Nicholas Carlini is an American researcher affiliated with Anthropic and previously with Google DeepMind who has published research in the fields of computer security and machine learning. He is known for his work on adversarial machine learning, particularly his work on the Carlini & Wagner attack in 2016. This attack was particularly useful in defeating defensive distillation, a method used to increase model robustness, and has since been effective against other defenses against adversarial input. In 2018, Carlini demonstrated an attack on Mozilla's DeepSpeech model, showing that hidden commands could be embedded in speech inputs, which the model would execute even if they were inaudible to humans. He also led a team at UC Berkeley that successfully broke seven out of nine defenses against adversarial attacks presented at the 2018 International Conference on Learning Representations. In addition to his work on adversarial attacks, Carlini has made significant contributions to understanding the privacy risks of machine learning models. In 2020, he revealed that large language models, like GPT-2, could memorize and output personally identifiable information. His research demonstrated that this issue worsened with larger models, and he later showed similar vulnerabilities in generative image models, such as Stable Diffusion. == Life and career == Nicholas Carlini obtained his Bachelor of Arts in Computer Science and Mathematics from the University of California, Berkeley, in 2013. He then continued his studies at the same university, where he pursued a PhD under the supervision of David Wagner, completing it in 2018. Carlini became known for his work on adversarial machine learning. In 2016, he worked alongside Wagner to develop the Carlini & Wagner attack, a method of generating adversarial examples against machine learning models. The attack was proved to be useful against defensive distillation, a popular mechanism where a student model is trained based on the features of a parent model to increase the robustness and generalizability of student models. The attack gained popularity when it was shown that the methodology was also effective against most other defenses, rendering them ineffective. In 2018, Carlini demonstrated an attack against Mozilla Foundation's DeepSpeech model where he showed that by hiding malicious commands inside normal speech input the speech model would respond to the hidden commands even when the commands were not discernible by humans. In the same year, Carlini and his team at UC Berkeley showed that out of the 11 papers presenting defenses to adversarial attacks accepted in that year's ICLR conference, seven of the defenses could be broken. Since 2021, he and his team have been working on large language models, creating a questionnaire where humans typically scored 35% whereas AI models scored in the 40%, with GPT-3 getting 38% which could be improved to 40% through few shot prompting. The best performer in the test was UnifiedQA, a model developed by Google specifically for answer questions and answer sets. Carlini has also developed methods to cause large language models like ChatGPT to answer harmful questions like how to construct bombs. He is also known for his work studying the privacy of machine learning models. In 2020, he showed for the first time that large language models would memorize some of the text data that they were trained on. For example, he found that GPT-2 could output personally identifiable information. He then led an analysis of larger models and studied how memorization increased with model size. Then, in 2022 he showed the same vulnerability in generative image models, and specifically diffusion models, by showing that Stable Diffusion could output images of people's faces that it was trained on. Following on this, Carlini then showed that ChatGPT would also sometimes output exact copies of webpages it was trained on, including personally identifiable information. Some of these studies have since been referenced by the courts in debating the copyright status of AI models. == Other work == Carlini received the Best of Show award at the 2020 IOCCC for implementing a tic-tac-toe game entirely with calls to printf, expanding on work from a research paper of his from 2015. The judges commented on his submission "This year's Best of Show (carlini) is such a novel way of obfuscation that it would be worth of a special mention in the (future) Best of IOCCC list!". [sic] == Awards == Best Student Paper Award, IEEE S&P 2017 ("Towards Evaluating the Robustness of Neural Networks") Best Paper Award, ICML 2018 ("Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples") Distinguished Paper Award, USENIX 2021 ("Poisoning the Unlabeled Dataset of Semi-Supervised Learning") Distinguished Paper Award, USENIX 2023 ("Tight Auditing of Differentially Private Machine Learning") Best Paper Award, ICML 2024 ("Stealing Part of a Production Language Model") Best Paper Award, ICML 2024 ("Considerations for Differentially Private Learning with Large-Scale Public Pretraining")
Framebuffer
