AI Email Response

AI Email Response — independent reviews, comparisons, pricing and step-by-step guides on Aizhi.

Voyages: The Trans-Atlantic Slave Trade Database

Voyages: The Trans-Atlantic Slave Trade Database is a database hosted at Rice University that aims to present all documentary material pertaining to the transatlantic slave trade. It is a sister project to African Origins. The database breaks down the kingdoms and countries that engaged in the Atlantic trade. By 2008, the project had gathered data on nearly 35,000 transatlantic slave voyages from 1501 to 1867. For each voyage they sought to establish dates, owners, vessels, captains, African visits, American destinations, numbers of slaves embarked, and numbers landed. They have been able to find much of this material for an estimated 80 percent of the entire transatlantic African slave trade. With corrections for missing voyages, the Project has estimated the entire size of the transatlantic slave trade with more comprehension, precision, and accuracy than before. They reckon that in 366 years, slaving vessels embarked about 12.5 million captives in Africa, and landed 10.7 million in the New World. A horrific discovery is a careful estimate that the Middle Passage took a toll of more than 1.8 million African lives. In this quantitative database, the numbers are enslaved people.
Read more →
MIT Computer Science and Artificial Intelligence Laboratory

Computer Science and Artificial Intelligence Laboratory (CSAIL) is a research institute at the Massachusetts Institute of Technology (MIT) formed by the 2003 merger of the Laboratory for Computer Science (LCS) and the Artificial Intelligence Laboratory (AI Lab). Housed within the Ray and Maria Stata Center, CSAIL is the largest on-campus laboratory as measured by research scope and membership. It is part of the Schwarzman College of Computing but is also overseen by the MIT Vice President of Research. == Research activities == CSAIL's research activities are organized around a number of semi-autonomous research groups, each of which is headed by one or more professors or research scientists. These groups are divided up into seven general areas of research: Artificial intelligence Computational biology Graphics and vision Language and learning Theory of computation Robotics Systems (includes computer architecture, databases, distributed systems, networks and networked systems, operating systems, programming methodology, and software engineering, among others) == History == Computing Research at MIT began with Vannevar Bush's research into a differential analyzer and Claude Shannon's electronic Boolean algebra in the 1930s, the wartime MIT Radiation Laboratory, the post-war Project Whirlwind and the Research Laboratory of Electronics (RLE), and MIT Lincoln Laboratory's SAGE in the early 1950s. At MIT, research in the field of artificial intelligence began in the late 1950s. === Project MAC === On July 1, 1963, Project MAC (the Project on Mathematics and Computation, later backronymed to Multiple Access Computer, Machine Aided Cognitions, or Man and Computer) was launched with a $2 million grant from the Defense Advanced Research Projects Agency (DARPA). Project MAC's original director was Robert Fano of MIT's Research Laboratory of Electronics (RLE). Fano decided to call MAC a "project" rather than a "laboratory" for reasons of internal MIT politics – if MAC had been called a laboratory, then it would have been more difficult to raid other MIT departments for research staff. The program manager responsible for the DARPA grant was J. C. R. Licklider, who had previously been at MIT conducting research in RLE, and would later succeed Fano as director of Project MAC. Project MAC would become famous for groundbreaking research in operating systems, artificial intelligence, and the theory of computation. Its contemporaries included Project Genie at Berkeley, the Stanford Artificial Intelligence Laboratory, and (somewhat later) University of Southern California's (USC's) Information Sciences Institute. An "AI Group" including Marvin Minsky (the director), John McCarthy (inventor of Lisp), and a talented community of computer programmers were incorporated into Project MAC. They were interested principally in the problems of vision, mechanical motion and manipulation, and language, which they view as the keys to more intelligent machines. In the 1960s and 1970s the AI Group developed a time-sharing operating system called Incompatible Timesharing System (ITS) which ran on PDP-6 and later PDP-10 computers. The early Project MAC community included Fano, Minsky, Licklider, Fernando J. Corbató, and a community of computer programmers and enthusiasts among others who drew their inspiration from former colleague John McCarthy. These founders envisioned the creation of a computer utility whose computational power would be as reliable as an electric utility. To this end, Corbató brought the first computer time-sharing system, Compatible Time-Sharing System (CTSS), with him from the MIT Computation Center, using the DARPA funding to purchase an IBM 7094 for research use. One of the early focuses of Project MAC would be the development of a successor to CTSS, Multics, which was to be the first high availability computer system, developed as a part of an industry consortium including General Electric and Bell Laboratories. In 1966, Scientific American featured Project MAC in the September thematic issue devoted to computer science, that was later published in book form. At the time, the system was described as having approximately 100 TTY terminals, mostly on campus but with a few in private homes. Only 30 users could be logged in at the same time. The project enlisted students in various classes to use the terminals simultaneously in problem solving, simulations, and multi-terminal communications as tests for the multi-access computing software being developed. === AI Lab and LCS === In the late 1960s, Minsky's artificial intelligence group was seeking more space, and was unable to get satisfaction from project director Licklider. Minsky found that although Project MAC as a single entity could not get the additional space he wanted, he could split off to form his own laboratory and then be entitled to more office space. As a result, the MIT AI Lab was formed in 1970, and many of Minsky's AI colleagues left Project MAC to join him in the new laboratory, while most of the remaining members went on to form the Laboratory for Computer Science. Talented programmers such as Richard Stallman, who used TECO to develop EMACS, flourished in the AI Lab during this time. Those researchers who did not join the smaller AI Lab formed the Laboratory for Computer Science and continued their research into operating systems, programming languages, distributed systems, and the theory of computation. Two professors, Hal Abelson and Gerald Jay Sussman, chose to remain neutral—their group was referred to variously as Switzerland and Project MAC for the next 30 years. Among much else, the AI Lab led to the invention of Lisp machines and their attempted commercialization by two companies in the 1980s: Symbolics and Lisp Machines Inc. === CSAIL === On the fortieth anniversary of Project MAC's establishment, July 1, 2003, LCS was merged with the AI Lab to form the MIT Computer Science and Artificial Intelligence Laboratory, or CSAIL. This merger created the largest laboratory (over 600 personnel) on the MIT campus. In 2018, CSAIL launched a five-year collaboration program with IFlytek, a company sanctioned the following year for allegedly using its technology for surveillance and human rights abuses in Xinjiang. In October 2019, MIT announced that it would review its partnerships with sanctioned firms such as iFlyTek and SenseTime. In April 2020, the agreement with iFlyTek was terminated. CSAIL moved from the School of Engineering to the newly formed Schwarzman College of Computing by February 2020. == Offices == From 1963 to 2004, Project MAC, LCS, the AI Lab, and CSAIL had their offices at 545 Technology Square, taking over more and more floors of the building over the years. In 2004, CSAIL moved to the new Ray and Maria Stata Center, which was built specifically to house it and other departments. == Outreach activities == The IMARA (from Swahili word for "power") group sponsors a variety of outreach programs that bridge the global digital divide. Its aim is to find and implement long-term, sustainable solutions which will increase the availability of educational technology and resources to domestic and international communities. These projects are run under the aegis of CSAIL and staffed by MIT volunteers who give training, install and donate computer setups in greater Boston, Massachusetts, Kenya, Native American Indian tribal reservations in the American Southwest such as the Navajo Nation, the Middle East, and Fiji Islands. The CommuniTech project strives to empower under-served communities through sustainable technology and education and does this through the MIT Used Computer Factory (UCF), providing refurbished computers to under-served families, and through the Families Accessing Computer Technology (FACT) classes, it trains those families to become familiar and comfortable with computer technology. == Notable researchers == (Including members and alumni of CSAIL's predecessor laboratories) MacArthur Fellows Tim Berners-Lee, Erik Demaine, Dina Katabi, Daniela L. Rus, Regina Barzilay, Peter Shor, Richard Stallman, and Joshua Tenenbaum Turing Award recipients Leonard M. Adleman, Fernando J. Corbató, Shafi Goldwasser, Butler W. Lampson, John McCarthy, Silvio Micali, Marvin Minsky, Ronald L. Rivest, Adi Shamir, Barbara Liskov, and Michael Stonebraker IJCAI Computers and Thought Award recipients Terry Winograd, Patrick Winston, David Marr, Gerald Jay Sussman, Rodney Brooks Rolf Nevanlinna Prize recipients Madhu Sudan, Peter Shor, Constantinos Daskalakis Gödel Prize recipients Shafi Goldwasser (two-time recipient), Silvio Micali, Maurice Herlihy, Charles Rackoff, Johan Håstad, Peter Shor, and Madhu Sudan Grace Murray Hopper Award recipients Robert Metcalfe, Shafi Goldwasser, Guy L. Steele, Jr., Richard Stallman, and W. Daniel Hillis Textbook authors Harold Abelson and Gerald Jay Sussman, Richard Stallman, Thomas H. Cormen, Charles E. Leiserson, Patrick Winston, Ronald L.
Read more →
Common Crawl

