Local-first software is a software engineering approach in which an application stores its data primarily on the user's own device rather than on remote servers. Users can read and write data without an Internet connection, and changes are synchronized across devices in the background when connectivity is available. The approach differs from conventional cloud-based applications, where the server holds the authoritative copy of user data and the client acts as a thin client. The term was coined in a 2019 paper published by researchers at Ink & Switch, an independent research lab, and presented at the Onward! conference at ACM SIGPLAN. The paper, sometimes referred to as a manifesto, was authored by Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan. == Background == Before the widespread adoption of Internet-connected software in the 2000s, most desktop applications stored data as files on the user's local disk. Users had direct access to their files and could copy, back up, or delete them at will. The rise of software as a service (SaaS) and cloud-based applications like Google Docs shifted data storage to centralized servers. While cloud applications made real-time collaboration across devices straightforward, they introduced a dependency on the service provider: if the provider discontinued the service or experienced an outage, users could lose access to their data. A related concept, "offline-first," emerged in the early 2010s and focused on making web applications resilient to network interruptions. The local-first approach built on these earlier efforts while placing greater emphasis on long-term data ownership and end-to-end encryption. == Origins == === Ink & Switch manifesto === Ink & Switch is an industrial research lab co-founded by Adam Wiggins, who had earlier co-founded Heroku. Martin Kleppmann, an associate professor in the Department of Computer Science and Technology at the University of Cambridge, was a co-author of the 2019 paper. The manifesto proposed seven "ideals" for local-first software: Fast — Operations respond without network round-trips. Multi-device — Data synchronizes across a user's devices. Offline — Users can read and write data without a network connection. Collaboration — Multiple users can work on the same data concurrently. Longevity — Data remains accessible even if the software vendor ceases operation. Privacy — End-to-end encryption protects user data. User control — The vendor cannot restrict how users access or use their data. The paper surveyed existing approaches to data storage and collaboration — ranging from email attachments and Dropbox-style file synchronization to web applications and mobile backends — and argued that none of them satisfied all seven ideals simultaneously. === Role of CRDTs === The manifesto identified conflict-free replicated data types (CRDTs) as a promising technical foundation for local-first applications. CRDTs are data structures that allow multiple replicas to be edited independently and then merged without conflicts, a property first formalized in research by Marc Shapiro and colleagues around 2011. Kleppmann and collaborators at Ink & Switch developed Automerge, an open-source CRDT library for JSON documents, to make these algorithms available to application developers. == Adoption and community == Developer interest in the local-first approach grew after the 2019 paper spread on Hacker News and at developer conferences In August 2023, Wired published a feature article on the movement, describing it as an effort to reduce reliance on large cloud providers. The first Local-First Conf took place on 30 May 2024 in Berlin, with talks by Kleppmann and developers from companies including Linear and Anytype. The community has continued to expand, with regular "LoFi" meetups, a podcast (localfirst.fm), and a third edition of the conference planned for Berlin in July 2026. == Criticisms and limitations == Developers and commentators have pointed out practical difficulties with the local-first approach. Synchronizing data between multiple devices that may be offline for extended periods introduces complexity that cloud-based architectures avoid. Conflict resolution, even with CRDTs, can produce results that are technically consistent but semantically unexpected to users. Schema migrations across thousands of client devices running different application versions pose another difficulty that does not arise with server-side databases. Web browsers impose storage limits and may evict locally stored data. Safari, for instance, has been reported to clear IndexedDB data after seven days of inactivity on a given site, which undermines the assumption that local data is persistent. There is also disagreement within the local-first community about whether a fully decentralized architecture is required. The original manifesto described decentralization as the "logical end goal," but a number of products that identify as local-first still depend on centralized servers for authentication, backup, or synchronization. In a talk at Local-First Conf 2024, Kleppmann said the seven ideals are better understood as a "gradient" rather than a strict checklist.