A framebuffer (frame buffer, or sometimes framestore) is a portion of random-access memory (RAM) containing a bitmap that drives a video display. It is a memory buffer containing data representing all the pixels in a complete video frame. Modern video cards contain framebuffer circuitry in their cores. This circuitry converts an in-memory bitmap into a video signal that can be displayed on a computer monitor. In computing, a screen buffer is a part of computer memory used by a computer application for the representation of the content to be shown on the computer display. The screen buffer may also be called the video buffer, the regeneration buffer, or regen buffer for short. The phrase "screen buffer” refers to a logical function, while video memory refers to a hardware storage location. In particular, the screen buffer may be placed in the main RAM, the video memory, or some other hardware location. To reduce latency and avoid screen tearing, multiple frames can be buffered, and this technique is called multiple buffering. When this is so, at any time, only one frame would be visible, and the others would not be. The currently invisible frames are located in the off-screen buffer. The information in the buffer typically consists of color values for every pixel to be shown on the display. Color values are commonly stored in 1-bit binary (monochrome), 4-bit palettized, 8-bit palettized, 16-bit high color and 24-bit true color formats. An additional alpha channel is sometimes used to retain information about pixel transparency. The total amount of memory required for the framebuffer depends on the resolution of the output signal, and on the color depth or palette size. == History == Computer researchers had long discussed the theoretical advantages of a framebuffer but were unable to produce a machine with sufficient memory at an economically practicable cost. In 1947, the Manchester Baby computer used a Williams tube, later the Williams-Kilburn tube, to store 1024 bits on a cathode-ray tube (CRT) memory and displayed on a second CRT. Other research labs were exploring these techniques with MIT Lincoln Laboratory achieving a 4096 display in 1950. A color-scanned display was implemented in the late 1960s, called the Brookhaven RAster Display (BRAD), which used a drum memory and a television monitor. In 1969, A. Michael Noll of Bell Telephone Laboratories, Inc. implemented a scanned display with a frame buffer, using magnetic-core memory. A year or so later, the Bell Labs system was expanded to display an image with a color depth of three bits on a standard color TV monitor. The vector graphics used in the computer had to be converted for the scanned graphics of a TV display. In the early 1970s, the development of MOS memory (metal–oxide–semiconductor memory) integrated-circuit chips, particularly high-density DRAM (dynamic random-access memory) chips with at least 1 kb memory, made it practical to create, for the first time, a digital memory system with framebuffers capable of holding a standard video image. This led to the development of the SuperPaint system by Richard Shoup at Xerox PARC in 1972. Shoup was able to use the SuperPaint framebuffer to create an early digital video-capture system. By synchronizing the output signal to the input signal, Shoup was able to overwrite each pixel of data as it shifted in. Shoup also experimented with modifying the output signal using color tables. These color tables allowed the SuperPaint system to produce a wide variety of colors outside the range of the limited 8-bit data it contained. This scheme would later become commonplace in computer framebuffers. In 1974, Evans & Sutherland released the first commercial framebuffer, the Picture System, costing about $15,000. It was capable of producing resolutions of up to 512 by 512 pixels in 8-bit grayscale, and became a boon for graphics researchers who did not have the resources to build their own framebuffer. The New York Institute of Technology would later create the first 24-bit color system using three of the Evans & Sutherland framebuffers. Each framebuffer was connected to an RGB color output (one for red, one for green and one for blue), with a Digital Equipment Corporation PDP 11/04 minicomputer controlling the three devices as one. In 1975, the UK company Quantel produced the first commercial full-color broadcast framebuffer, the Quantel DFS 3000. It was first used in TV coverage of the 1976 Montreal Olympics to generate a picture-in-picture inset of the Olympic flaming torch while the rest of the picture featured the runner entering the stadium. The rapid improvement of integrated-circuit technology made it possible for many of the home computers of the late 1970s to contain low-color-depth framebuffers. Today, nearly all computers with graphical capabilities utilize a framebuffer for generating the video signal. Amiga computers, created in the 1980s, featured special design attention to graphics performance and included a unique Hold-And-Modify framebuffer capable of displaying 4096 colors. Framebuffers also became popular in high-end workstations and arcade system boards throughout the 1980s. SGI, Sun Microsystems, HP, DEC and IBM all released framebuffers for their workstation computers in this period. These framebuffers were usually of a much higher quality than could be found in most home computers, and were regularly used in television, printing, computer modeling and 3D graphics. Framebuffers were also used by Sega for its high-end arcade boards, which were also of a higher quality than on home computers. == Display modes == Framebuffers used in personal and home computing often had sets of defined modes under which the framebuffer can operate. These modes reconfigure the hardware to output different resolutions, color depths, memory layouts and refresh rate timings. In the world of Unix machines and operating systems, such conveniences were usually eschewed in favor of directly manipulating the hardware settings. This manipulation was far more flexible in that any resolution, color depth and refresh rate was attainable – limited only by the memory available to the framebuffer. An unfortunate