The Common Crawl Foundation (Common Crawl) is a nonprofit 501(c)(3) organization that crawls the web and freely provides its archives and datasets to the public. Common Crawl was founded by Gil Elbaz. The data had mostly been primarily used by researchers and some startups until the 2020s, when AI companies started training large language models using the data. In November 2025, an investigation by The Atlantic revealed that Common Crawl misled publishers when it claimed it respected paywalls in its scraping and it was not honoring requests from publishers to have their content removed from its databases. == History == Common Crawl was founded in 2007 in San Francisco. It began publishing its crawls in 2011. By 2013, sites like TinEye were building their products off of Common Crawl. The crawl reduces the reliance of companies and researchers on Google, which has the biggest dataset. Common Crawl was designed to have more and fresher data that was more efficient to analyze and utilize than the Wayback Machine created by the Internet Archive. By 2015, 1.8 billion webpages were on the Common Crawl, which started by crawling a list of URLs donated by the search engine Blekko. They use Amazon Web Services, which provides some of its services for free, allowing computing costs to average $2-4000/month. The Common Crawl website listed 30 studies based on Common Crawl data. Before 2023, Common Crawl was not very well known outside of academic researchers who utilize the data. Common Crawl received its first requests to redact information in 2023 and increasingly started seeing its crawler, CCBot, blocked. In 2023, it began receiving significant financial support from AI companies, including Anthropic and OpenAI, each of which donated $250,000. It was also used to train Google DeepMind's large language model Gemini. By April 2023, Common Crawl was capturing 3.1 billion webpages, with an estimated 5% of pages before 2021 containing hate speech or slurs. As of 2024, Common Crawl had been cited in more than 10,000 academic studies. By 2024, The Pile and Common Crawl had been the two main training datasets being used to train AI models. In November 2025, an investigation by technology journalist Alex Reisner for The Atlantic revealed that Common Crawl misled publishers when it claimed it respected paywalls in its scraping and when it said that it was honoring requests from publishers to have their content removed from its databases. It included misleading results in the public search function on its website that showed no entries for websites that had requested their archives be removed, when in fact those sites were still included in its scrapes used by AI companies. As of 2025, Reisner found that CCBot was the most widely-blocked bot by the top 1000 websites. A 2026 article in LWN.net discussed an advantage to services like Common Crawl being that it can limit the scraping costs to websites by allowing companies and researchers to download the data from Common Crawl instead of scraping it themselves. In April 2026, Common Crawl experimentally began to distribute its data through Hugging Face Storage Bucket, in addition to its standard storage on Amazon S3. == Organization == Peter Norvig and Joi Ito have served on the advisory board. Rich Skrenta is the executive director. It has received funding almost exclusively from the Elbaz Family Foundation Trust until 2023 when it started receiving donations from the AI industry. == Refined versions == A number of organizations take raw Common Crawl data and refine it into datasets that exclude edgy content or are otherwise higher-quality for their purposes, such as FineWeb, DCLM and C4. === Colossal Clean Crawled Corpus === Google version of the Common Crawl is called the Colossal Clean Crawled Corpus, or C4 for short. It was constructed for the training of the T5 language model series in 2019. As of 2023, there were some concerns over copyrighted content in the C4 as well as racist content. A 2024 study found that 45% of content was explicitly restricted by websites' terms of service to be used for purposes like AI training by for-profit companies.
Read more →
Aurora (supercomputer)