Gallery software
Gallery software is software that helps the user publish or share photos, pictures, videos or other digital media. Most galleries are located on Web servers, where users are allowed to register and publish their pictures. Gallery software usually features automatic image resizing, allows digital media be categorized into sets, and allows comments. == Types == Early digital media publishing and sharing was done with imageboards. The boards are by topics, sometimes called "chan". Each discussion in a "chan" are started with a piece of digital media, and follow-up discussions can contain another piece too. Software works in this way: Futallaby, Danbooru. Traditionally, galleries are managed. An administrator maintains a set of or hierarchy of albums. The users can upload their digital media in one of the existing albums defined by an administrator, or create their own albums. The users with sufficient permission can re-categorise the digital media others uploaded. Often, the site's administrator can define which album the users are allowed to categorise their media into, or delete other user's content. Examples are open source galleries Coppermine, Gallery Project. There are decentralised gallery software that does not have an administrator for managing contents. Pinterest, Flickr and DeviantArt has been successful with this model. Open source gallery software MediaGoblin works in this way. Each user can create their own "collections", to categorise theirs or other users' media. However users cannot put media into other user's collections. Each user's category is separate. There is no centralised theme or hierarchy for the media.
ML.NET
ML.NET is a free software machine learning library for the C# and F# programming languages. It also supports Python models when used together with NimbusML. The preview release of ML.NET included transforms for feature engineering like n-gram creation, and learners to handle binary classification, multi-class classification, and regression tasks. Additional ML tasks like anomaly detection and recommendation systems have since been added, and other approaches like deep learning will be included in future versions. == Machine learning == ML.NET brings model-based Machine Learning analytic and prediction capabilities to existing .NET developers. The framework is built upon .NET Core and .NET Standard inheriting the ability to run cross-platform on Linux, Windows and macOS. Although the ML.NET framework is new, its origins began in 2002 as a Microsoft Research project named TMSN (text mining search and navigation) for use internally within Microsoft products. It was later renamed to TLC (the learning code) around 2011. ML.NET was derived from the TLC library and has largely surpassed its parent says Dr. James McCaffrey, Microsoft Research. Developers can train a Machine Learning Model or reuse an existing Model by a 3rd party and run it on any environment offline. This means developers do not need to have a background in Data Science to use the framework. Support for the open-source Open Neural Network Exchange (ONNX) Deep Learning model format was introduced from build 0.3 in ML.NET. The release included other notable enhancements such as Factorization Machines, LightGBM, Ensembles, LightLDA transform and OVA. The ML.NET integration of TensorFlow is enabled from the 0.5 release. Support for x86 & x64 applications was added to build 0.7 including enhanced recommendation capabilities with Matrix Factorization. A full roadmap of planned features have been made available on the official GitHub repo. The first stable 1.0 release of the framework was announced at Build (developer conference) 2019. It included the addition of a Model Builder tool and AutoML (Automated Machine Learning) capabilities. Build 1.3.1 introduced a preview of Deep Neural Network training using C# bindings for Tensorflow and a Database loader which enables model training on databases. The 1.4.0 preview added ML.NET scoring on ARM processors and Deep Neural Network training with GPU's for Windows and Linux. === Performance === Microsoft's paper on machine learning with ML.NET demonstrated it is capable of training sentiment analysis models using large datasets while achieving high accuracy. Its results showed 95% accuracy on Amazon's 9GB review dataset. === Model builder === The ML.NET CLI is a Command-line interface which uses ML.NET AutoML to perform model training and pick the best algorithm for the data. The ML.NET Model Builder preview is an extension for Visual Studio that uses ML.NET CLI and ML.NET AutoML to output the best ML.NET Model using a GUI. === Model explainability === AI fairness and explainability has been an area of debate for AI Ethicists in recent years. A major issue for Machine Learning applications is the black box effect where end users and the developers of an application are unsure of how an algorithm came to a decision or whether the dataset contains bias. Build 0.8 included model explainability API's that had been used internally in Microsoft. It added the capability to understand the feature importance of models with the addition of 'Overall Feature Importance' and 'Generalized Additive Models'. When there are several variables that contribute to the overall score, it is possible to see a breakdown of each variable and which features had the most impact on the final score. The official documentation demonstrates that the scoring metrics can be output for debugging purposes. During training & debugging of a model, developers can preview and inspect live filtered data. This is possible using the Visual Studio DataView tools. === Infer.NET === Microsoft Research announced the popular Infer.NET model-based machine learning framework used for research in academic institutions since 2008 has been released open source and is now part of the ML.NET framework. The Infer.NET framework utilises probabilistic programming to describe probabilistic models which has the added advantage of interpretability. The Infer.NET namespace has since been changed to Microsoft.ML.Probabilistic consistent with ML.NET namespaces. === NimbusML Python support === Microsoft acknowledged that the Python programming language is popular with Data Scientists, so it has introduced NimbusML the experimental Python bindings for ML.NET. This enables users to train and use machine learning models in Python. It was made open source similar to Infer.NET. === Machine learning in the browser === ML.NET allows users to export trained models to the Open Neural Network Exchange (ONNX) format. This establishes an opportunity to use models in different environments that don't use ML.NET. It would be possible to run these models in the client side of a browser using ONNX.js, a JavaScript client-side framework for deep learning models created in the Onnx format. === AI School Machine Learning Course === Along with the rollout of the ML.NET preview, Microsoft rolled out free AI tutorials and courses to help developers understand techniques needed to work with the framework.