side-effect of this method was that the display device could be driven beyond its capabilities. In some cases, this resulted in hardware damage to the display. More commonly, it simply produced garbled and unusable output. Modern CRT monitors fix this problem through the introduction of protection circuitry. When the display mode is changed, the monitor attempts to obtain a signal lock on the new refresh frequency. If the monitor is unable to obtain a signal lock or if the signal is outside the range of its design limitations, the monitor will ignore the framebuffer signal and possibly present the user with an error message. LCD monitors tend to contain similar protection circuitry, but for different reasons. Since the LCD must digitally sample the display signal (thereby emulating an electron beam), any signal that is out of range cannot be physically displayed on the monitor. == Color palette == Framebuffers have traditionally supported a wide variety of color modes. Due to the expense of memory, most early framebuffers used 1-bit (2 colors per pixel), 2-bit (4 colors), 4-bit (16 colors) or 8-bit (256 colors) color depths. The problem with such small color depths is that a full range of colors cannot be produced. The solution to this problem was indexed color, which adds a lookup table to the framebuffer. Each color stored in framebuffer memory acts as a color index. The lookup table serves as a palette with a limited number of different colors, while the rest is used as an index table. Here is a typical indexed 256-color image and its own palette (shown as a rectangle of swatches): In some designs, it was also possible to write data to the lookup table (or switch between existing palettes) on the fly, allowing dividing the picture into horizontal bars with their own palette and thus rendering an image that had a far wider palette. For example, viewing an outdoor shot photograph, the picture could be divided into four bars: the top one with emphasis on sky tones, the next with foliage tones, the next with skin and clothing tones, and the bottom one with ground colors. This required each palette to have overlapping colors, but, carefully done, allowed great flexibility. == Memory access == While framebuffers are commonly accessed via a memory mapping directly to the CPU memory space, this is not the only method by which they may be accessed. Framebuffers have varied widely in the methods used to access memory. Some of the most common are: Mapping the entire framebuffer to a given memory range. Port commands to set each pixel, range of pixels or palette entry. Mapping a memory range smaller than the framebuffer memory, then bank switching as necessary. The framebuffer organization may be packed pixel or planar. The framebuffer may be all
AI effect
The AI effect is a phenomenon in which advances in artificial intelligence lead to a redefinition of what is considered intelligence, such that capabilities achieved by AI systems are no longer regarded as examples of "real" intelligence. The concept has been used to describe both a cognitive tendency and a sociotechnical pattern, in which successful AI techniques are reclassified as routine computation or absorbed into other domains. Historian Pamela McCorduck described this as a recurring feature of AI research, noting in her 2004 book Machines Who Think that once a problem is solved, it is no longer considered evidence of intelligence. Researcher Rodney Brooks similarly observed in 2002 that once systems are understood, they are often regarded as "just computation". == Definition == The AI effect refers to a shift in how intelligence is defined as machines acquire new capabilities. Tasks such as playing chess, recognizing speech, or interpreting images were historically considered indicators of intelligence, but after successful automation they are often reclassified as routine computation. McCorduck described this as an "odd paradox", in which successful AI systems are assimilated into other domains, leaving AI researchers to focus on unsolved problems. The phenomenon is often interpreted as an instance of moving the goalposts. A commonly cited formulation is Tesler's theorem, often expressed as "AI is whatever hasn't been done yet". When problems are not fully formalised, they may be described using models involving human computation, such as human-assisted Turing machines. == Historical examples == === Game playing === Early AI systems capable of playing games such as checkers and chess were initially regarded as demonstrations of machine intelligence. As these systems improved and became better understood, their achievements were often reinterpreted as examples of computation rather than intelligence. The victory of IBM's Deep Blue over Garry Kasparov in 1997 is a frequently cited example. Critics argued that the system relied on brute-force methods rather than genuine understanding. === Pattern recognition === Technologies such as optical character recognition and speech recognition were once considered core problems in artificial intelligence. As these systems became reliable and widely deployed, they were increasingly treated as standard engineering solutions. === Integration into applications === Many techniques originally developed within AI research have been incorporated into broader technological systems, including marketing, automation, and software applications. Michael Swaine reported in 2007 that AI advances are often presented as developments in other fields. Marvin Minsky observed that successful AI innovations often evolve into separate disciplines. Nick Bostrom noted in 2006 that widely