Aurora is an exascale supercomputer that was sponsored by the United States Department of Energy (DOE) and designed by Intel and Cray for Argonne National Laboratory. It was briefly the second fastest supercomputer in the world from November 2023 to June 2024. The cost was estimated in 2019 to be US$500 million. Olivier Franza is the chief architect and principal investigator of this design. == History == In 2013 DOE presented a proposal for an "exascale" supercomputer, capable of speeds in the neighborhood of 1 exaFLOP (1018 floating point mathematical operations per second) with a maximum power consumption of 20 megawatts (MW) by 2020. Aurora was first announced in 2015 and to be finished in 2018. It was expected to have a speed of 180 petaFLOPS which would be around the speed of Summit. Aurora was meant to be the most powerful supercomputer at the time of its launch and to be built by Cray with Intel processors. Later, in 2017, Intel announced that Aurora would be delayed to 2021 but scaled up to 1 exaFLOP. In March 2019, DOE said that it would build the first supercomputer with a performance of one exaFLOP in the United States in 2021. In October 2020, DOE said that Aurora would be delayed again for a further six months, and would no longer be the first exascale computer in the US. In late October 2021 Intel announced that Aurora would now exceed 2 exaFLOPS in peak double-precision compute – That claim however never was realized. The system was fully installed on June 22, 2023. In May 2024, Aurora appeared at number two on the Top500 supercomputer list, with a performance of 1.012 exaFLOPS, marking the second entry of an exascale capable system on the Top500. == Usage == Functions include research on brain structure, nuclear fusion, low carbon technologies, subatomic particles, cancer and cosmology. It will also develop new materials that will be useful for batteries and more efficient solar cells. It is to be available to the general scientific community. == Architecture == Aurora has 10,624 nodes, with each node being composed of two Intel Xeon Max processors, six Intel Max series GPUs and a unified memory architecture, providing a maximum computing power of 130 teraFLOPS per node. It has around 10 petabytes of memory and 230 petabytes of storage. The machine is stated to consume around 39 MW of power. For comparison, the fastest computer in the world today, El Capitan uses 30 MW, while another Top 500 System, Frontier uses 24 MW.
Read more →
My Drama

My Drama (also may be stylised as MyDrama) is a global streaming service specializing in vertical video series for Duanju. It is owned by the company Holywater Tech. The platform focuses on short-form, emotional storytelling optimized for smartphone viewing, offering content in over 30 languages across 190 countries. == History == My Drama was launched in 2024 by Holywater Tech, founded by Ukrainian entrepreneur Bogdan Nesvit and Anatolii Kasianov. The service gained international traction as part of a growing market for short-form vertical storytelling, influenced by mobile-first entertainment trends. My Drama primarily streams serialized vertical dramas, which are short-form episodes around 1-2 minutes in length designed for mobile consumption. Many series are adaptations of successful stories originally published on Holywater Tech's book platform My Passion. The platform employs AI technology in areas such as content recommendation and story generation, and is one of several Holywater apps focused on interactive entertainment. In 2024, My Drama won a People's Voice award at the 28th Annual Webby Awards. In 2025, My Drama received a Gold Award at the MUSE Creative Awards in the Mobile App: Video Streaming Services category. In 2025, the company received strategic investment from Fox Entertainment, aimed at expanding content creation capabilities and producing over 200 vertical video series. As of 2025, My Drama has produced over 56 titles and reached more than 40 million lifetime users, according to media reports. In January 2026, Holywater Tech raised $22 million in funding to expand its microdrama business in the United States. The investment round was led by Horizon Capital, with participation from U.S.-based investors including Endeavor Catalyst and Wheelhouse. The funding is intended to support the development of Holywater Tech's mobile-first vertical video platform, My Drama, as well as the company's AI-driven content initiatives, such as AI-assisted comics and anime. In February 2026, Holywater bought Jeynix, a studio that uses AI for special effects. This deal helps the company make better-quality shows and translate them into different languages much faster. == Partnerships == In 2024, Holywater Tech entered a partnership with Latin American studio Elefantec Global to distribute vertical dramas in Spanish-language markets. In early 2026, Fox Entertainment entered into a partnership with content creator Dhar Mann to produce a slate of 40 original vertical microdrama series. Under the agreement, the series debut exclusively on the My Drama platform, while global distribution is managed by Fox Entertainment Global. == Reception == My Drama has been highlighted in discussions of the global rise of vertical short drama platforms and has been compared with similar apps such as ReelShort and DramaBox.
Read more →
Linde–Buzo–Gray algorithm

The Linde–Buzo–Gray algorithm (named after its creators Yoseph Linde, Andrés Buzo and Robert M. Gray, who designed it in 1980) is an iterative vector quantization algorithm to improve a small set of vectors (codebook) to represent a larger set of vectors (training set), such that it will be locally optimal. It combines Lloyd's Algorithm with a splitting technique in which larger codebooks are built from smaller codebooks by splitting each code vector in two. The core idea of the algorithm is that by splitting the codebook such that all code vectors from the previous codebook are present, the new codebook must be as good as the previous one or better. == Description == The Linde–Buzo–Gray algorithm may be implemented as follows: algorithm linde-buzo-gray is input: set of training vectors training, codebook to improve old-codebook output: codebook that is twice the size and better or as good as old-codebook new-codebook ← {} for each old-codevector in old-codebook do insert old-codevector into new-codebook insert old-codevector + 𝜖 into new-codebook where 𝜖 is a small vector return lloyd(new-codebook, training) algorithm lloyd is input: codebook to improve, set of training vectors training output: improved codebook do previous-codebook ← codebook clusters ← divide training into |codebook| clusters, where each cluster contains all vectors in training who are best represented by the corresponding vector in codebook for each cluster cluster in clusters do the corresponding code vector in codebook ← the centroid of all training vectors in cluster while difference in error representing training between codebook and previous-codebook > 𝜖 return codebook
Read more →
PyTorch