Clinical quality management system
Clinical quality management systems (CQMS) are systems used in the life sciences sector (primarily in the pharmaceutical, biologics and medical device industries) designed to manage quality management best practices throughout clinical research and clinical study management. A CQMS system is designed to manage all of the documents, activities, tasks, processes, quality events, relationships, audits and training that must be administered and controlled throughout the life of a clinical trial. The premise of a CQMS is to bring together the activities led by two sectors of clinical research, Clinical Quality and Clinical Operations, to facilitate cross-functional activities to improve efficiencies and transparency and to encourage the use of risk mitigation and risk management practices at the clinical study level. Based on the principles of quality management systems (QMS) which are used in many industries to create a framework for defining and delivering quality outcomes, managing risk, and continual improvement. Many guidelines and governance bodies have been established to ensure a common approach within a given industry to a set of parameters used to identify the minimally acceptable standard for that industry. The pharmaceutical industry is no exception, with several trade groups (e.g. PhRMA, EFPIA, RQA, etc.) coming together to enhance collaboration. However, as noted by the Academy of Medical Sciences, there are increasingly complex and bureaucratic legal and ethical frameworks that innovators must work within to develop new medicines for patients. The historical pharmaceutical QMS applies primarily to good manufacturing practice as described in existing ISO (International Organization for Standardization) and ICH (International Committee on Harmonization) guidelines. "Good Manufacturing Practices (GMP) relate to quality control and quality assurance enabling companies in the pharmaceutical sector to minimize or eliminate instances of contamination, mix-ups, and errors. This in turn, protects the customer from purchasing a product which is ineffective or even dangerous." These standards have historically been applied to the manufacturing environment, appropriate to how they have been written. However, according to FDA as well as other regulatory bodies, "Implementation of ICH Q10 throughout the product lifecycle should facilitate innovation and continual improvement", implying that the same standards that apply to the manufacturing environment should also be applied to the clinical research space, earlier in the lifecycle of an investigational or marketed product. Accordingly, a CQMS is any system developed to apply these principles to clinical operations within an organization.
