learning for privacy preservation are used extensively in the multimedia standards of ITU-T Study Group 16 (VCEG) such as JPEG, MP3, H.264, and H.265 (commonly known as MPEG). === Zero-knowledge proof === Previous pre-standardization work on the topic of zero-knowledge proof has been conducted in the ITU-T Focus Group on Digital Ledger Technologies. === Differential privacy === The application of differential privacy in the preservation of privacy was examined at several of the "Day 0" machine learning workshops at AI for Good Global Summits. == Mozilla "Rebel Alliance" == In January 2026, the Mozilla Foundation and its subsidiaries announced a strategic shift to deploy their entire $1.4 billion reserve into building what foundation president Mark Surman termed a "rebel alliance" for trustworthy AI. Framed by Surman as a mission-driven alternative to the market dominance of OpenAI and Anthropic, the initiative seeks to establish an open-source AI stack by 2028. The alliance includes several startups funded via Mozilla Ventures, specifically focusing on decentralized governance and transparency: Trail: A firm developing AI compliance frameworks for regulated industries. Transformer Lab: A developer of open-source tools for AI model management. Oumi: A platform for training and deploying open-source models. The "rebel alliance" terminology is a historical reference to Mozilla's efforts in 1998 to challenge Microsoft's browser monopoly. While the $1.4 billion in funding is significant, it has been contrasted with the tens of billions in capital raised by proprietary competitors like OpenAI.
Generative AI Copyright Disclosure Act
The Generative AI Copyright Disclosure Act is a piece of legislation introduced by California Representative Adam Schiff in the United States Congress on April 9, 2024. It concerns the transparency of companies regarding their use of copyrighted work to train their generative artificial intelligence (AI) models. The legislation requires the submission of a notice regarding the identity and the uniform resource locator (URL) address of the copyrighted works used in the training data to the Register of Copyrights at least 30 days before the public release of the new or updated version of the AI model; it does not ban the use of copyrighted works for AI training. The bill's requirements would apply retroactively to prior AI models. Violation penalties would start at US$5,000. The legislation does not have a maximum penalty assessment that can be charged. The bill by Schiff was introduced a few days after The New York Times published an article regarding the business activities of major tech firms, including Google and Meta, in the training of their generative AI platforms on April 6, 2024. The legislation is supported by the Professional Photographers of America (PPA), SAG-AFTRA, the Writers Guild of America, the International Alliance of Theatrical Stage Employees (IATSE), the Recording Industry Association of America (RIAA), and others.
Liang Wenfeng
Liang Wenfeng (Chinese: 梁文锋; pinyin: Liáng Wénfēng; born 1985) is a Chinese entrepreneur and businessman who is the co-founder of the quantitative hedge fund High-Flyer, as well as the founder and CEO of its artificial intelligence company DeepSeek. Liang attended Zhejiang University, and began his career by applying machine learning methods to quantitative finance. Through High-Flyer, he built large-scale computing infrastructure that was later used to support artificial intelligence research, leading to the creation of DeepSeek in 2023. DeepSeek gained international attention following the release of DeepSeek-R1, which analysts described as demonstrating high-level performance with comparatively limited compute resources. In 2025, Liang was named to Time magazine's list of 100 Most Influential People in AI and Fortune's list of the Most Powerful People in Business. == Early life == Liang was born in 1985 in the village of Mililing (米历岭村), Qinba town (覃巴镇), Wuchuan city (吴川市), Guangdong. His parents were both primary school teachers. Liang was routinely praised by both locals and teachers alike. Even since middle school, Liang was recalled for being well-known for reading comic books, while also being very proficient in mathematics. == Education == After elementary school, Liang attended Wuchuan No. 1 Middle School. There, he quickly excelled in class and ranked highly amongst his peers. He taught himself high school and university-level mathematics courses. Liang then attended Wuchaun No. 1 High School. In these years, he developed hobbies of mathematical modeling and conducting research projects. Compared to his peers, he was always ranked highly. For every mathematics exam, he always ranked within the top three. He was also the top scorer in the Zhanjiang region of Guangdong for the college entrance exam. Thus, in 2002, Liang left high school early to further pursue his education at the university level at the young age of 17. Attending Zhejiang University at the age of 17, Liang earned a Bachelor of Engineering in Electronic Information Engineering in 2007 and his Master of Engineering in Information & Communication Engineering in 2010. His master's dissertation was titled "Study on Object Tracking Algorithm Based on Low-Cost PTZ camera" (基于低成本PTZ摄像机的目标跟踪算法研究). In his college years, DJI founder Wang Tao asked Liang to join as a co-founder. Liang declined the invitation to pursue artificial intelligence methodologies in financial markets. While he states that those around him had entrepreneurial mindsets, he himself valued academics. == Career == === Early career (2008–2016) === During the 2008 financial crisis, Liang formed a team with his classmates to accumulate data related to financial markets. He also led the team to explore quantitative trading using machine learning and other technologies. After his graduation, Liang moved to a cheap flat in Chengdu, Sichuan, where he experimented with ways to apply AI to various fields. These ventures failed, until he tried applying AI to finance. In 2013, Liang attempted to integrate artificial intelligence with quantitative trading and founded Hangzhou Yakebi Investment Management Co Ltd with Xu Jin, an alumnus of Zhejiang University. In 2015, they co-founded Hangzhou Huanfang Technology Co Ltd, which is today's Zhejiang Jiuzhang Asset Management Co Ltd. === High-Flyer (2016–2023) === In February 2016, Liang and two other engineering classmates co-founded Ningbo High-Flyer Quantitative Investment Management Partnership (Limited Partnership). The team relied on mathematics and AI to make investments. Much of the early startup culture was described by former employees to be "geeky" and "quirky," often seen as contrary to the existing culture in large Chinese tech companies. In 2019, Liang founded High-Flyer AI which was dedicated to research on AI algorithms and its basic applications. By this time, High-Flyer had over 10 billion yuan in assets under management. On 30 August 2019, Liang Wenfeng delivered a keynote speech entitled "The Future of Quantitative Investment in China from a Programmer's Perspective" at the Private Equity Golden Bull Award ceremony held by China Securities Journal, and sparked heated discussions. Liang stated that the criterion for determining what is quantitative or non-quantitative is whether the investment decision is made by quantitative methods or by people. Quantitative funds do not have portfolio managers making the decisions and instead are just servers. He also stated High-Flyer's mission is to improve the effectiveness of China's secondary market. In February 2021, Gregory Zuckerman's book The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution was published. Liang wrote the preface for the Chinese edition of the book where he stated that whenever he encountered difficulties at work, he would think of Simons' words "There must be a way to model prices". In January 2025, Zuckerman wrote in The Wall Street Journal where he acknowledged this fact and stated he has been trying to get in touch with Liang but much like Simons, Liang is very secretive and difficult to contact. During 2021, Liang started buying thousands of Nvidia GPUs for his AI side project while running High-Flyer. Liang wanted to build something and it will be a game changer which his business partners thought was only possible from giants such as ByteDance and Alibaba Group. === DeepSeek (since 2023) === ==== DeepSeek begins ==== In May 2023, Liang announced High-Flyer would pursue the development of artificial general intelligence and launched DeepSeek. During that month in an interview with 36Kr, Liang stated that High-Flyer had acquired 10,000 Nvidia A100 GPUs before the US government imposed AI chip restrictions on China. That laid the foundation for DeepSeek to operate as an LLM developer. Liang also stated DeepSeek gets funding from High-Flyer. This was because when DeepSeek was founded, venture capital firms were reluctant in providing funding as it was unlikely that it would be able to generate an exit in a short period of time. Liang only personally holds 1% of the company, with 99% of the company being held by Ningbo High-Flyer Quantitative Investment Management Partnership (Limited Partnership). With DeepSeek's funding model, it lacks commercial pressure and rigid key performance indicators, enabling the company to deviate from previously established model architectures. ==== Early development ==== In July 2024, Liang was interviewed again by 36Kr. He stated that when DeepSeek-V2 was released and triggered an AI price war in China, it came as a huge surprise as the team did not expect pricing to be so sensitive. Liang's aggressive pricing of the language model forced domestic tech giants including Alibaba and Baidu to cut their own rates by over 95%. He also stated that as China's economy develops, it should gradually become a contributor instead of freeriding. What is lacking in China's innovation is not capital but a lack of confidence and knowledge on organizing talent into it. DeepSeek has not hired anyone particularly special and employees tend to be locally educated. When it comes to disruptive technologies, closed source approaches can only temporarily delay others in catching up. As the goal was long-term, DeepSeek sought employees who had ability and passion rather than experience. To retain a high talent density relative to larger firms like Bytedance or Baidu, DeepSeek aimed to maintain a low-hierarchy corporate culture, with members working in project-based groups, as well as competitive