PyTorch is an open-source deep learning library, originally developed by Meta Platforms and currently developed with support from the Linux Foundation. The successor to Torch, PyTorch provides a high-level API that builds upon optimised, low-level implementations of deep learning algorithms and architectures, such as the Transformer, or SGD. Notably, this API simplifies model training and inference to a few lines of code. PyTorch allows for automatic parallelization of training and, internally, implements CUDA bindings that speed training further by leveraging GPU resources. PyTorch utilises the tensor as a fundamental data type, similarly to NumPy. Training is facilitated by a reversed automatic differentiation system, Autograd, that constructs a directed acyclic graph of the operations (and their arguments) executed by a model during its forward pass. With a loss, backpropagation is then undertaken. As of 2025, PyTorch remains one of the most popular deep learning libraries, alongside others such as TensorFlow and Keras. It can be installed using Anaconda package managers. A number of commercial deep learning architectures are built on top of PyTorch, including ChatGPT, Tesla Autopilot, Uber's Pyro, and Hugging Face's Transformers. == History == In 2001, Torch was written and released under a GPL. It was a machine-learning library written in C++ and CUDA, supporting methods including neural networks, support vector machines (SVM), hidden Markov models, etc. Around 2010, it was rewritten by Ronan Collobert, Clement Farabet and Koray Kavuckuoglu. This was known as Torch7 or LuaTorch. This was written so that the backend was in C and the frontend was in Lua. In mid-2016, some developers refactored it to decouple the frontend and the backend, with strong influence from torch-autograd and Chainer. In turn, torch-autograd was influenced by HIPS/autograd. Development on Torch7 ceased in 2018 and was subsumed by the PyTorch project. Meta (formerly known as Facebook) operates both PyTorch and Convolutional Architecture for Fast Feature Embedding (Caffe2), but models defined by the two frameworks were mutually incompatible. The Open Neural Network Exchange (ONNX) project was created by Meta and Microsoft in September 2017 to decouple deep learning frameworks from hardware-specific runtimes, allowing models to be converted between frameworks and optimized for execution providers like NVIDIA’s TensorRT. Caffe2 was merged into PyTorch at the end of March 2018. In September 2022, Meta announced that PyTorch would be governed by the independent PyTorch Foundation, a newly created subsidiary of the Linux Foundation. PyTorch 2.0 was released on 15 March 2023, introducing TorchDynamo, a Python-level compiler that makes code run up to two times faster, along with significant improvements in training and inference performance across major cloud platforms. == PyTorch tensors == PyTorch defines a class called Tensor (torch.Tensor) to store and operate on homogeneous multidimensional rectangular arrays of numbers. PyTorch supports various sub-types of multi-dimensional arrays, or Tensors. PyTorch Tensors are similar to NumPy Arrays, but can also be operated on by a CUDA-capable NVIDIA GPU. PyTorch has also been developing support for other GPU platforms, for example, AMD's ROCm and Apple's Metal Framework. == PyTorch neural networks == PyTorch defines a module called nn (torch.nn) to describe neural networks and to support training. This module offers a comprehensive collection of building blocks for neural networks, including various layers and activation functions, enabling the construction of complex models. Networks are built by inheriting from the torch.nn module and defining the sequence of operations in the forward() function. == PyTorch Serialized File Format == Pytorch can save and load models using its own file format, which is a ZIP64 archive containing the model weights in a Python pickle file, and other information such as the byte order. The file extensions .pt and .pth are commonly used for these files. == Example == The following program shows the low-level functionality of the library with a simple example. The following code block defines a neural network with linear layers using the nn module.
Read more →
Layer (deep learning)

A layer in a deep learning model is a structure or network topology in the model's architecture, which takes information from the previous layers and then passes it to the next layer. == Layer types == The first type of layer is the Dense layer, also called the fully-connected layer, and is used for abstract representations of input data. In this layer, neurons connect to every neuron in the preceding layer. In multilayer perceptron networks, these layers are stacked together. The Convolutional layer is typically used for image analysis tasks. In this layer, the network detects edges, textures, and patterns. The outputs from this layer are then fed into a fully-connected layer for further processing. See also: CNN model. The Pooling layer is used to reduce the size of data input. The Recurrent layer is used for text processing with a memory function. Similar to the Convolutional layer, the output of recurrent layers are usually fed into a fully-connected layer for further processing. See also: RNN model. The Normalization layer adjusts the output data from previous layers to achieve a regular distribution. This results in improved scalability and model training. A Hidden layer is any of the layers in a Neural Network that aren't the input or output layers. == Differences with layers of the neocortex == There is an intrinsic difference between deep learning layering and neocortical layering: deep learning layering depends on network topology, while neocortical layering depends on intra-layers homogeneity.
Read more →
Spatial anti-aliasing