AI Dungeon
AI Dungeon is a single-player/multiplayer text adventure game which uses artificial intelligence (AI) to generate content and allows players to create and share adventures and custom prompts. The game's first version was made available in May 2019, and its second version (initially called AI Dungeon 2) was released on Google Colaboratory in December 2019. It was later ported that same month to its current cross-platform web application. The AI model was then reformed in July 2020. == Gameplay == AI Dungeon is a text adventure game that uses artificial intelligence to generate random storylines in response to player-submitted stimuli. In the game, players are prompted to choose a setting for their adventure (e.g. fantasy, mystery, apocalyptic, cyberpunk, zombies), followed by other options relevant to the setting (such as character class for fantasy settings). After beginning an adventure, four main interaction methods can be chosen for the player's text input: Do: Must be followed by a verb, allowing the player to perform an action. Say: Must be followed by dialogue sentences, allowing players to communicate with other characters. Story: Can be followed by sentences describing something that happens to progress the story, or that players want the AI to know for future events. See: Must be followed by a description, allowing the player to perceive events, objects, or characters. Using this command creates an AI generated image, and does not affect gameplay. The game adapts and responds to most actions the player enters. Providing blank inputs can be used to prompt the AI to generate further content, and the game also provides players with options to undo or redo or modify recent events to improve the game's narrative. Players can also tell the AI what elements to "remember" for reference in future parts of their playthrough. === User-generated content === In addition to AI Dungeon's pre-configured settings, players can create custom "adventures" from scratch by describing the setting in text format, which the AI will then generate a setting from. These custom adventures can be published for others to play, with an interface for browsing published adventures and leaving comments under them. === Multiplayer === AI Dungeon includes a multiplayer mode in which different players each have their own character and take turns interacting with the AI within the same game session. Multiplayer supports both online play across multiple devices or local play using a shared device. The game's hosts are able to supervise the AI and modify its output. Unlike the single-player game, in which actions and stories use second person narration, multiplayer game stories are presented using third-person narration. === Worlds === AI Dungeon allows players to set their adventures within specific "Worlds" that give context to the broader environment where the adventure takes place. This feature was first released with two different worlds available for selection: Xaxas, a "world of peace and prosperity"; and Kedar, a "world of dragons, demons, and monsters". == Development == === AI Dungeon Classic (Early GPT-2) === The first version of AI Dungeon (sometimes referred to as AI Dungeon Classic) was designed and created by Nick Walton of Brigham Young University's "Perception, Control, and Cognition" deep learning laboratory in March 2019 during a hackathon. Before this, Walton had been working as an intern for several companies in the field of autonomous vehicles. This creation used an early version of the GPT-2 natural-language-generating neural network, created by OpenAI, allowing it to generate its original adventure narratives. During his first interactions with GPT-2, Walton was partly inspired by the tabletop game Dungeons & Dragons (D&D), which he had played for the first time with his family a few months earlier: I realized that there were no games available that gave you the same freedom to do anything that I found in [Dungeons & Dragons] ... You can be so creative compared to other games. This led him to wonder if an AI could function as a dungeon master. Unlike later versions of AI Dungeon, the original did not allow players to specify any action they wanted. Instead, it generated a finite list of possible actions to choose from. This first version of the game was released to the public in May 2019. It is not to be confused with another GPT-2-based adventure game, GPT Adventure, created by Northwestern University neuroscience postgraduate student Nathan Whitmore, also released on Google Colab several months after the public release of AI Dungeon. === AI Dungeon 2 (Full GPT-2) === In November 2019, a new, "full" version of GPT-2 was released by OpenAI. This new model included support for 1.5 billion parameters (which determine the accuracy with which a machine learning model can perform a task), compared with the 126 million parameter version used in the earliest stages of AI Dungeon's development. The game was recreated by Walton, leveraging this new version of the model, and temporarily rebranded as AI Dungeon 