compensation. Liang emphasized his vision for DeepSeek employees to bring their "unique experience and ideas" instead of needing to be explicitly directed, with an overall bottom-up approach to division of labor. Liang noted that a significant outcome of this approach was the multi-head latent attention training architecture, which was attributed directly to a young DeepSeek researcher's personal interest. This advancement played a core role in reducing the cost of training the DeepSeek-V3 model, released in December 2024. ==== Release of DeepSeek-R1 ==== Also on 20 January 2025, DeepSeek, the company Liang founded and served as the CEO, released DeepSeek-R1, a 671-billion-parameter open-source reasoning AI model, alongside the publication of a detailed technical paper explaining its architecture and training methodology. The model was built using just 2,048 Nvidia H800 GPUs at a cost of $5.6 million, showcasing a resource-efficient approach that contrasted sharply with the billion-dollar budgets of Western competitors. The development of DeepSeek-R1 occurred amidst U.S. sanctions where Trump limited sales of Nvidia chips to China. By 27 January, DeepSeek surpassed ChatGPT to become the #1 free app on the United States iOS App Store. U.S. stocks plummeted, as more than $1 trillion was erased in market capitalization amid panic over DeepSeek. Technology journ
Large language model
A large language model (LLM) is a neural network trained on a vast amount of text for natural language processing tasks, especially language generation. LLMs can typically generate, summarize, translate and analyze text in many contexts, and are a foundational technology behind modern chatbots. Biased or inaccurate training data can make an LLM's output less reliable. As of 2026, the most capable LLMs are based on transformer architectures, which, according to the 2017 paper "Attention Is All You Need", can be more efficient and parallelizable than earlier statistical and recurrent neural network models. Benchmark evaluations for LLMs attempt to measure model reasoning, factual accuracy, alignment, and safety. == History == Before the emergence of transformer-based models in 2017, some language models were considered large relative to the computational and data constraints of their time. In the early 1990s, IBM's statistical models pioneered word alignment techniques for machine translation, laying the groundwork for corpus-based language modeling. In 2001, a smoothed n-gram model, such as those employing Kneser–Ney smoothing, trained on 300 million words, achieved state-of-the-art perplexity on benchmark tests. During the 2000s, with the rise of widespread internet access, researchers began compiling massive text datasets from the web ("web as corpus") to train statistical language models. Moving beyond n-gram models, researchers started in 2000 to use neural networks as language models. Following the breakthrough of deep neural networks in image classification around 2012, similar architectures were adapted for language tasks. This shift was marked by the development of word embeddings (e.g., Word2Vec by Mikolov in 2013) and sequence-to-sequence (seq2seq) models using LSTM. In 2016, Google transitioned its translation service to neural machine translation (NMT), replacing statistical phrase-based models with deep recurrent neural networks. These early NMT systems used LSTM-based encoder-decoder architectures, as they preceded the invention of transformers. At the 2017 NeurIPS conference, Google researchers introduced the transformer architecture in their landmark paper "Attention Is All You Need". This paper's goal was to improve upon 2014 seq2seq technology, and was based mainly on the attention mechanism developed by Bahdanau et al. in 2014. The following year in 2018, BERT was introduced and quickly became "ubiquitous". Though the original transformer has both encoder and decoder blocks, BERT is an encoder-only model. Academic and research usage of BERT began to decline in 2023, following rapid improvements in the abilities of decoder-only models (such as GPT) to solve tasks via prompting. Although decoder-only GPT-1 was introduced in 2018, it was GPT-2 in 2019 that caught widespread attention because OpenAI claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use. GPT-3 in 2020 went a step further and as of 2025 is available only via API with no offering of downloading the model to execute locally. But it was the consumer-facing chatbot ChatGPT in late 2022 that received extensive media coverage and public attention by 2023. The 2023 GPT-4 was praised for its increased accuracy and as a "holy grail" for its multimodal capabilities. OpenAI did not reveal the high-level architecture and the number of parameters of GPT-4. The release of ChatGPT led to an uptick in LLM usage across several research subfields of computer science, including robotics, software engineering, and societal impact work. In 2024, OpenAI released the reasoning model OpenAI o1, which generates long chains of thought before returning a final answer. Many LLMs with parameter counts comparable to those of OpenAI's GPT series have been developed. Since 2022, weights-available models have been gaining popularity, especially at first with BLOOM and LLaMA, though both have