In digital signal processing, spatial anti-aliasing is a technique for minimizing the distortion artifacts (aliasing) when representing a high-resolution image at a lower resolution. Anti-aliasing is used in digital photography, computer graphics, digital audio, and many other applications. Anti-aliasing means removing signal components that have a higher frequency than is able to be properly resolved by the recording (or sampling) device. This removal is done before (re)sampling at a lower resolution. When sampling is performed without removing this part of the signal, it causes undesirable artifacts such as black-and-white noise. In signal acquisition and audio, anti-aliasing is often done using an analog anti-aliasing filter to remove the out-of-band component of the input signal prior to sampling with an analog-to-digital converter. In digital photography, optical anti-aliasing filters made of birefringent materials smooth the signal in the spatial optical domain. The anti-aliasing filter essentially blurs the image slightly in order to reduce the resolution to or below that achievable by the digital sensor (the larger the pixel pitch, the lower the achievable resolution at the sensor level). == Examples == In computer graphics, anti-aliasing improves the appearance of "jagged" polygon edges, or "jaggies", so they are smoothed out on the screen. However, it incurs a performance cost for the graphics card and uses more video memory. The level of anti-aliasing determines how smooth polygon edges are (and how much video memory it consumes). Near the top of an image with a receding checker-board pattern, the image is difficult to recognise and often not considered aesthetically pleasing. In contrast, when anti-aliased the checker-board near the top blends into grey, which is usually the desired effect when the resolution is insufficient to show the detail. Even near the bottom of the image, the edges appear much smoother in the anti-aliased image. Multiple methods exist, including the sinc filter, which is considered a better anti-aliasing algorithm. When magnified, it can be seen how anti-aliasing interpolates the brightness of the pixels at the boundaries to produce grey pixels since the space is occupied by both black and white tiles. These help make the sinc filter antialiased image appear much smoother than the original. In a simple diamond image, anti-aliasing blends the boundary pixels; this reduces the aesthetically jarring effect of the sharp, step-like boundaries that appear in the aliased graphic. Anti-aliasing is often applied in rendering text on a computer screen, to suggest smooth contours that better emulate the appearance of text produced by conventional ink-and-paper printing. Particularly with fonts displayed on typical LCD screens, it is common to use subpixel rendering techniques like ClearType. Sub-pixel rendering requires special colour-balanced anti-aliasing filters to turn what would be severe colour distortion into barely-noticeable colour fringes. Equivalent results can be had by making individual sub-pixels addressable as if they were full pixels, and supplying a hardware-based anti-aliasing filter as is done in the OLPC XO-1 laptop's display controller. Pixel geometry affects all of this, whether the anti-aliasing and sub-pixel addressing are done in software or hardware. == Simplest approach to anti-aliasing == The most basic approach to anti-aliasing a pixel is determining what percentage of the pixel is occupied by a given region in the vector graphic - in this case a pixel-sized square, possibly transposed over several pixels - and using that percentage as the colour. A Python program producing a basic plot of a single, white-on-black anti-aliased point using the method is as follows: This method is generally best suited for simple graphics, such as basic lines or curves, and applications that would otherwise have to convert absolute coordinates to pixel-constrained coordinates, such as 3D graphics. It is a fairly fast function, but it is relatively low-quality, and gets slower as the complexity of the shape increases. For purposes requiring very high-quality graphics or very complex vector shapes, this will probably not be the best approach. Note: The plot_antialiased_point routine above cannot blindly set the colour value to the percent calculated. It must add the new value to the existing value at that location up to a maximum of 1. Otherwise, the brightness of each pixel will be equal to the darkest value calculated in time for that location which produces a very bad result. For example, if one point sets a brightness level of 0.90 for a given pixel and another point calculated later barely touches that pixel and has a brightness of 0.05, the final value set for that pixel should be 0.95, not 0.05. For more sophisticated shapes, the algorithm may be generalized as rendering the shape to a pixel grid with higher resolution than the target display surface (usually a multiple that is a power of 2 to reduce distortion), then using bicubic interpolation to determine the average intensity of each real pixel on the display surface. == Signal processing approach to anti-aliasing == In this approach, the ideal image is regarded as a signal. The image displayed on the screen is taken as samples, at each (x,y) pixel position, of a filtered version of the signal. Ideally, one would understand how the human brain would process the original signal, and provide an on-screen image that will yield the most similar response by the brain. The most widely accepted analytic tool for such problems is the Fourier transform; this decomposes a signal into basis functions of different frequencies, known as frequency components, and gives us the amplitude of each frequency component in the signal. The waves are of the form: cos ⁡ ( 2 j π x ) cos ⁡ ( 2 k π y ) {\displaystyle \ \cos(2j\pi x)\cos(2k\pi y)} where j and k are arbitrary non-negative integers. There are also frequency components involving the sine functions in one or both dimensions, but for the purpose of this discussion, the cosine will suffice. The numbers j and k together are the frequency of the component: j is the frequency in the x direction, and k is the frequency in the y direction. The goal of an anti-aliasing filter is to greatly reduce frequencies above a certain limit, known as the Nyquist frequency, so that the signal will be accurately represented by its samples, or nearly so, in accordance with the sampling theorem; there are many different choices of detailed algorithm, with different filter transfer functions. Current knowledge of human visual perception is not sufficient, in general, to say what approach will look best. == Two dimensional considerations == The previous discussion assumes that the rectangular mesh sampling is the dominant part of the problem. The filter usually considered optimal is not rotationally symmetrical, as shown in this first figure; this is because the data is sampled on a square lattice, not using a continuous image. This sampling pattern is the justification for doing signal processing along each axis, as it is traditionally done on one dimensional data. Lanczos resampling is based on convolution of the data with a discrete representation of the sinc function. If the resolution is not limited by the rectangular sampling rate of either the source or target image, then one should ideally use rotationally symmetrical filter or interpolation functions, as though the data were a two dimensional function of continuous x and y. The sinc function of the radius has too long a tail to make a good filter (it is not even square-integrable). A more appropriate analog to the one-dimensional sinc is the two-dimensional Airy disc amplitude, the 2D Fourier transform of a circular region in 2D frequency space, as opposed to a square region. One might consider a Gaussian plus enough of its second derivative to flatten the top (in the frequency domain) or sharpen it up (in the spatial domain), as shown. Functions based on the Gaussian function are natural choices, because convolution with a Gaussian gives another Gaussian whether applied to x and y or to the radius. Similarly to wavelets, another of its properties is that it is halfway between being localized in the configuration (x and y) and in the spectral (j and k) representation. As an interpolation function, a Gaussian alone seems too spread out to preserve the maximum possible detail, and thus the second derivative is added. As an example, when printing a photographic negative with plentiful processing capability and on a printer with a hexagonal pattern, there is no reason to use sinc function interpolation. Such interpolation would treat diagonal lines differently from horizontal and vertical lines, which is like a weak form of aliasing. == Practical real-time anti-aliasing approximations == There are only a handful of primitives used at the lowest level in a real-time rend
Read more →
GuideGeek