2. AI Dungeon 2's AI was given more focused training compared to its predecessor, using genre-specific text. This training material included approximately 30 megabytes of content web-scraped from chooseyourstory.com (an online community website of content inspired by interactive gamebooks, written by contributors of multiple skill levels, using logic of differing complexity) and multiple D&D rulebooks and adventures. The new version was released in December 2019 as open-source software available on GitHub. It was accessible via Google Colab, an online tool for data scientists and AI researchers that allows for free execution of code on Google-hosted machines. It could also be run locally on a PC, but in both cases, it required players to download the full model, around 5 gigabytes of data. Within days of the initial release, this mandatory download resulted in bandwidth charges of over $20,000, forcing the temporary shut-down of the game until a peer-to-peer alternative solution was established. Due to the game's sudden and explosive growth that same month, however, it became closed-source, proprietary software and was relaunched by Walton's start-up development team, Latitude (with Walton taking on the role of CTO). This relaunch constituted mobile apps for iOS and Android (built by app developer Braydon Batungbacal) on December 17. Other members of this team included Thorsten Kreutz for the game's long-term strategy and the creator's brother, Alan Walton, for hosting infrastructure. At this time, Nick Walton also established a Patreon campaign to support the game's further growth (such as the addition of multiplayer and voice support, along with longer-term plans to include music and image content) and turn the game into a commercial endeavor, which Walton felt was necessary to cover the costs of delivering a higher-quality version of the game. AI Dungeon was one of the only known commercial applications to be based upon GPT-2. Following its first announcement in December 2019, a multiplayer mode was added to the game in April 2020. Hosting a game in this mode was originally restricted to premium subscribers, although any players could join a hosted game. === Dragon model release (GPT-3) === In July 2020, the developers introduced a premium-exclusive version of the AI model, named Dragon, which uses OpenAI's API for leveraging the GPT-3 model without maintaining a local copy (released on June 11, 2020). GPT-3 was trained with 570 gigabytes of text content (approximately one trillion words, with a $12 million development cost) and can support 175 billion parameters, compared to the 40 gigabytes of training content and 1.5 billion parameters of GPT-2. The free model was also upgraded to a less-advanced version of GPT-3 and was named Griffin. Speaking shortly after this release, on the differences between GPT-2 and GPT-3, Walton stated: [GPT-3 is] one of the most powerful AI models in the world... It's just much more coherent in terms of understanding who the characters are, what they're saying, what's going on in the story and just being able to write an interesting and believable story. In the latter half of 2020, the "Worlds" feature was added to AI Dungeon, providing players with a selection of overarching worlds in which their adventures can take place. In February 2021, it was announced that AI Dungeon's developers, Latitude, had raised $3.3 million in seed funding (led by NFX, with participation from Album VC and Griffin Gaming Partners) to "build games with 'infinite' story possibilities." This funding intended to move AI content creation beyond the purely text-based nature of AI Dungeon as it existed at the time. After its announcement on August 20, a new "See" interaction mode was made available for all players and added to the game on August 30, 2022. AI Dungeon was retired from Steam on March 12, 2024. == Reception == Approximate
Neural scaling law
In machine learning, a neural scaling law is an empirical scaling law that describes how neural network performance changes as key factors are scaled up or down. These factors typically include the number of parameters, training dataset size, and training cost. Some models also exhibit performance gains by scaling inference through increased test-time compute (TTC), extending neural scaling laws beyond training to the deployment phase. == Introduction == In general, a deep learning model can be characterized by four parameters: model size, training dataset size, training cost, and the post-training error rate (e.g., the test set error rate). Each of these variables can be defined as a real number, usually written as N , D , C , L {\displaystyle N,D,C,L} (respectively: parameter count, dataset size, computing cost, and loss). A neural scaling law is a theoretical or empirical statistical law between these parameters. There are also other parameters with other scaling laws. === Size of the model === In most cases, the model's size is simply the number of parameters. However, one complication arises with the use of sparse models, such as mixture-of-expert