adopted technologies are often no longer labeled as AI. == Contemporary discussion == The AI effect continues to be discussed in the context of recent advances in machine learning, particularly large language models and other generative AI systems. As these systems have become more widely used, some researchers and commentators have noted that their capabilities are frequently described as statistical or mechanical once understood, rather than as intelligence. A 2016 survey of artificial intelligence also noted that AI systems are increasingly embedded in everyday applications, reinforcing earlier observations that successful AI technologies tend to become normalized and no longer identified as AI. At the same time, the widespread commercial use of artificial intelligence has led to greater visibility of the field, contrasting with earlier periods in which AI techniques were often present but unacknowledged. == Interpretations == === Cognitive bias === Some authors describe the AI effect as a cognitive bias in which expectations of intelligence shift as machines achieve new capabilities. === Sociotechnical perspective === Another interpretation emphasizes how technologies are reclassified over time as they become widespread and commercially successful. === Philosophical debate === Some philosophers argue that reclassification reflects genuine conceptual distinctions rather than bias. == Historical context == During periods such as the AI winter, researchers sometimes avoided the term "artificial intelligence" due to negative perceptions. In the 21st century, however, the term "AI" has become widely used in public discourse and marketing. == Broader implications == The AI effect has been linked to broader questions about human uniqueness and the nature of intelligence. Michael Kearns suggested that people may seek to preserve a special role for humans. Similar patterns have been observed in studies of animal cognition. Herbert A. Simon noted that artificial intelligence can provoke strong emotional reactions.
Machine unlearning
Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin
Weak artificial intelligence
Weak artificial intelligence (weak AI) is artificial intelligence that implements a limited part of the mind, or, as narrow AI, artificial narrow intelligence (ANI), is focused on one narrow task. Weak AI is contrasted with strong AI, which can be interpreted in various ways: Artificial general intelligence (AGI): a machine with the ability to apply intelligence to any problem, rather than just one specific problem. Artificial superintelligence (ASI): a machine with a vastly superior intelligence to the average human being. Artificial consciousness: a machine that has consciousness, sentience and mind (John Searle uses "strong AI" in this sense). Narrow AI can be classified as being "limited to a single, narrowly defined task. Most modern AI systems would be classified in this category." Artificial general intelligence is conversely the opposite. == Applications and risks == Some examples of narrow AI are AlphaGo, self-driving cars, robot systems used in the medical field, and diagnostic doctors. Narrow AI systems are sometimes dangerous if unreliable. And the behavior that it follows can become inconsistent. It could be difficult for the AI to grasp complex patterns and get to a solution that works reliably in various environments. This "brittleness" can cause it to fail in unpredictable ways. Narrow AI failures can sometimes have significant consequences. It could for example cause disruptions in the electric grid, damage nuclear power plants, cause global economic problems, and misdirect autonomous vehicles. Medicines could be incorrectly sorted and distributed. Also, medical diagnoses can ultimately have serious and sometimes deadly consequences if the AI is faulty or biased. Simple AI programs have already worked their way into society, oftentimes unnoticed by the public. Autocorrection for typing, speech recognition for speech-to-text programs, and vast expansions in the data science fields are examples. Narrow AI has also been the subject of some controversy, including resulting in unfair prison sentences, discrimination against women in the workplace for hiring, resulting in death via autonomous driving, among other cases. Despite being "narrow" AI, recommender systems are efficient at predicting user reactions based on their posts, patterns, or trends. For instance, TikTok's "For You" algorithm can determine a user's interests or preferences in less than an hour. Some other social media AI systems are used to detect bots that may be involved in propaganda or other potentially malicious activities. == Weak AI versus strong AI == John Searle contests the possibility of strong AI (by which he means conscious AI). He further believes that the Turing test (created by Alan Turing and originally called the "imitation game", used to assess whether a machine can converse indistinguishably from a human) is not accurate or appropriate for testing whether an AI is "strong". Scholars such as Antonio Lieto have argued that the current research on both AI and cognitive modelling are perfectly aligned with the weak-AI hypothesis (that should not be confused with the "general" vs "narrow" AI distinction) and that the popular assumption that cognitively inspired AI systems espouse the strong AI hypothesis is ill-posed and problematic since "artificial models of brain and mind can be used to understand mental phenomena without pretending that that they are the real phenomena that they are modelling" (as, on the other hand, implied by the strong AI assumption).