GuideGeek is an AI-powered travel assistant that was launched by travel publisher Matador Network in April 2023 and is accessed by users through Instagram, WhatsApp and Facebook Messenger to plan itineraries or provide travel tips and recommendations. It uses generative artificial intelligence technology from OpenAI. Matador Network is a San Francisco-based digital media company and online travel publication with millions of monthly visitors and social media followers. == Features == Users message GuideGeek questions about travel and receive customized answers and itineraries that are pulled from ChatGPT in addition to over 1,000 additional travel-specific integrations such as live flight, hotel and vacation rental data. Travelers can specify their budget and needs to generate custom itineraries. GuideGeek is not an app and does not require the user to download anything, instead relying on messaging apps such as Instagram to connect users with the AI. GuideGeek is free to use, doesn't include ads, and doesn't sell user data. Matador Network has a team of staff members monitoring conversations to correct them if the AI makes a false statement; for example, one user incorrectly inputted “Crete Freeze” instead of “Crete, Greece”, and the AI made up a fictional soft serve company. Using a technique known as reinforcement learning from human feedback (RLHF), the accuracy of GuideGeek increased to 98%, according to Matador Network CEO, Ross Borden. == Destination partnerships == Matador Network is monetizing GuideGeek via white-label partnerships with tourism bureaus and destination marketing organizations (DMOs). As of March 2024, it had over a dozen such clients. Estes Park, Colorado, was one of the first DMOs to partner with Matador for a custom version of GuideGeek called “Rocky Mountain Roamer.” For Discover Greece, Matador created Pythia, a custom AI named after the high priestess of the Temple of Apollo at Delphi. As Borden explained to Travel + Leisure, “Visitors to the Discover Greece website will find Pythia in the bottom right corner, and they can converse with the AI like a friend who knows everything about Greece.” Other DMOs who have partnerships with GuideGeek include the Aruba Tourism Authority, Visit Reno Tahoe, Illinois Office of Tourism, and Tourism Richmond. == Awards == In recognition of GuideGeek, Fast Company named Matador Network to its 2024 list of Most Innovative Companies. Following growth driven by the launch of GuideGeek, Matador Network was ranked on the 2024 Inc. 5000 list of fastest-growing private companies in America. The 2024 Skift IDEA Awards recognized Matador Network as a finalist in the category of Best Use of AI for GuideGeek's customized AI for the travel industry. == Michael Motamedi experiment == Travel influencer and chef Michael Motamedi traveled the world with his wife Vanessa Salas and their 2-year-old daughter on a six-month trip (which was later extended to a full year) led by GuideGeek. The family started off in Morocco before heading to Spain and continuing east. The experiment became the basis of a web series called “No Fixed Address.” Motamedi used GuideGeek's AI to select countries the family visited, where they ate, and what sites they saw. Motamedi and Salas first tested out the technology in April 2023 while using the chatbot to plan a date night in Mexico City. GuideGeek provided speakeasy and drink recommendations as well as local history facts.
Read more →
Learning rule

An artificial neural network's learning rule or learning process is a method, mathematical logic or algorithm which improves the network's performance and/or training time. Usually, this rule is applied repeatedly over the network. It is done by updating the weight and bias levels of a network when it is simulated in a specific data environment. A learning rule may accept existing conditions (weights and biases) of the network, and will compare the expected result and actual result of the network to give new and improved values for the weights and biases. Depending on the complexity of the model being simulated, the learning rule of the network can be as simple as an XOR gate or mean squared error, or as complex as the result of a system of differential equations. The learning rule is one of the factors which decides how fast or how accurately the neural network can be developed. Depending on the process to develop the network, there are three main paradigms of machine learning: supervised learning, unsupervised learning, and reinforcement learning. == Background == A lot of the learning methods in machine learning work similar to each other, and are based on each other, which makes it difficult to classify them in clear categories. But they can be broadly understood in 4 categories of learning methods, though these categories don't have clear boundaries and they tend to belong to multiple categories of learning methods - Hebbian - Neocognitron, Brain-state-in-a-box Gradient Descent - ADALINE, Hopfield Network, Recurrent Neural Network Competitive - Learning Vector Quantisation, Self-Organising Feature Map, Adaptive Resonance Theory Stochastic - Boltzmann Machine, Cauchy Machine Though these learning rules might appear to be based on similar ideas, they do have subtle differences, as they are a generalisation or application over the previous rule, and hence it makes sense to study them separately based on their origins and intents. === Hebbian Learning === Developed by Donald Hebb in 1949 to describe biological neuron firing. In the mid-1950s it was also applied to computer simulations of neural networks. Δ w i = η x i y {\displaystyle \Delta w_{i}=\eta x_{i}y} Where η {\displaystyle \eta } represents the learning rate, x i {\displaystyle x_{i}} represents the input of neuron i, and y is the output of the neuron. It has been shown that Hebb's rule in its basic form is unstable. Oja's Rule, BCM Theory are other learning rules built on top of or alongside Hebb's Rule in the study of biological neurons. ==== Perceptron Learning Rule (PLR) ==== The perceptron learning rule originates from the Hebbian assumption, and was used by Frank Rosenblatt in his perceptron in 1958. The net is passed to the activation (transfer) function and the function's output is used for adjusting the weights. The learning signal is the difference between the desired response and the actual response of a neuron. The step function is often used as an activation function, and the outputs are generally restricted to -1, 0, or 1. The weights are updated with w new = w old + η ( t − o ) x i {\displaystyle w_{\text{new}}=w_{\text{old}}+\eta (t-o)x_{i}} where "t" is the target value and "o" is the output of the perceptron, and η {\displaystyle \eta } is called the learning rate. The algorithm converges to the correct classification if: the training data is linearly separable η {\displaystyle \eta } is sufficiently small (though smaller η {\displaystyle \eta } generally means a longer learning time and more epochs) It should also be noted that a single layer perceptron with this learning rule is incapable of working on linearly non-separable inputs, and hence the XOR problem cannot be solved using this rule alone === Backpropagation === Seppo Linnainmaa in 1970 is said to have developed the Backpropagation Algorithm but the origins of the algorithm go back to the 1960s with many contributors. It is a generalisation of the least mean squares algorithm in the linear perceptron and the Delta Learning Rule. It implements gradient descent search through the space possible network weights, iteratively reducing the error, between the target values and the network outputs. ==== Widrow-Hoff Learning (Delta Learning Rule) ==== Similar to the perceptron learning rule but with different origin. It was developed for use in the ADALINE network, which differs from the Perceptron mainly in terms of the training. The weights are adjusted according to the weighted sum of the inputs (the net), whereas in perceptron the sign of the weighted sum was useful for determining the output as the threshold was set to 0, -1, or +1. This makes ADALINE different from the normal perceptron. Delta rule (DR) is similar to the Perceptron Learning Rule (PLR), with some differences: Error (δ) in DR is not restricted to having values of 0, 1, or -1 (as in PLR), but may have any value DR can be derived for any differentiable output/activation function f, whereas in PLR only works for threshold output function Sometimes only when the Widrow-Hoff is applied to binary targets specifically, it is referred to as Delta Rule, but the terms seem to be used often interchangeably. The delta rule is considered to a special case of the back-propagation algorithm. Delta rule also closely resembles the Rescorla-Wagner model under which Pavlovian conditioning occurs. === Competitive Learning === Competitive learning is considered a variant of Hebbian learning, but it is special enough to be discussed separately. Competitive learning works by increasing the specialization of each node in the network. It is well suited to finding clusters within data. Models and algorithms based on the principle of competitive learning include vector quantization and self-organizing maps (Kohonen maps).
Read more →
Yale shooting problem