models. With sparse models, during inference, only a fraction of their parameters are used. In comparison, most other kinds of neural networks, such as transformer models, always use all their parameters during inference. === Size of the training dataset === The size of the training dataset is usually quantified by the number of data points within it. Larger training datasets are typically preferred, as they provide a richer and more diverse source of information from which the model can learn. This can lead to improved generalization performance when the model is applied to new, unseen data. However, increasing the size of the training dataset also increases the computational resources and time required for model training. With the "pretrain, then finetune" method used for most large language models, there are two kinds of training dataset: the pretraining dataset and the finetuning dataset. Their sizes have different effects on model performance. Generally, the finetuning dataset is less than 1% the size of pretraining dataset. In some cases, a small amount of high quality data suffices for finetuning, and more data does not necessarily improve performance. Many scaling laws, due to their inherent diminishing returns nature, value data based on a submodular set function which was shown in a paper on this topic. === Cost of training === Training cost is typically measured in terms of time (how long it takes to train the model) and computational resources (how much processing power and memory are required). It is important to note that the cost of training can be significantly reduced with efficient training algorithms, optimized software libraries, and parallel computing on specialized hardware such as GPUs or TPUs. The cost of training a neural network model is a function of several factors, including model size, training dataset size, the training algorithm complexity, and the computational resources available. In particular, doubling the training dataset size does not necessarily double the cost of training, because one may train the model for several times over the same dataset (each being an "epoch"). === Performance === The performance of a neural network model is evaluated based on its ability to accurately predict the output given some input data. Common metrics for evaluating model performance include: Negative log-likelihood per token (logarithm of perplexity) for language modeling; Accuracy, precision, recall, and F1 score for classification tasks; Mean squared error (MSE) or mean absolute error (MAE) for regression tasks; Elo rating in a competition against other models, such as gameplay or preference by a human judge. Performance can be improved by using more data, larger models, different training algorithms, regularizing the model to prevent overfitting, and early stopping using a validation set. When the performance is a number bounded within the range of [ 0 , 1 ] {\displaystyle [0,1]} , such as accuracy, precision, etc., it often scales as a sigmoid function of cost, as seen in the figures. == Examples == === (Hestness, Narang, et al, 2017) === The 2017 paper is a common reference point for neural scaling laws fitted by statistical analysis on experimental data. Previous works before the 2000s, as cited in the paper, were either theoretical or orders of magnitude smaller in scale. Whereas previous works generally found the scaling exponent to scale like L ∝ D − α {\displaystyle L\propto D^{-\alpha }} , with α ∈ { 0.5 , 1 , 2 } {\displaystyle \alpha \in \{0.5,1,2\}} , the paper found that α ∈ [ 0.07 , 0.35 ] {\displaystyle \alpha \in [0.07,0.35]} . Of the factors they varied, only task can change the exponent α {\displaystyle \alpha } . Changing the architecture optimizers, regularizers, and loss functions, would only change the proportionality factor, not the exponent. For example, for the same task, one architecture might have L = 1000 D − 0.3 {\displaystyle L=1000D^{-0.3}} while another might have L = 500 D − 0.3 {\displaystyle L=500D^{-0.3}} . They also found that for a given architecture, the number of parameters necessary to reach lowest levels of loss, given a fixed dataset size, grows like N ∝ D β {\displaystyle N\propto D^{\beta }} for another exponent β {\displaystyle \beta } . They studied machine translation with LSTM ( α ∼ 0.13 {\displaystyle \alpha \sim 0.13} ), generative language modelling with LSTM ( α ∈ [ 0.06 , 0.09 ] , β ≈ 0.7 {\displaystyle \alpha \in [0.06,0.09],\beta \approx 0.7} ), ImageNet classification with ResNet ( α ∈ [ 0.3 , 0.5 ] , β ≈ 0.6 {\displaystyle \alpha \in [0.3,0.5],\beta \approx 0.6} ), and speech recognition with two hybrid (LSTMs complemented by either CNNs or an attention decoder) architectures ( α ≈ 0.3 {\displaystyle \alpha \approx 0.3} ). === (Henighan, Kaplan, et al, 2020) === A 2020 analysis studied statistical relations between C , N , D , L {\displaystyle C,N,D,L} over a wide range of values and found similar scaling laws, over the range of N ∈ [ 10 3 , 10 9 ] {\displaystyle N\in [10^{3},10^{9}]} , C ∈ [ 10 12 , 10 21 ] {\displaystyle