AltStore
AltStore is an alternative app store for the iOS and iPadOS[1] mobile operating systems, which allows users to download applications that are not available on the App Store, most commonly tweaked apps, jailbreak apps, and apps including paid apps on the app store. It was publicly announced on September 25, 2019, and launched on September 28. == History == Riley Testut is an American developer who began to work on AltStore after Apple declined to allow his Nintendo emulator Delta on the App Store. Since Xcode allowed him to temporarily install his Delta app to his iOS device for 7 days of testing, he created AltStore in 2019 to replicate this functionality, which could be extended to other .ipa files. As of 2022, AltStore had been downloaded 1.5 million times. In the following years, AltStore expanded beyond its initial sideloading functionality. The platform was founded by Testut, with Shane Gill later joining as co-founder. AltStore was initially supported through Patreon contributions from its user community, and later saw increased adoption following regulatory developments in the European Union that enabled broader third-party app distribution. The project has also been involved in notable industry collaborations, including a partnership with Epic Games. == Features == AltStore exploits a loophole in the Xcode developer platform, which allows developers to sideload their own apps which they are working on without needing to jailbreak. Sideloaded apps are signed like a developer project for testing and will expire after 7 days with a free account or one year with a paid developer account, by which they will need to be refreshed or reinstalled.
CHAOS (chess)
CHAOS (Chess Heuristics and Other Stuff) is a chess playing program that was developed by programmers working at the RCA Systems Programming division in the late 1960s. It played competitively in computer chess competitions in the 1970s and 1980s. It differed from other programs of that era in its look-ahead philosophy, choosing to use chess knowledge to evaluate fewer positions and continuations as opposed to simple evaluations that relied on deep look-ahead to avoid bad moves. == Introduction == CHAOS was originally developed by Ira Ruben, Fred Swartz, Victor Berman, Joe Winograd and William Toikka while working at RCA in Cinnaminson, NJ. Its name is an acronym for 'Chess Heuristics and Other Stuff.' Program development moved to the Computing Center of the University of Michigan when Swartz changed jobs, and Mike Alexander joined the development group. Swartz, Alexander and Berman were continuously group members from that point onward in CHAOS' evolution, as others of the original authors left and new members contributed episodically. Chess Senior Master Jack O'Keefe contributed to CHAOS' development from about 1980 onwards. CHAOS was written in Fortran, except for low-level board representation manipulations written in assembly language or C. Due to this portability, it ran on RCA, Univac and IBM-compatible mainframes in its lifetime. CHAOS heralds from the mainframe computing era when only machines of that capacity were able to play at a high level. Consequently, development and testing could only take place at off-peak times for production use of the machine. In a competition, CHAOS had to run on a dedicated mainframe with a telephone link to the match venue. In its later years, CHAOS ran on computers on the machine assembly floor of Amdahl Corporation on MTS. == Background == === Chess and artificial intelligence === Mathematicians Claude Shannon and Alan Turing, working separately, were the first to view playing chess as a challenge to machines. Working for AT&T / Bell Labs with its access to telephone switching equipment, Shannon built a relay-based machine that learned how to work its way through a two-dimensional, 5x5 cell maze in 1949. Shannon viewed this as an analogue of the way that organisms learn things about their natural environment. There is a random element to searching it, a memory element to benefit from the search outcome, and a reward element that reinforces learning when the global outcome is favorable to the organism. Soon afterward, Shannon wrote a mathematical analysis of the game of chess, published in 1950. Like with the maze, he broke down game play into the necessary elements for reinforcement learning. Associated with each board configuration a move will be made from, there is a numerical score. To decide what move to make, a player wants to maximize their own position's score after the move and to minimize their opponent's score (a minimax view). Since there are about 32 possible moves at each of the early stages of the game, and about 40 moves and responses in each game, then there are about 32 80 {\displaystyle 32^{80}} or about 10 120 {\displaystyle 10^{120}} possible games - an impossibly large set to evaluate completely. Therefore, there must be a way to limit the number of moves to look ahead for to find the best one. Reducing the game to these few key elements provided a way to think about human intelligence in general. Shannon became part of a wider group using computing machines to mimic aspects of human intelligence that grew into the general idea of artificial intelligence. (Other members of this group were John McCarthy, Herbert Simon, Allen