The Yale shooting problem is a conundrum or scenario in formal situational logic on which early logical solutions to the frame problem fail. The name of this problem comes from a scenario proposed by its inventors, Steve Hanks and Drew McDermott, working at Yale University when they proposed it. In this scenario, Fred (later identified as a turkey) is initially alive and a gun is initially unloaded. Loading the gun, waiting for a moment, and then shooting the gun at Fred is expected to kill Fred. However, if inertia is formalized in logic by minimizing the changes in this situation, then it cannot be uniquely proved that Fred is dead after loading, waiting, and shooting. In one solution, Fred indeed dies; in another (also logically correct) solution, the gun becomes mysteriously unloaded and Fred survives. Technically, this scenario is described by two fluents (a fluent is a condition that can change truth value over time): a l i v e {\displaystyle alive} and l o a d e d {\displaystyle loaded} . Initially, the first condition is true and the second is false. Then, the gun is loaded, some time passes, and the gun is fired. Such problems can be formalized in logic by considering four time points 0 {\displaystyle 0} , 1 {\displaystyle 1} , 2 {\displaystyle 2} , and 3 {\displaystyle 3} , and turning every fluent such as a l i v e {\displaystyle alive} into a predicate a l i v e ( t ) {\displaystyle alive(t)} depending on time. A direct formalization of the statement of the Yale shooting problem in logic is the following one: a l i v e ( 0 ) {\displaystyle alive(0)} ¬ l o a d e d ( 0 ) {\displaystyle \neg loaded(0)} t r u e → l o a d e d ( 1 ) {\displaystyle true\rightarrow loaded(1)} l o a d e d ( 2 ) → ¬ a l i v e ( 3 ) {\displaystyle loaded(2)\rightarrow \neg alive(3)} The first two formulae represent the initial state. The third formula formalizes the effect of loading the gun at time 1 {\displaystyle 1} . The fourth formula formalizes the effect of shooting at Fred at time 2 {\displaystyle 2} . This is a simplified formalization in which action names are neglected and the effects of actions are directly specified for the time points in which the actions are executed. See situation calculus for details. The formulae above, while being direct formalizations of the known facts, do not suffice to correctly characterize the domain. Indeed, ¬ a l i v e ( 1 ) {\displaystyle \neg alive(1)} is consistent with all these formulae, although there is no reason to believe that Fred dies before the gun has been shot. The problem is that the formulae above only include the effects of actions, but do not specify that all fluents not changed by the actions remain the same. In other words, a formula a l i v e ( 0 ) ≡ a l i v e ( 1 ) {\displaystyle alive(0)\equiv alive(1)} must be added to formalize the implicit assumption that loading the gun only changes the value of l o a d e d {\displaystyle loaded} and not the value of a l i v e {\displaystyle alive} . The necessity of a large number of formulae stating the obvious fact that conditions do not change unless an action changes them is known as the frame problem. An early solution to the frame problem was based on minimizing the changes. In other words, the scenario is formalized by the formulae above (that specify only the effects of actions) and by the assumption that the changes in the fluents over time are as minimal as possible. The rationale is that the formulae above enforce all effect of actions to take place, while minimization should restrict the changes to exactly those due to the actions. In the Yale shooting scenario, one possible evaluation of the fluents in which the changes are minimized is the following one. This is the expected solution. It contains two fluent changes: l o a d e d {\displaystyle loaded} becomes true at time 1 and a l i v e {\displaystyle alive} becomes false at time 3. The following evaluation also satisfies all formulae above. In this evaluation, there are still two changes only: l o a d e d {\displaystyle loaded} becomes true at time 1 and false at time 2. As a result, this evaluation is considered a valid description of the evolution of the state, although there is no valid reason to explain l o a d e d {\displaystyle loaded} being false at time 2. The fact that minimization of changes leads to wrong solution is the motivation for the introduction of the Yale shooting problem. While the Yale shooting problem has been considered a severe obstacle to the use of logic for formalizing dynamical scenarios, solutions to it have been known since the late 1980s. One solution involves the use of predicate completion in the specification of actions: in this solution, the fact that shooting causes Fred to die is formalized by the preconditions: alive and loaded, and the effect is that alive changes value (since alive was true before, this corresponds to alive becoming false). By turning this implication into an if and only if statement, the effects of shooting are correctly formalized. (Predicate completion is more complicated when there is more than one implication involved.) A solution proposed by Erik Sandewall was to include a new condition of occlusion, which formalizes the “permission to change” for a fluent. The effect of an action that might change a fluent is therefore that the fluent has the new value, and that the occlusion is made (temporarily) true. What is minimized is not the set of changes, but the set of occlusions being true. Another constraint specifying that no fluent changes unless occlusion is true completes this solution. The Yale shooting scenario is also correctly formalized by the Reiter version of the situation calculus, the fluent calculus, and the action description languages. In 2005, the 1985 paper in which the Yale shooting scenario was first described received the AAAI Classic Paper award. In spite of being a solved problem, that example is still sometimes mentioned in recent research papers, where it is used as an illustrative example (e.g., for explaining the syntax of a new logic for reasoning about actions), rather than being presented as a problem.
Read more →
Automated medical scribe