C\in [10^{12},10^{21}]} , and over multiple modalities (text, video, image, text to image, etc.). In particular, the scaling laws it found are (Table 1 of ): For each modality, they fixed one of the two C , N {\displaystyle C,N} , and varying the other one ( D {\displaystyle D} is varied along using D = C / 6 N {\displaystyle D=C/6N} ), the achievable test loss satisfies L = L 0 + ( x 0 x ) α {\displaystyle L=L_{0}+\left({\frac {x_{0}}{x}}\right)^{\alpha }} where x {\displaystyle x} is the varied variable, and L 0 , x 0 , α {\displaystyle L_{0},x_{0},\alpha } are parameters to be found by statistical fitting. The parameter α {\displaystyle \alpha } is the most important one. When N {\displaystyle N} is the varied variable, α {\displaystyle \alpha } ranges from 0.037 {\displaystyle 0.037} to 0.24 {\displaystyle 0.24} depending on the model modality. This corresponds to the α = 0.34 {\displaystyle \alpha =0.34} from the Chinchilla scaling paper. When C {\displaystyle C} is the varied variable, α {\displaystyle \alpha } ranges from 0.048 {\displaystyle 0.048} to 0.19 {\displaystyle 0.19} depending on the model modality. This corresponds to the β = 0.28 {\displaystyle \beta =0.28} from the Chinchilla scaling paper. Given fixed computing budget, optimal model parameter count is consistently around N o p t ( C ) = ( C 5 × 10 − 12 petaFLOP-day ) 0.7 = 9.0 × 10 − 7 C 0.7 {\displaystyle N_{opt}(C)=\left({\frac {C}{5\times 10^{-12}{\text{petaFLOP-day}}}}\right)^{0.7}=9.0\times 10^{-7}C^{0.7}} The parameter 9.0 × 10 − 7 {\displaystyle 9.0\times 10^{-7}} varies by a factor of up to 10 for different modalities. The exponent parameter 0.7 {\displaystyle 0.7} varies from 0.64 {\displaystyle 0.64} to 0.75 {\displaystyle 0.75} for different modalities. This exponent corresponds to the ≈ 0.5 {\displaystyle \approx 0.5} from the Chinchilla scaling paper. It's "strongly suggested" (but not statistically checked) that D o p t ( C ) ∝ N o p t ( C ) 0.4 ∝ C 0.28 {\displaystyle D_{opt}(C)\propto N_{opt}(C)^{0.4}\propto C^{0.28}} . This exponent corresponds to the ≈ 0.5 {\displaystyle \approx 0.5} from the Chinchilla scaling paper. The scaling law of L = L 0 + ( C 0 / C ) 0.048 {\displaystyle L=L_{0}+(C_{0}/C)^{0.048}} was confirmed during the training of GPT-3 (Figure 3.1 ). === Chinchilla scaling (Hoffmann, et al, 2022) === One particular scaling law ("Chinchilla scaling") states that, for a large language model (LLM) autoregressively trained for one epoch, with a cosine learning rate schedule, we have: { C = C 0 N D L = A N α + B D β + L 0 {\displaystyle {\begin{cases}C=C_{0}ND\\L={\frac {A}{N^{\alpha }}}+{\frac {B}{D^{\beta }}}+L_{0}\end{cases}}} where the variables are C {\displaystyle C} is the cost o
Eline Van der Velden
Eline van der Velden is a Dutch comedian, writer, actress and producer based in London, England. She is best known for her work creating Tilly Norwood, an AI-generated "actress". == Early life == Van der Velden was born on the Dutch island of Curaçao, Netherlands Antilles to Dutch businessman Steven van der Velden and physiotherapist Quirine van der Velden. She moved to the United Kingdom at age 14 to study drama and musical theatre at Tring Park School for the Performing Arts. She graduated with an MSc in physics from Imperial College London in 2008. == Career == She was nominated by the International Academy of Digital Arts and Sciences for the Lovie Awards and won Best Online Comedy in 2013 for two of her submitted entries. She has created multiple online shows such as Sketch My Life with London Hughes and Emily Hartridge and Match.com Parody. She became managing director of Makers Channel (makerschannel.co.uk), the first curated video platform in Europe in 2015. Makers Channel has been recently acquired by a Belgian media company De Persgroep, due to its success in the Netherlands. In 2016, she appeared in adverts for the Dutch shampoo brand Andrelon. Miss Holland, a comedy character created by Van der Velden, made headlines in 2016 as she asked the British public to teach her the national anthem. As an actress, she has starred in Dutch TV series De Troon, Beatrix and the Golden Calf-winning series Overspel. In Belgium, she appeared opposite Jamie Dornan in Flying Home. Van der Velden starred in the BBC Three series Putting It Out There, in which she challenges social perceptions of body hair, heels, spit, personal space, and authority figures. In 2018, she starred in the BBC One comedy series Soft Border Patrol and the BBC Three comedy series Miss Holland. In 2025, Particle6 Group, which Van der Velden founded in 2016, introduced Tilly Norwood, an AI-generated "actress" at the Zurich Film Festival. The announcement was met with outrage and a condemnation by the American actors' union SAG-AFTRA. == Awards and recognition == Miss Holland won the Best Online Comedy at the 2013 Lovie Awards, judged by Stephen Fry. The Match.com Parody video won Best Online Comedy People's Lovie Award, the people's vote. Miss Holland and Match.com Parody Date 1 were also featured in the 2013 Google Lovie Letters.