restrictions on usage and deployment. Mistral AI's open-weight models Mistral 7B and Mixtral 8x7B have a more permissive Apache License. In January 2025, DeepSeek released DeepSeek R1, a 671-billion-parameter open-weight model that performs comparably to OpenAI o1 but at a much lower price per token for users. Since 2023, many LLMs have been trained to be multimodal, having the ability to also process or generate other types of data, such as images, audio, or 3D meshes. Open-weight LLMs have become more influential since 2023. Per Vake et al. (2025), community-driven contributions to open-weight models improve their efficiency and performance via collaborative platforms such as Hugging Face. == Dataset preprocessing == === Tokenization === As machine learning algorithms process numbers rather than text, the text must be converted to numbers. In the first step, a vocabulary is decided upon, then integer indices are arbitrarily but uniquely assigned to each vocabulary entry, and finally, an embedding is associated with the integer index. Algorithms include byte-pair encoding (BPE) and WordPiece. There are also special tokens serving as control characters, such as [MASK] for masked-out token (as used in BERT), and [UNK] ("unknown") for characters not appearing in the vocabulary. Also, some special symbols are used to denote special text formatting. For example, "Ġ" denotes a preceding whitespace in RoBERTa and GPT and "##" denotes continuation of a preceding word in BERT. For example, the BPE tokenizer used by the legacy version of GPT-3 would split tokenizer: texts -> series of numerical "tokens" as Tokenization also compresses the datasets. Because LLMs generally require input to be an array that is not jagged, the shorter texts must be "padded" until they match the length of the longest one. ==== Byte-pair encoding ==== As an example, consider a tokenizer based on byte-pair encoding. In the first step, all unique characters (including blanks and punctuation marks) are treated as an initial set of n-grams (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) n-grams that most frequently occur together are then again merged into even lengthier n-gram, until a vocabulary of prescribed size is obtained. After a tokenizer is trained, any text can be tokenized by it, as long as it does not contain characters not appearing in the initial-set of uni-grams. === Dataset cleaning === In the context of training LLMs, datasets are typically cleaned by removing low-quality, duplicated, or toxic data. Cleaned datasets can increase training efficiency and lead to improved downstream performance. A trained LLM can be used to clean datasets for training a further LLM. With the increasing proportion of LLM-generated content on the web, data cleaning in the future may include filtering out such content. LLM-generated content can pose a problem if the content is similar to human text (making filtering difficult) but of lower quality (degrading performance of models trained on it). === Synthetic data === Training of largest language models might need more linguistic data than naturally available, or that the naturally occurring data is of insufficient quality. In these cases, synthetic data might be used. == Training == An LLM is a type of foundation model (large X model) trained on language. LLMs can be trained in different ways. In particular, GPT models are first pretrained to predict the next word on a large amount of data, before being fine-tuned. === Cost === Substantial infrastructure is necessary for training the largest models. The tendency towards larger models is visible in the list of large language models. For example, the training of GPT-2 (i.e. a 1.5-billion-parameter model) in 2019 cost $50,000, while training of the PaLM (i.e. a 540-billion-parameter model) in 2022 cost $8 million, and Megatron-Turing NLG 530B (in 2021) cost around $11 million. The qualifier "large" in "large language model" is inherently vague, as there is no definitive threshold for the number of parameters required to qualify as "large". === Fine-tuning === Before being fine-tuned, most LLMs are next-token predictors. The fine-tuning shapes the LLM's behavior via techniques like reinforcement learning from human feedback (RLHF) or constitutional AI. Instruction fine-tuning is a form of supervised learning used to teach LLMs to follow user instructions. In 2022, OpenAI demonstrated InstructGPT, a version of GPT-3 similarly fine-tuned to follow instructions. Reinforcement learning from human feedback (RLHF) involves training a reward model to predict which text humans prefer. Then, the LLM can be fine-tuned through reinforcement learning to better satisfy this reward model. Since humans typically prefer truthful, helpful and harmless answers, RLHF favors such answers. == Architecture == LLMs are generally based on the tra
Juergen Pirner
Juergen Pirner (born 1956) is the German creator of Jabberwock, a chatterbot that won the 2003 Loebner prize. Pirner created Jabberwock modelling the Jabberwocky from Lewis Carroll's poem of the same name. Initially, Jabberwock would just give rude or fantasy-related answers; but over the years, Pirner has programmed better responses into it. As of 2007 he has taught it 2.7 million responses. Pirner lives in Hamburg, Germany.