Automated medical scribes (also called artificial intelligence scribes, AI scribes, digital scribes, virtual scribes, ambient AI scribes, AI documentation assistants, and digital/virtual/smart clinical assistants) are tools for transcribing medical speech, such as patient consultations and dictated medical notes. Many also produce summaries of consultations. Automated medical scribes based on large language models (LLMs, commonly called "AI", short for "artificial intelligence") increased drastically in popularity in 2024. There are privacy and antitrust concerns. Accuracy concerns also exist, and intensify in situations in which tools try to go beyond transcribing and summarizing, and are asked to format information by its meaning, since LLMs do not deal well with meaning (see weak artificial intelligence). Medics using these scribes are generally expected to understand the ethical and legal considerations, and supervise the outputs. The privacy protections of automated medical scribes vary widely. While it is possible to do all the transcription and summarizing locally, with no connection to the internet, most closed-source providers require that data be sent to their own servers over the internet, processed there, and the results sent back (as with digital voice assistants). Some retailers say their tools use zero-knowledge encryption (meaning that the service provider can't access the data). Others explicitly say that they use patient data to train their AIs, or rent or resell it to third parties; the nature of privacy protections used in such situations is unclear, and they are likely not to be fully effective. Most providers have not published any safety or utility data in academic journals, and are not responsive to requests from medical researchers studying their products. == Privacy == Some providers unclear about what happens to user data. Some may sell data to third parties. Some explicitly send user data to for-profit tech companies for secondary purposes, which may not be specified. Some require users to sign consents to such reuse of their data. Some ingest user data to train the software, promising to anonymize it; however, deanonymization may be possible (that is, it may become obvious who the patient is). It is intrinsically impossible to prevent an LLM from correlating its inputs; they work by finding similar patterns across very large data sets. Some information on the patient will be known from other sources (for instance, information that they were injured in an incident on a certain day might be available from the news media; information that they attended specific appointment locations at specific times is probably available to their cellphone provider/apps/data brokers; information about when they had a baby is probably implied by their online shopping records; and they might mention lifestyle changes to their doctor and on a forum or blog). The software may correlate such information with the "anonymized" clinical consultation record, and, asked about the named patient, provide information which they only told their doctor privately. Because a patient's record is all about the same patient, it is all unavoidably linked; in very many cases, medical histories are intrinsically identifiable. Depending on how common a condition and what other data is available, K-anonymity may be useless. Differential privacy could theoretically preserve privacy. Data broker companies like Google, Amazon, Apple and Microsoft have produced or bought up medical scribes, some of which use user data for secondary purposes, which has led to antitrust concerns. Transfer of patient records for AI training has, in the past, prompted legal action. Open-source programs typically do all the transcription locally, on the doctor's own computer. Open-source software is widely used in healthcare, with some national public healthcare bodies holding hack days. === Data resale and commercialization === Several AI medical scribe providers include terms in their service agreements that allow the reuse, sale, or commercialization of de-identified or user-submitted data. Although such data are generally described as anonymized or aggregated, these practices have raised ethical concerns among clinicians and privacy advocates regarding secondary uses of medical information beyond clinical documentation. Freed, an AI transcription and scribe platform, states in its Terms of Use that it may "collect, use, publish, disseminate, sell, transfer, and otherwise exploit" de-identified and aggregated data derived from user inputs. OpenEvidence similarly states that it may "collect, use, transfer, sell, and disclose non-personal information and customer usage data for any purpose including commercial uses." Doximity, which offers an AI-enabled medical scribe as part of its physician platform, grants itself a "nonexclusive, irrevocable, worldwide, perpetual, unlimited, assignable, sublicensable, royalty-free" license to "copy, prepare derivative works from, improve, distribute, publish, ... analyze, index, tag, [and] commercialize" content submitted by users, subject to its privacy policy. Because these terms allow broad secondary use—including sale, licensing, model-training, derivative works, and commercial exploitation of de-identified or user-submitted data—some commentators have recommended that clinicians review data-handling provisions carefully when adopting AI-scribe tools, particularly in clinical environments where patient privacy and regulatory compliance are critical. === Encryption === Multifactor authentication for access to the data is expected practice. Typically, Diffie–Hellman key exchange is used for encryption; this is the standard method commonly used for things like online banking. This encryption is expensive but not impossible to break; it is not generally considered safe against eavesdroppers with the resources of a nation-state. If content is encrypted between the client and the service provider's remote server (transport cryptography), then the server has an unencrypted copy. This is necessary if the data is used by the service provider (for instance, to train the software). Zero-knowledge encryption implies that the only unencrypted copy is at the client, and the server cannot decrypt the data any more easily than a monster-in-the-middle attacker. == Platforms == Scribes may operate on desktops, laptop, or mobile computers, under a variety of operating systems. These vary in their risks; for instance, mobiles can be lost. The underlying mobile or desktop operating systems are also part of the trusted computing base, and if they are not secure, the software relying on them cannot be secure either. Some AI medical scribe platforms are designed to operate as cloud-based applications that generate structured clinical documentation from clinician–patient conversations. These systems may offer features such as real-time transcription, document generation, and integration with electronic health record (EHR) systems. == Confabulation, omissions, and other errors == Like other LLMs, medical-scribe LLMs are prone to hallucinations, where they make up content based on statistically associations between their training data and the transcription audio. LLMs do not distinguish between trying to transcribe the audio and guessing what words will come next, but perform both processes mixed together. They are especially likely to take short silences or non-speech noises and invent some sort of speech to transcribe them as. LLM medical scribes have been known to confabulate racist and otherwise prejudiced content; this is partly because the training datasets of many LLMs contain pseudoscientific texts about medical racism. They may misgender patients. A survey found that most doctors preferred, in principle, that scribes be trained on data reviewed by medical subject experts. Relevant, accurate training data increases the probability of an accurate transcription, but does not guarantee accuracy. Software trained on thousands of real clinical conversations generated transcripts with lower word error rates. Software trained on manually-transcribed training data did better than software trained with automatically transcribed training data such as YouTube captions. Autoscribes omit parts of the conversation classes as irrelevant. The may wrongly classify pertinent information as irrelevant and omit it. They may also confuse historic and current symptoms, or otherwise misclassify information. They may also simply wrongly transcribe the speech, writing something incorrect instead. If clinicians do not carefully check the recording, such mistakes could make their way into their medical records and cause patient harms. == Patient consent == Professional organizations generally require that scribes be used only with patient consent; some bodies may require written consent. Medics must also abide by local surveillance laws, which may criminalize recording pri
Read more →
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
Read more →
HTK (software)

HTK (Hidden Markov Model Toolkit) is a proprietary software toolkit for handling HMMs. It is mainly intended for speech recognition, but has been used in many other pattern recognition applications that employ HMMs, including speech synthesis, character recognition and DNA sequencing. Originally developed at the Machine Intelligence Laboratory (formerly known as the Speech Vision and Robotics Group) of the Cambridge University Engineering Department (CUED), HTK is now being widely used among researchers who are working on HMMs.
Read more →