Newell, Alan Kotok, Alex Bernstein and Richard Greenblatt.) The paradigm that evolved was that there was a quantification of the position on the board into a score, an evaluation method to find favorable outcomes (minimax, later alpha-beta pruning), and a strategy to manage the combinatorial explosion of the look-ahead possibilities. By the early 1960s, there were computer programs that played chess at a rudimentary level. They used very simple evaluation functions for each position and tried to search as far forward as was practical given the time constraints and available compute power. Naturally, programmers optimized their code to use the available computing resources. This led to a major philosophical divide among chess programs: those that tried to evaluate as many positions as possible, and those that tried to evaluate the most promising move sequences as deeply as possible. CHAOS was firmly in the camp believing only the most promising moves should be evaluated in depth. Said Swartz, "The 'brute force people' ... look at every (possible move) no matter what garbage it is. Most moves are just terrible, terrible moves, and most computing time is being spent on pure garbage." The program spent more time evaluating each board position in the expectation that it would find the most promising lines of play to explore in depth. In 1983, the then-fastest chess program (Belle) evaluated 110,000 positions per second, and typical programs 1000–50,000 per second, whereas CHAOS evaluated about 50-100 per second. === Machine learning and strategies to manage search === From about 1949 onward, Arthur Samuel began work for IBM on machine learning, culminating in a checkers-playing program in 1952 and publications on the topic. Concurrently, Christopher Strachey created Checkers, a program to play the board game of checkers in 1951, but it had no capacity to learn from its play. Checkers was chosen by both authors because it was simpler than chess yet contained the basic characteristics of an intellectual activity, and, in Samuel's view, was a test-bed in which heuristic procedures and learning processes could be evaluated quickly. Checker playing programs introduced the notion of the game tree and evaluating play to various depths to choose the best move. The complexity of chess, however, promoted it to the status of an analogue for human intelligence, and it attracted computer scientists' attention, who referred to it as research into artificial intelligence (AI). Like checkers, it required a numerical assessment of each arrangement of chess pieces on a board. It also required looking ahead to future moves to decide how to play the present position. Due to the enormous number of possible moves, there had to be a way to confine the look-ahead search to the most promising lines of play. From these factors, the notion of minimax score evaluation developed and, later, alpha-beta tree pruning to abandon looking at positions worse than any that have already been examined. === Chess search strategies === The AI community viewed artificial intelligence as comprising two parts: a way to symbolically quantify the knowledge in hand (a chess board position), and a set of heuristics to limit look-ahead to the consequences of a move. The early chess playing programs attempted to look forward as far as possible, perhaps to 3 moves ahead by each player, and to choose the best outcome. This led to the horizon effect, whereby a key move 4 or more moves ahead would be unexamined and therefore missed. Consequently, the programs were quite weak and heuristics to manage the search became important in their development. CHAOS used a selective search strategy with iterative widening. As chess programs evolved, they incorporated books of opening lines of play from historic sources. Nowadays, book moves are catalogued in machine-readable form, but originally programmers had to type them in. CHAOS had an extensive book for its time of around 10,000 moves that O'Keefe helped to develop. A problem with play from an opening book is the behavior of the program when the play leaves the book: the positional advantage may be so subtle that the evaluation scheme may be unable to understand it, leading to very wide and shallow searches to establish a line of play. The horizon effect again plagues move selection after leaving the book. CHAOS mitigated these problems by only using book lines that it could understand, and by relying on cached analyses of continuations out of the book made while the opponent's clock was running. == Game Play History == CHAOS played in twelve ACM computer chess tournaments and four World Computer Chess Championships (WCCC). Its debut was the ACM computer chess tournament in 1973, taking 2nd place. In 1974, it again won 2nd place in the WCCC, defeating the tournament favorite Chess 4.0 but losing to Kaissa. CHAOS was close to winning the 1980 WCCC, but lost to Belle in a playoff. The 1985 ACM computer chess tournament was CHAOS' last competition. One of CHAOS' notable victories was over Chess 4.0 at the 1974 WCCC tournament. Chess 4.0 was unbeaten by any other program up until then. Playing as white, CHAOS made a knight sacrifice (16 Nd4-e6!!) that traded material for open lines of attack and eventually won the game. CHAOS’ authors thought the move was due to a